> <> <> Where <<[·; ·]>> denotes concatenation of a scalar onto a column vector. The parameters are <>, <> and <> , <> . Here we use 100 hidden units, i.e.<< h = 100>>. We have <>, the dimension of an MNIST image. We train with a batch size of 100 for 160 epochs. We use the standard training set of 60,000 images, and the standard test set of 10,000 images as a validation/test set. We optimize our model using SGD with momentum with β = 0.9. Our learning rate schedule is 1e-1 for the first 60 epochs, 1e-2 until epoch 100, 1e-3 until epoch 140, and 1e-4 for the final 20 epochs. B.3 Continuous Generative Modelling of Time-Series The PhysioNet dataset consists of observations of 41 distinct traits over a time period of 48 hours. We remove the parameters “Age”, “Gender”, “Height”, and “ICUType” as these attributes do not vary in time. We also quantize the measurements for each attribute by the hour by averaging multiple measurements within the same hour. This leaves 49 unique time stamps (the extra time stamp for observations at exactly the endpoint of the 48 hour observation period). We report all our losses on this quantized data. We performed this rather coarse quantization for computational reasons having to do with our particular implementation of this model. The validation split was obtained by taking a random split of 20% of the trajectories from the full dataset. In total there are 8000 trajectories. Code is included for processing the dataset, and links to downloading the data may be found in the code for Rubanova et al. (2019). All other experimental details may be found in the main body and appendices of Rubanova et al. (2019). B.4 Continuous Normalizing Flows For the model trained on the MINIBOONE tabular dataset from Papamakarios et al. (2017), we used the same architecture as in Table 4 in the appendix of Grathwohl et al. (2019). We chose the number of epochs and a learning rate schedule based on manual tuning on the validation set, in contrast to Grathwohl et al. (2019) who tuned these automatically using early stopping and an automatic heuristic for the learning rate decay using evaluation on a validation set. In particular, we trained for 500 epochs with a learning rate of 1e-3 for the first 300 epochs, 1e-4 until epoch 425, and 1e-5 for the remaining 75 epochs. The number of epochs and learning rate schedule was determined by evaluating the model on the validation set every 10 epochs, and decaying the learning rate by a factor of 10 once the loss on the validation set stopped improving for several evaluations, with the goal of matching or improving upon the log-likelihood reported in Grathwohl et al. (2019). The data was obtained as made available from Papamakarios et al. (2017), which was already processed and split into train/validation/test. In particular, the training set has 29556 examples, the validation set has 3284 examples, and the test set has 3648 examples, which consist of 43 features. It is important to note that we implemented a single-flow model for the MNIST dataset, while the original comparison in Finlay et al. (2020) was on a multi-flow model. This accounts for discrepancy in bits/dim and NFE reported in Finlay et al. (2020). All other experimental details are as in Grathwohl et al. (2019). B.5 Hardware MNIST Supervised learning, Physionet Time-series, and MNIST FFJORD experiments were trained and evaluated on NVIDIA Tesla P100 GPU. Tabular data FFJORD experiments were evaluated on NVIDIA Tesla P100 GPU but trained on NVIDIA Tesla T4 GPU. All experiments except for MNIST FFJORD were trained with double precision for purposes of reproducibility. Appendix C Additional Results C.1 Overfitting of NFE <

> Figure 10: The difference in NFE is tracked by the variance of NFE. In fig. 10 we note that there is a striking correspondence in the variance of NFE across individual examples (in both the train set (dark red) and test set (light red)) and the absolute difference in NFE between examples in the training set and test set. This suggests that any difference in the average NFE between training examples and test examples is explained by noise in the estimate of the true average NFE. It is also interesting that speed regularization does not have a monotonic relationship with the variance of NFE, and we speculate as to how this might interact between the correspondence of NFE for a particular example and the difficulty in the model correctly classifying it. C.2 Trading off function evaluations with a surrogate loss In fig. 11 and fig. 12 we confirm that our method poses a suitable tradeoff not only on the loss being optimized, but also on the potentially non-differentiable loss which we truly care about. On MNIST, we get a similar pareto curve when plotting classification error as opposed to cross-entropy loss, and similarly on the time-series modelling task we see that we get a similar pareto curve on MSE loss as compared to IWAE loss. The pareto curves are plotted for R3 , R2 respectively. <

> Figure 11: MNIST Classification <

> Figure 12: Physionet Time-Series C.3 Wall-clock Time We include additional tables with wall-clock time and training with fixed grid solvers in table 3 and table 4. Appendix D Comparison to How to Train Your Neural ODE The terms from Finlay et al. (2020) are <> and an estimate of <> Table 3: Classification on MNIST <> These are combined with a weighted average and integrated along the solution trajectory. These terms are motivated by the expansion <> Namely, eq. (3) regularizes the first total derivative of the solution, f (z(t), t), along the trajectory, and eq. (4) regularizes a stochastic estimate of the Frobenius norm of the spatial derivative, ∇z f (z(t), t), along the solution trajectory. In contrast, R2 regularizes the norm of the second total derivative directly. In particular, this takes into account the ∂f ∂t term. In other words, this accounts for the explicit dependence of f on time, while eq. (3) and eq. (4) capture only the implicit dependence on time through z(t). Even in the case of an autonomous system, that is, where ∂f ∂t is identically 0 and the dynamics f only depend implicitly on time, these terms still differ. Namely, R2 integrates the following along the solution trajectory: <> while Finlay et al. (2020) penalizes the respective norms of the matrix ∇z f (z(t), t) and vector f (z(t), t) separately. Table 4: Density Estimation on Tabular Data (MINIBOONE) <> <> <> <> < <> <> How to Train Your Neural ODE: the World of Jacobian and Kinetic Regularization Chris Finlay 1 Jörn-Henrik Jacobsen 2 Levon Nurbekyan 3 Adam M Oberman 1 Abstract Training neural ODEs on large datasets has not been tractable due to the necessity of allowing the adaptive numerical ODE solver to refine its step size to very small values. In practice this leads to dynamics equivalent to many hundreds or even thousands of layers. In this paper, we overcome this apparent difficulty by introducing a theoretically-grounded combination of both op- timal transport and stability regularizations which encourage neural ODEs to prefer simpler dynam- ics out of all the dynamics that solve a problem well. Simpler dynamics lead to faster conver- gence and to fewer discretizations of the solver, considerably decreasing wall-clock time without loss in performance. Our approach allows us to train neural ODE-based generative models to the same performance as the unregularized dynamics, with significant reductions in training time. This brings neural ODEs closer to practical relevance in large-scale applications. <

> Figure 1. Optimal transport map and a generic normalizing flow. Indeed, it was observed that there is a striking similarity 1. Introduction between ResNets and the numerical solution of ordinary differential equations (E, 2017; Haber & Ruthotto, 2017; Recent research has bridged dynamical systems, a Ruthotto & Haber, 2018; Chen et al., 2018; 2019). In these workhorse of mathematical modeling, with neural networks, works, deep networks are interepreted as discretizations of the defacto function approximator for high dimensional data. an underlying dynamical system, where time indexes the The great promise of this pairing is that the vast mathemat- “depth” of the network and the parameters of the discretized ical machinery stemming from dynamical systems can be dynamics are learned. An alternate viewpoint was taken by leveraged for modelling high dimensional problems in a neural ODEs (Chen et al., 2018), where the dynamics of dimension-independent fashion. the neural network are approximated by an adaptive ODE Connections between neural networks and ordinary differ- solver on the fly. This latter approach is quite compelling ential equations (ODEs) were almost immediately noted as it does not require specifying the number of layers of the after residual networks (He et al., 2016) were first proposed. network beforehand. Furthermore, it allows the learning of homeomorphisms without any structural constraints on the function computed by the residual block. Neural ODEs have shown great promise in the physical sciences (Köhler et al., 2019), in modeling irregular time series (Rubanova et al., 2019), mean field games (Ruthotto et al., 2019), continuous-time modeling (Yildiz et al., 2019; Kanaa et al., 2019), and for generative modeling through normaliz- ing flows with free-form Jacobians (Grathwohl et al., 2019). Recent work has even adapted neural ODEs to the stochas- based on (ODE) which abstain from a priori fixing step-size. tic setting (Li et al., 2020). Despite these successes, some Chen et al.’s method is a continuous-time generalization of hurdles still remain. In particular, although neural ODEs are residual networks, where the dynamics are generated by an memory efficient, they can take a prohibitively long time to adaptive ODE solver that chooses step-size on-the-fly. train, which is arguably one of the main stumbling blocks Because of their adaptive nature, neural ODEs can be more towards their widespread adoption. flexible than ResNets in certain scenarios, such as when In this work we reduce the training time of neural ODEs trading between model speed and accuracy. Moreover given by regularizing the learned dynamics, complementing other a fixed network depth, the memory footprint of neural ODEs recent approaches to this end such as augmented neural is orders of magnitude smaller than a standard ResNet dur- ODEs (Dupont et al., 2019). Without further constraints on ing training. They therefore show great potential on a host their dynamics, high dimensional neural ODEs may learn of applications, including generative modeling and density dynamics which minimize an objective function, but which estimation. An apparent drawback of neural ODEs is their generate irregular solution trajectories. See for example long training time: although a learned function f (· ; θ) may Figure 1b, where an unregularized flow exhibits undesirable generate a map that solves a problem particularly well, the properties due to unnecessarily fluctuating dynamics. As computational cost of numerically integrating (ODE) may a solution, we propose two theoretically motivated regular- be so prohibitive that it is not tractable in practice. In this ization terms arising from an optimal transport viewpoint paper we demonstrate this need not be so: with proper reg- of the learned map, which encourage well-behaved dynam- ularization, it is possible to learn f (· ; θ) so that (ODE) is ics (see 1a left). We empirically demonstrate that proper easily and quickly solved. regularization leads to significant speed-up in training time without loss in performance, thus bringing neural ODEs 2.1. FFJORD closer to deployment on large-scale datasets. Our methods are validated on the problem of generative modelling and In density estimation and generative modeling, we wish density estimation, as an example of where neural ODEs to estimate an unknown data distribution p(x) from which have shown impressive results, but could easily be applied we have drawn N samples. Maximum likelihood seeks to elsewhere. approximate p(x) with a parameterized distribution pθ (x) by minimizing the Kullback-Leibler divergence between the In summary, our proposed regularized neural ODE (RN- two, or equivalently minimizing ODE) achieves the same performance as the baseline, while reducing the wall-clock training time by many hours or even days. <> (1) 2. Neural ODEs & Continuous normalizing Continuous normalizing flows (Grathwohl et al., 2019; Chen flows et al., 2018) parameterize pθ (x) using a vector field f : Rd × R 7→ Rd as follows. Let z(x, T ) be the solution map Neural ODEs simplify the design of deep neural networks given by running the dynamics (ODE) for fixed time T . by formulating the forward pass of a deep network as the Suppose we are given a known distribution q at final time T , solution of a ordinary differential equation. Initial work such as the normal distribution. Change of variables tells us along these lines was motivated by the similarity of the eval- that the distribution pθ (x) may be evaluated through uation of one layer of a ResNet and the Euler discretization of an ODE. Suppose the block in the t-th layer of a ResNet <> (2) is given by the function f (x, t; θ), where θ are the block’s parameters. Then the evaluation of this layer of the ResNet Evaluating the log determinant of the Jacobian is difficult. is simply xt+1 = xt + f (xt , t; θ). Now, instead consider Grathwohl et al. (2019) exploit the following identity from the following ODE fluid mechanics (Villani, 2003, p 114) <> (ODE) <> (3) The Euler discretization of this ODE with step-size <<τ>> is where <> is the divergence operator, <>, which is nearly identical to the i ∂xi fi (x)>>. By the fundamental theorem of calculus, we forward evaluation of the ResNet’s layer (setting step-size 1 In the normalizing flow literature divergence is typically writ- <<τ = 1>> gives equality). Armed with this insight, Chen et al. ten explicitly as the trace of the Jacobian, however we use div (·) (2018) suggested a method for training neural networks which is more common elsewhere. <

> Figure 2. Log-likelihood (measured in bits/dim) on the validation set as a function of wall-clock time. Rolling average of three hours, with 90% confidence intervals. may then rewrite (2) in integral form From this simple motivating example, the need for regular- ity of the vector field is apparent. Without placing demands on the vector field f , it is entirely possible that the learned <> dynamics will be poorly conditioned. This is not just a theo- (4) retical exercise: because the dynamics must be solved with Remark 2.1 (Divergence trace estimate). In (Grathwohl a numerical integrator, poorly conditioned dynamics will et al., 2019), the divergence is estimated using an unbiased lead to difficulties during numerical integration of (ODE). Monte-Carlo trace estimate (Hutchinson, 1990; Avron & Indeed, later we present results demonstrating a clear corre- Toledo, 2011), lation between the number of time steps an adaptive solver takes to solve (ODE), and the regularity of f .  <> (5) How can the regularity of the vector field be measured? One motivating approach is to measure the force experienced by a particle z(t) under the dynamics generated by the vector By using the substitution (4), the task of maximizing log- field f , which is given by the total derivative of f with likelihood shifts from choosing pθ to minimize (1), to learn- respect to time ing the flow generated by a vector field f . This results in a normalizing flow with a free-form Jacobian and reversible dynamics, and was named FFJORD by Grathwohl et al.. <> (6) 2.2. The need for regularity <> (7) The vector field learned through FFJORD that maximizes Well conditioned flows will place constant, or nearly con- the log-likelihood is not unique, and raises troubling prob- stant, force on particles as they travel. Thus, in this work we lems related to the regularity of the flow. For a simple propose regularizing the dynamics with two penalty terms, example, refer to Figure 1, where we plot two normaliz- one term regularizing f and the other ∇ f . The first penalty, ing flows, both mapping a toy one-dimensional distribution presented in Section 3, is a measure of the distance travelled to the unit Gaussian, and where both maximize the log- under the flow f , and can alternately be interpreted as the likelihood of exactly the same sample of particles. Figure kinetic energy of the flow. This penalty term is based off 1a presents a “regular” flow, where particles travel in straight of numerical methods in optimal transport, and encourages lines that travel with constant speed. In contrast, Figure 1b particles to travel in straight lines with constant speed. The shows a flow that still maximizes the log-likelihood, but second penalty term, discussed in Section 4, performs regu- that has undesirable properties, such as rapidly varying local larization on the Jacobian of the vector field. Taken together trajectories and non-constant speed. the two terms ensure that the force experienced by a particle under the flow is constant or nearly so. 3.1. Linking normalizing flows to optimal transport These two regularizers will promote dynamics that follow Now suppose we wish to minimize (18a), with q(z) a unit numerically easy-to-integrate paths, thus greatly speeding normal distribution, and p(x) a data distribution, unknown up training time. to us, but from which we have drawn N samples, and which we model as a discrete distribution of Dirac masses. Enforc- 3. Optimal transport maps & ing the initial condition is trivial because we have sampled from p directly. The continuity equation (18b) need not be Benamou-Brenier enforced because we are tracking a finite number of sam- There is a remarkable similarity between density estimation pled particles. However the final time condition ρT = q using continuous time normalizing flows, and the calcula- cannot be implemented directly, since we do not have di- tion of the optimal transport map between two densities rect control on the form ρT (z) takes. Instead, introduce using the Benamou-Brenier formulation (Benamou & Bre- a Kullback-Leibler term to (18a) penalizing discrepancy nier, 2000; Santambrogio, 2015). While a review of optimal between ρT and q. This penalty term has an elegant simpli- transport theory is far outside the scope of this paper, here fication when p(x) is modeled as a distribution of a finite we provide an informal summary of key ideas relevant to number of masses, as is done in generative modeling. Set- continuous normalizing flows. The quadratic-cost optimal ting ρ0 = pθ a brief derivation yields transport map between two densities p(x) and q(x) is a map z : Rd 7→ Rd minimizing the transport cost <> (10) <> (8) With this simplification (18a) becomes subject to the constraint that A q(z) dz = z−1 (A) p(x) dx, in other words that the measure of any set A is preserved under the map z. In a seminal work, Benamou & Brenier <> (11) (2000) showed that rather than solving for minimizers of (8) directly, an indirect (but computationally efficient) method is available by writing z(x, T ) as the solution map of a flow under a vector field f (as in (ODE)) for time T , by For further details on this derivation consult the supplemen- minimizing tary materials. The connection between the Benamou-Brenier formulation <> (9a) of the optimal transport problem on a discrete set of points and continuous normalizing flows is apparent: the optimal transport problem (11) is a regularized form of the continu- <> (9b) ous normalizing flow optimization problem (1). We there- <<ρ0 (x) = p>>, (9c) fore expect that adding a kinetic energy regularization term <<ρT (z) = q>>. (9d) to FFJORD will encourage solution trajectories to prefer straight lines with constant speed. The objective function (18a) is a measure of the kinetic energy of the flow. The constraint (18b) ensures probability mass is conserved. The latter two constraints guarantee the 4. Unbiased Frobenius norm regularization of learned distribution agrees with the source p and target q. the Jacobian Note that the kinetic energy (18a) is an upper bound on the Refering to equation (7), one can see that even if f is regu- transport cost, with equality only at optimality. larized to be small, via a kinetic energy penalty term, if the The optimal flow f minimizing (18) has several particularly Jacobian is large then the force experienced by a particle appealing properties. First, particles induced by the opti- may also still be large. As a result, the error of the numerical mal flow f travel in straight lines. Second, particles travel integrator can be large, which may lead an adaptive solver with constant speed. Moreover, under suitable conditions to make many function evaluations. This relationship is on the source and target distributions, the optimal solution apparent in Figure 3, where we empirically demonstrate the map is unique (Villani, 2008). Therefore the solution map correlation between the number of function evaluations of z(x, t) is entirely characterized by the initial and final posi- f taken by the adaptive solver, and the size of the Jacobian tions: z(x, t) = (1 − Tt )z(x, 0) + Tt z(x, T ). Consequently, norm of f . The correlation is remarkably strong: dynamics given an optimal f it is extraordinarily easy to solve (ODE) governed by a poorly conditioned Jacobian matrix require numerically with minimal computational effort. the adaptive solver to take many small time steps. Algorithm 1 RNODE: regularized neural ODE training of FFJORD <> <

> Figure 3. Number of function evaluations vs Jacobian Frobenius norm of flows on CIFAR10 during training with vanilla FFJORD, using an adaptive ODE solver. \ Avron & Toledo, 2011). For real matrix B, an unbiased <> estimate of the trace is given by <> (14) where <> is drawn from a unit normal distribution. Thus the squared Frobenius norm can be easily estimated by setting B = AAT. Moreover, in particle-based methods, the kinetic energy Turning to the Jacobian <> of a vector valued func- term forces dynamics to travel in straight lines only on tion f : Rd 7→ Rd , recall that the vector-Jacobian product data seen during training, and so the regularity of the map <> may be quickly computed through reverse-mode is only guaranteed on trajectories taken by training data. automatic differentiation. Therefore an unbiased Monte- The issue here is one of generalization: the map may be Carlo estimate of the Frobenius norm of the Jacobian is irregular on off-distribution or perturbed images, and cannot readily available be remedied by the kinetic energy term during training alone. In the context of generalization, Jacobian regularization is <> (15) analagous to gradient regularization, which has been shown to improve generalization (Drucker & LeCun, 1992; Novak <> (16) et al., 2018). For these reasons, we also propose regularizing the Jacobian Conveniently, in the FFJORD framework the quantity through its Frobenius norm. The Frobenius norm k · kF of a <> must be computed during the estimate of the prob- real matrix A can be thought of as the `2 norm of the matrix ability distribution under the flow, in the Monte-Carlo esti- A vectorized mate of the divergence term (5). Thus Jacobian Frobenius <> (12) norm regularization is available with essentially no extra computational cost. Equivalently it may be computed as 5. Algorithm description <> (13) All together, we propose modifying the objective function of the FFJORD continuous normalizing flow (Grathwohl and is the Euclidean norm of the singular values of a matrix. et al., 2019) with the two regularization penalties of Sec- In trace form, the Frobenius norm lends itself to estimation tions 3 & 4. The proposed method is called RNODE, short using a Monte-Carlo trace estimator (Hutchinson, 1990; for regularized neural ODE. Pseudo-code of the method is <

> Table 1. Log-likelihood (in bits/dim) and training time (in hours) on validation images with uniform dequantization. Results on clean images are found in the supplemental materials. For comparison we report both the results of the original FFJORD paper (Grathwohl et al., 2019) and our own independent run of FFJORD (“vanilla”) on CIFAR10 and MNIST. Vanilla FFJORD did not train on ImageNet64 (denoted by “x”). Also reported are results for other flow-based generative modeling papers. Our method (FFJORD with RNODE) has comparable log-likelihood as FFJORD but is significantly faster. <

> Figure 4. Quality of generated samples samples on 5bit CelebA-HQ64 with RNODE. Here temperature annealing (Kingma & Dhariwal, 2018) with T = 0.7 was used to generate visually appealing images. For full sized CelebA-HQ256 samples, consult the supplementary materials. presented in Algorithm 1. The optimization problem to be Here E, l, and n are respectively the kinetic energy, the solved is log determinant of the Jacobian, and the integral of the Frobenius norm of the Jacobian. Both the divergence term and the Jacobian Frobenius norm are approximated with Monte-Carlo trace estimates. In our <> implementation, the Jacobian Frobenius estamate reuses the computatian T ∇ f from the divergence estimate for efficiency. We remark that the kinetic energy term only <> requires the computation of a dot product. Thus just as in FFJORD, our implementation scales linearly with the <> (17) number of time steps taken by the ODE solver. Gradients of the objective function with respect to the net- where z(x, t) is determined by numerically solving (ODE). work parameters are computed using the adjoint sensitivity Note that we take the mean over number of samples and method (Pontryagin et al., 1962; Chen et al., 2018). input dimension. This is to ensure that the choice of regu- larization strength λK and λJ is independent of dimension size and sample size. 6. Experimental design To compute the three integrals and the log-probability under Here we demonstrate the benefits of regularizing neural q of z(x, T ) at final time T , we augment the dynamics of ODEs on generative models, an application where neu- the ODE with three extra terms, so that the entire system ral ODEs have shown strong empirical performance. We solved by the numerical integrator is use four datasets: CIFAR10 (Krizhevsky & Hinton, 2009), MNIST (LeCun & Cortes, 1998), downsampled ImageNet (64x64) (van den Oord et al., 2016), and 5bit CelebA-HQ (256x256) (Karras et al., 2017). We use an identical neural <> (RNODE) architecture to that of Grathwohl et al. (2019). The dynamics (Kingma & Dhariwal, 2018) trained with 40 GPUs for a week; in contrast we train with four GPUs in just under a week. <

> Figure 5. Ablation study of the effect of the two regularizers, comparing two measures of flow regularity during training with a fixed step-size ODE solver. Figure 5a: mean Jacobian Frobenius norm as a function of training epoch. Figure 5b: mean kinetic energy of the flow as a function of training epoch. Figure 5c: number of function evaluations. are defined by a neural network <> where <<θ(t)>> is piecewise constant in time. On MNIST we stable and achieved good performance. The Runge-Kutta use 10 pieces; CIFAR10 uses 14; downsampled ImageNet 4(5) adaptive solver was used on the two larger datasets. We uses 18; and CelebA-HQ uses 26 pieces. Each piece is a have also observed that RNODE improves the training time 4-layer deep convolutional network comprised of 3x3 ker- of the adaptive solvers as well, requiring many fewer func- nels and softplus activation functions. Intermediary layers tion evaluations; however in Python we have found that the have 64 hidden dimensions, and time t is concatenated to fixed grid solver is typically quicker at a specified number the spatial input z. The integration time of each piece is of function evaluations. At test time RNODE uses the same [0, 1]. Weight matrices are chosen to imitate the multi-scale adaptive solver as FFJORD. architecture of Real NVP (Dinh et al., 2017), in that im- We always initialize RNODE so that <>; thus train- ages are ‘squeezed’ via a permutation to halve image height ing begins with an initial identity map. This is done by zero- and width but quadruple the number of channels. Diver- ing the parameters of the last layer in each piece (block), gence of f is estimated using the Gaussian Monte-Carlo following Goyal et al. (2017). The identity map is an ap- trace estimator with one sample of fixed noise per solver propriate choice because it has zero transport cost and zero time-step. Frobenius norm. Moreover the identity map is trivially On MNIST and CIFAR10 we train with a batch size of solveable for any numerical solver, thus training begins 200 and train for 100 epochs on a single GPU3 , using the without any effort required on the solver’s behalf. Adam optimizer (Kingma & Ba, 2015) with a learning rate On all datasets we set both the kinetic energy regularization of 1e−3. On the two larger datasets, we train with four coefficient λK and the Jacobian norm coefficient λJ to 0.01. GPUs, using a per-GPU batch size of respectively 3 and 50 for CelebA-HQ and ImageNet. Data is preprocessed by per- turbing with uniform noise followed by the logit transform. 7. Results The reference implementation of FFJORD solves the dy- A comparison of RNODE against FFJORD and other flow- namics using a Runge-Kutta 4(5) adaptive solver (Dormand based generative models is presented in Table 1. We report & Prince, 1980) with error tolerances 1e−5 and initial step both our running of “vanilla” FFJORD and the results as size 1e−2. We have found that using less accurate solvers originally reported in (Grathwohl et al., 2019). We highlight on the reference implementation of FFJORD results in nu- that RNODE runs roughly 2.8x faster than FFJORD on both merically unstable training dynamics. In contrast, a simple datasets, while achieving or surpassing the performance of fixed-grid four stage Runge-Kutta solver suffices for RN- FFJORD. This can further be seen in Figure 2 where we plot ODE during training on MNIST and CIFAR10, using a bits per dimension ( − d1 log2 p(x), a normalized measure step size of 0.25. The step size was determined based on of log-likelihood) on the validation set as a function of a simple heuristic of starting with 0.5 and decreasing the training epoch, for both datasets. Visual inspection of the sample quality reveals no qualitative difference between <

> Figure 6. Quality of generated samples samples with and without regularization on MNIST, left, and CIFAR10, right. regularized and unregularized approaches; refer to Figure 6. encourages flows to travel a minimal distance. In addition, Generated images for downsampled ImageNet and CelebA- we see that the Jacobian norm alone also has a beneficial HQ are deferred to the supplementary materials; we provide effect on the distance particles travel. Overall, the results smaller generated images for networks trained on CelebA- support our theoretical reasoning empirically. HQ 64x64 in Figure 4. Surprisingly, our run of “vanilla” FFJORD achieved slightly 8. Previous generative flows inspired by better performance than the results reported in (Grathwohl optimal transport et al., 2019). We suspect the discrepancy in performance and run times between our implementation of FFJORD and Zhang et al. (2018) define a neural ODE flow where the that of the original paper is due to batch size: Grathwohl dynamics are given as the gradient of a scalar potential func- et al. use a batch size of 900 and train on six GPUs, whereas tion. This interpretation has deep connections to optimal on MNIST and CIFAR10 we use a batch size of 200 and transport: the optimal transport map is the gradient of a train on a single GPU. convex potential function. Yang & Karniadakis (2019) con- tinue along these lines, and define an optimal transport again We were not able to train vanilla FFJORD on ImageNet64, as a scalar potential gradient. Yang & Karniadakis (2019) due to numerical underflow in the adaptive solver’s time step. enforce that the learned map is in fact an optimal trans- This issue cannot be remedied by increasing the solver’s port map by penalizing their objective function with a term error tolerance, for this would bias the log-likelihood esti- measuring violations of the continuity equation. Ruthotto mates on validation. et al. (2019) place generative flows within a broader context of mean field games, and as an example consider a neural 7.1. Ablation study on MNIST ODE gradient potential flow solving the optimal transport problem in up to 100 dimensions. We also note the recent In Figure 5, we compare the effect of each regularizer by work of Twomey et al. (2019), who proposed regularizing itself on the training dynamics with the fixed grid ODE neural ODEs with an Euler-step discretization of the kinetic solver on the MNIST dataset. Without any regularization at energy term to enforce ‘straightness’, although connections all, training dynamics are numerically unstable and fail after to optimal transport were not discussed. just under 50 epochs. This is precisely when the Jacobian norm grows large; refer to Figure 5a. Figure 5a demonstrates When a flow is the gradient of a scalar potential, the change that each regularizer by itself is able to control the Jacobian of variables formula (4) simplifies so that the divergence norm. The Jacobian regularizer is better suited to this task, term is replaced by the Laplacian of the scalar potential. although it is interesting that the kinetic energy regularizer Although mathematically parsimonious and theoretically also improves the Jacobian norm. Unsurprisingly Figure 5b well-motivated, we chose not to implement our flow as the demonstrates the addition of the kinetic energy regularizer gradient of a scalar potential function due to computational How to Train Your Neural ODE: the World of Jacobian and Kinetic Regularization constraints: such an implementation would require ‘triple through CIFAR, and companies sponsoring the Vector Insti- backprop’ (twice to compute or approximate the Laplacian, tute (www.vectorinstitute.ai/#partners). and once more for the parameter gradient). Ruthotto et al. (2019) circumvented this problem by utilizing special struc- References tural properties of residual networks to efficiently compute the Laplacian. Avron, H. and Toledo, S. Randomized algorithms for esti- mating the trace of an implicit symmetric positive semi- definite matrix. J. ACM, 58(2):8:1–8:34, 2011. doi: 9. Discussion 10.1145/1944345.1944349. URL https://doi.org/ In practice, RNODE is simple to implement, and only re- 10.1145/1944345.1944349. quires augmenting the dynamics (ODE) with two extra scalar equations (one for the kinetic energy term, and an- Behrmann, J., Grathwohl, W., Chen, R. T. Q., Duve- other for the Jacobian penalty). In the setting of FFJORD, naud, D., and Jacobsen, J. Invertible residual networks. because we may recycle intermediary terms used in the In Chaudhuri, K. and Salakhutdinov, R. (eds.), Pro- divergence estimate, the computational cost of evaluating ceedings of the 36th International Conference on Ma- these two extra equations is minimal. RNODE introduces chine Learning, ICML 2019, 9-15 June 2019, Long two extra hyperparameters related to the strength of the reg- Beach, California, USA, volume 97 of Proceedings ularizers; we have found these required almost no tuning. of Machine Learning Research, pp. 573–582. PMLR, 2019. URL http://proceedings.mlr.press/ Although the problem of classification was not considered v97/behrmann19a.html. in this work, we believe RNODE may offer similar im- provements both in training time and the regularity of the Benamou, J.-D. and Brenier, Y. A computational fluid me- classifier learned. In the classification setting we expect the chanics solution to the Monge-Kantorovich mass transfer computional overhead of calculating the two extra terms problem. Numerische Mathematik, 84(3):375–393, 2000. should be marginal relative to gains made in training time. Chen, T. Q., Rubanova, Y., Bettencourt, J., and Duve- naud, D. Neural Ordinary Differential Equations. In 10. Conclusion Advances in Neural Information Processing Systems 31: We have presented RNODE, a regularized method for neu- Annual Conference on Neural Information Processing ral ODEs. This regularization approach is theoretically Systems 2018, NeurIPS 2018, 3-8 December 2018, well-motivated, and encourages neural ODEs to learn well- Montréal, Canada, pp. 6572–6583, 2018. URL http: behaved dynamics. As a consequence, numerical integration //papers.nips.cc/paper/7892-neural- of the learned dynamics is straight forward and relatively ordinary-differential-equations. easy, which means fewer discretizations are needed to solve Chen, T. Q., Behrmann, J., Duvenaud, D., and Jacobsen, the dynamics. In many circumstances, this allows for the re- J. Residual flows for invertible generative modeling. placement of adaptive solvers with fixed grid solvers, which In Wallach, H. M., Larochelle, H., Beygelzimer, can be more efficient during training. This leads to a sub- A., d’Alché-Buc, F., Fox, E. B., and Garnett, R. stantial speed up in training time, while still maintaining (eds.), Advances in Neural Information Processing the same empirical performance, opening the use of neural Systems 32: Annual Conference on Neural Information ODEs to large-scale applications. Processing Systems 2019, NeurIPS 2019, 8-14 Decem- ber 2019, Vancouver, BC, Canada, pp. 9913–9923, Acknowledgements 2019. URL http://papers.nips.cc/paper/ 9183-residual-flows-for-invertible- C. F. and A. O. were supported by a grant from the Innova- generative-modeling. tive Ideas Program of the Healthy Brains and Healthy Lives initiative (HBHL) through McGill University. Dinh, L., Sohl-Dickstein, J., and Bengio, S. Den- L. N. was supported by AFOSR MURI FA9550-18-1-0502, sity estimation using real NVP. In 5th International AFOSR Grant No. FA9550-18-1-0167, and ONR Grant No. Conference on Learning Representations, ICLR 2017, N00014-18-1-2527. Toulon, France, April 24-26, 2017, Conference Track Pro- ceedings, 2017. URL https://openreview.net/ A. O. was supported by the Air Force Office of Scientific forum?id=HkpbnH9lx. Research under award number FA9550-18-1-0167 Dormand, J. R. and Prince, P. J. A family of embedded Resources used in preparing this research were provided, in Runge-Kutta formulae. Journal of computational and part, by the Province of Ontario, the Government of Canada applied mathematics, 6(1):19–26, 1980. How to Train Your Neural ODE: the World of Jacobian and Kinetic Regularization Drucker, H. and LeCun, Y. Improving generalization per- Hutchinson, M. F. A stochastic estimator of the trace of the formance using double backpropagation. IEEE Trans. influence matrix for Laplacian smoothing splines. Com- Neural Networks, 3(6):991–997, 1992. doi: 10.1109/ munications in Statistics-Simulation and Computation, 72.165600. URL https://doi.org/10.1109/ 19(2):433–450, 1990. 72.165600. Kanaa, D., Voleti, V., Kahou, S., and Pal, C. Simple video Dupont, E., Doucet, A., and Teh, Y. W. Augmented generation using neural ODEs. Workshop on Learning neural ODEs. In Wallach, H. M., Larochelle, H., with Rich Experience, Advances in Neural Information Beygelzimer, A., d’Alché-Buc, F., Fox, E. B., and Gar- Processing Systems 32: Annual Conference on Neural nett, R. (eds.), Advances in Neural Information Pro- Information Processing Systems 2019, NeurIPS 2019, cessing Systems 32: Annual Conference on Neural 8-14 December 2019, Vancouver, BC, Canada, 2019. Information Processing Systems 2019, NeurIPS 2019, 8-14 December 2019, Vancouver, BC, Canada, pp. Karras, T., Aila, T., Laine, S., and Lehtinen, J. Pro- 3134–3144, 2019. URL http://papers.nips.cc/ gressive growing of gans for improved quality, stabil- paper/8577-augmented-neural-odes. ity, and variation. CoRR, abs/1710.10196, 2017. URL http://arxiv.org/abs/1710.10196. E, W. A Proposal on Machine Learning via Dynam- ical Systems. Communications in Mathematics and Kingma, D. P. and Ba, J. Adam: A method for stochastic op- Statistics, 5(1):1–11, March 2017. ISSN 2194-671X. timization. In 3rd International Conference on Learning doi: 10.1007/s40304-017-0103-z. URL https:// Representations, ICLR 2015, San Diego, CA, USA, May doi.org/10.1007/s40304-017-0103-z. 7-9, 2015, Conference Track Proceedings, 2015. URL http://arxiv.org/abs/1412.6980. Goyal, P., Dollár, P., Girshick, R. B., Noordhuis, P., Kingma, D. P. and Dhariwal, P. Glow: Generative flow Wesolowski, L., Kyrola, A., Tulloch, A., Jia, Y., and with invertible 1x1 convolutions. In Bengio, S., Wallach, He, K. Accurate, large minibatch SGD: training ima- H. M., Larochelle, H., Grauman, K., Cesa-Bianchi, N., genet in 1 hour. CoRR, abs/1706.02677, 2017. URL and Garnett, R. (eds.), Advances in Neural Information http://arxiv.org/abs/1706.02677. Processing Systems 31: Annual Conference on Neural Grathwohl, W., Chen, R. T. Q., Bettencourt, J., Sutskever, Information Processing Systems 2018, NeurIPS 2018, I., and Duvenaud, D. FFJORD: free-form continu- 3-8 December 2018, Montréal, Canada, pp. 10236– ous dynamics for scalable reversible generative mod- 10245, 2018. URL http://papers.nips.cc/ els. In 7th International Conference on Learning Rep- paper/8224-glow-generative-flow-with- resentations, ICLR 2019, New Orleans, LA, USA, May invertible-1x1-convolutions. 6-9, 2019, 2019. URL https://openreview.net/ Köhler, J., Klein, L., and Noé, F. Equivariant flows: sam- forum?id=rJxgknCcK7. pling configurations for multi-body systems with sym- Haber, E. and Ruthotto, L. Stable architectures for deep metric energies. arXiv preprint arXiv:1910.00753, 2019. neural networks. Inverse Problems, 34(1):014004, 2017. Krizhevsky, A. and Hinton, G. Learning multiple He, K., Zhang, X., Ren, S., and Sun, J. Deep resid- layers of features from tiny images. Technical re- ual learning for image recognition. In 2016 IEEE port, University of Toronto, 2009. URL http:// Conference on Computer Vision and Pattern Recogni- www.cs.toronto.edu/ ̃kriz/cifar.html. tion, CVPR 2016, Las Vegas, NV, USA, June 27-30, LeCun, Y. and Cortes, C. The MNIST database of handwrit- 2016, pp. 770–778. IEEE Computer Society, 2016. doi: ten digits. 1998. URL http://yann.lecun.com/ 10.1109/CVPR.2016.90. URL https://doi.org/ exdb/mnist/. 10.1109/CVPR.2016.90. Li, X., Wong, T. L., Chen, R. T. Q., and Duvenaud, D. Scal- Ho, J., Chen, X., Srinivas, A., Duan, Y., and Abbeel, P. able gradients for stochastic differential equations. CoRR, Flow++: Improving flow-based generative models with abs/2001.01328, 2020. URL http://arxiv.org/ variational dequantization and architecture design. In abs/2001.01328. Chaudhuri, K. and Salakhutdinov, R. (eds.), Proceedings of the 36th International Conference on Machine Learn- Novak, R., Bahri, Y., Abolafia, D. A., Pennington, J., and ing, ICML 2019, 9-15 June 2019, Long Beach, California, Sohl-Dickstein, J. Sensitivity and generalization in neural USA, volume 97 of Proceedings of Machine Learning networks: an empirical study. In 6th International Con- Research, pp. 2722–2730. PMLR, 2019. URL http: ference on Learning Representations, ICLR 2018, Van- //proceedings.mlr.press/v97/ho19a.html. couver, BC, Canada, April 30 - May 3, 2018, Conference How to Train Your Neural ODE: the World of Jacobian and Kinetic Regularization Track Proceedings. OpenReview.net, 2018. URL https: Processing Systems 2019, NeurIPS 2019, 8-14 December //openreview.net/forum?id=HJC2SzZCW. 2019, Vancouver, BC, Canada, pp. 13412–13421, 2019. URL http://papers.nips.cc/paper/9497- Pontryagin, L. S., Mishchenko, E., Boltyanskii, V., and ode2vae-deep-generative-second-order- Gamkrelidze, R. The mathematical theory of optimal odes-with-bayesian-neural-networks. processes. 1962. Zhang, L., E, W., and Wang, L. Monge-Ampère flow for Rubanova, Y., Chen, T. Q., and Duvenaud, D. K. Latent or- generative modeling. CoRR, abs/1809.10188, 2018. URL dinary differential equations for irregularly-sampled time http://arxiv.org/abs/1809.10188. series. In Advances in Neural Information Processing Systems, pp. 5321–5331, 2019. Ruthotto, L. and Haber, E. Deep neural networks motivated by partial differential equations. Journal of Mathematical Imaging and Vision, pp. 1–13, 2018. Ruthotto, L., Osher, S. J., Li, W., Nurbekyan, L., and Fung, S. W. A machine learning framework for solv- ing high-dimensional mean field game and mean field control problems. CoRR, abs/1912.01825, 2019. URL http://arxiv.org/abs/1912.01825. Santambrogio, F. Benamou-Brenier and other continu- ous numerical methods, pp. 219–248. Springer Interna- tional Publishing, Cham, 2015. ISBN 978-3-319-20828- 2. doi: 10.1007/978-3-319-20828-2 6. URL https: //doi.org/10.1007/978-3-319-20828-2 6. Twomey, N., Kozlowski, M., and Santos-Rodrı́guez, R. Neu- ral ODEs with stochastic vector field mixtures. CoRR, abs/1905.09905, 2019. URL http://arxiv.org/ abs/1905.09905. van den Oord, A., Kalchbrenner, N., and Kavukcuoglu, K. Pixel recurrent neural networks. CoRR, abs/1601.06759, 2016. URL http://arxiv.org/ abs/1601.06759. Villani, C. Topics in Optimal Transportation. Graduate studies in mathematics. American Mathematical Society, 2003. ISBN 9780821833124. Villani, C. Optimal Transport: Old and New. Grundlehren der mathematischen Wissenschaften. Springer Berlin Hei- delberg, 2008. ISBN 9783540710509. URL https:// books.google.ca/books?id=hV8o5R7 5tkC. Yang, L. and Karniadakis, G. E. Potential flow gener- ator with L2 Optimal Transport regularity for gener- ative models. CoRR, abs/1908.11462, 2019. URL http://arxiv.org/abs/1908.11462. Yildiz, C., Heinonen, M., and Lähdesmäki, H. ODE2VAE: deep generative second order ODEs with Bayesian neural networks. In Wallach, H. M., Larochelle, H., Beygelz- imer, A., d’Alché-Buc, F., Fox, E. B., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information A. Details of Section 3.1: Benamou-Brenier Hence, multiplying the objective function in (20) by λ and formulation in Lagrangian coordinates ignoring the f -independent term Ex∼p log p(x) we obtain an equivalent objective function The Benamou-Brenier formulation of the optimal transporta- tion (OT) problem in Eulerian coordinates is <> (21) <> (18a) Finally, if we assume that {xi }N i=1 are iid sampled from p, <> (18b) we obtain the empirical objective function <<ρ0 (x) = p>>, (18c) <<ρT (z) = q>>. (18d) <> (22) The connection between continuous normalizing flows (CNF) and OT becomes transparent once we rewrite (18) in Lagrangian coordinates. Indeed, for regular enough velocity B. Additional results fields f one has that the solution of the continuity equation Here we present additional generated samples on the two (18b), (18c) is given by ρt = z(·, t)]p where z is the flow larger datasets considered, CelebA-HQ and ImageNet64. In addition bits/dim on clean images are reported in Table 2. <> The relation ρt = z(·, t)]p means that for arbitrary test function φ we have that <<φ(x)ρt (x, t)dx = φ(z(x, t))p(x)dx>> Therefore (18) can be rewritten as <> (19a) <>, (19b) <>, (19c) <>. (19d) Note that ρt is eliminated in this formulation. The terminal condition (18d) is trivial to implement in Eulerian coordi- nates (grid-based methods) but not so simple in Lagrangian ones (19d) (grid-free methods). To enforce (19d) we intro- duce a penalty term in the objective function that measures the deviation of z(·, T )]p from q. Thus, the penalized ob- jective function is <> (20) where λ > 0 is the penalization strength. Next, we observe that this objective function can be written as an expectation with respect to x ∼ p. Indeed, the Kullback-Leibler di- vergence is invariant under coordinate transformations, and therefore <> <

> Figure 7. Quality of FFJORD RNODE generated images on ImageNet-64. <

> Figure 8. Quality of FFJORD RNODE generated images on CelebA-HQ. We use temperature annealing, as described in (Kingma & Dhariwal, 2018), to generate visually appealing images, with T = 0.5, . . . , 1. Table 2. Additional results and model statistics of FFJORD RNODE. Here we report validation bits/dim on both validation images, and on validation images with uniform variational dequantization (ie perturbed by uniform noise). We also report number of trainable model parameters. <

> <> <> <> <> <> <> A guide to convolution arithmetic for deep learning The authors of this guide would like to thank David Warde-Farley, Guillaume Alain and Caglar Gulcehre for their valuable feedback. We are likewise grateful to all those who helped improve this tutorial with helpful comments, constructive criticisms and code contributions. Keep them coming! Special thanks to Ethan Schoonover, creator of the Solarized color scheme, 1 whose colors were used for the figures. Feedback Your feedback is welcomed! We did our best to be as precise, infor- mative and up to the point as possible, but should there be any thing you feel might be an error or could be rephrased to be more precise or com- prehensible, please don’t refrain from contacting us. Likewise, drop us a line if you think there is something that might fit this technical report and you would like us to discuss – we will make our best effort to update this document. Source code and animations The code used to generate this guide along with its figures is available on GitHub. 2 There the reader can also find an animated version of the figures. 1 Introduction 5 1.1 Discrete convolutions . . . . . . . . . . . . . . . . . . . . . . . . .6 1.2 Pooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10 2 Convolution arithmetic 12 2.1 No zero padding, unit strides . . . . . . . . . . . . . . . . . . . .12 2.2 Zero padding, unit strides . . . . . . . . . . . . . . . . . . . . . .13 2.2.1 Half (same) padding . . . . . . . . . . . . . . . . . . . . .13 2.2.2 Full padding . . . . . . . . . . . . . . . . . . . . . . . . .13 2.3 No zero padding, non-unit strides . . . . . . . . . . . . . . . . . .15 2.4 Zero padding, non-unit strides . . . . . . . . . . . . . . . . . . . .15 3 Pooling arithmetic 18 4 Transposed convolution arithmetic 19 4.1 Convolution as a matrix operation . . . . . . . . . . . . . . . . .20 4.2 Transposed convolution . . . . . . . . . . . . . . . . . . . . . . .20 4.3 No zero padding, unit strides, transposed . . . . . . . . . . . . .21 4.4 Zero padding, unit strides, transposed . . . . . . . . . . . . . . .22 4.4.1 Half (same) padding, transposed . . . . . . . . . . . . . .22 4.4.2 Full padding, transposed . . . . . . . . . . . . . . . . . . .22 4.5 No zero padding, non-unit strides, transposed . . . . . . . . . . .24 4.6 Zero padding, non-unit strides, transposed . . . . . . . . . . . . .24 5 Miscellaneous convolutions 28 5.1 Dilated convolutions . . . . . . . . . . . . . . . . . . . . . . . . .28 Chapter 1 Introduction Deep convolutional neural networks (CNNs) have been at the heart of spectac- ular advances in deep learning. Although CNNs have been used as early as the nineties to solve character recognition tasks (Le Cunet al., 1997), their current widespread application is due to much more recent work, when a deep CNN was used to beat state-of-the-art in the ImageNet image classification challenge (Krizhevskyet al., 2012). Convolutional neural networks therefor e constitute a very useful tool for ma- chine learning practitioners. However, learning to use CNNs for the first time is generally an intimidating experience. A convolutional layer’s output shape is affected by the shape of its input as well as the choice of kernel shape, zero padding and strides, and the relationship between these properties is not triv- ial to infer. This contrasts with fully-connected layers, whose output size is independent of the input size. Additionally, CNNs also usually feature apool- ingstage, adding yet another level of complexity with respect to fully-connected networks. Finally, so-called transposed convolutional layers (also known as frac- tionally strided convolutional layers) have been employed in more and more work as of late (Zeileret al., 2011; Zeiler and Fergus, 2014; Longet al., 2015; Rad- for det al., 2015; Visinet al., 2015; Imet al., 2016), and their relationship with convolutional layers has been explained with various degrees of clarity. This guide’s objective is twofold: 1.Explain the relationship between convolutional layers and transposed con- volutional layers. 2.Provide an intuitive underst and ing of the relationship between input shape, kernel shape, zero padding, strides and output shape in convolutional, pooling and transposed convolutional layers. In order to remain broadly applicable, the results shown in this guide are independent of implementation details and apply to all commonly used machine learning frameworks, such as Theano (Bergstraet al., 2010; Bastienet al., 2012), Torch (Collobertet al., 2011), Tensorflow (Abadiet al., 2015) and Caffe (Jia et al., 2014). This chapter briefly reviews the main building blocks of CNNs, namely dis- crete convolutions and pooling. for an in-depth treatment of the subject, see Chapter 9 of the Deep Learning textbook (Goodfellowet al., 2016). 1.1 Discrete convolutions The bread and butter of neural networks is affine transformations: a vector is received as input and is multiplied with a matrix to produce an output (to which a bias vector is usually added before passing the result through a non- linearity). This is applicable to any type of input, be it an image, a sound clip or an unordered collection of features: whatever their dimensionality, their representation can always be flattened into a vector before the transfomation. Images, sound clips and many other similar kinds of data have an intrinsic structure. More formally, they share these important properties: They are stored as multi-dimensional arrays. They feature one or more axes for which ordering matters (e.g., width and height axes for an image, time axis for a sound clip). One axis, called the channel axis, is used to access different views of the data (e.g., the red, green and blue channels of a color image, or the left and right channels of a stereo audio track). These properties are not exploited when an affine transformation is applied; in fact, all the axes are treated in the same way and the topological information is not taken into account. Still, taking advantage of the implicit structure of the data may prove very h and y in solving some tasks, like computer vision and speech recognition, and in these cases it would be best to preserve it. This is where discrete convolutions come into play. A discrete convolution is a linear transformation that preserves this notion of ordering. It is sparse (only a few input units contribute to a given output unit) and reuses parameters (the same weights are applied to multiple locations in the input). Figure 1.1 provides an example of a discrete convolution. The light blue grid is called the input feature map. To keep the drawing simple, a single input feature map is represented, but it is not uncommon to have multiple feature maps stacked one onto another. 1 A kernel(shaded area) of value <

> Figure 1.1: Computing the output values of a discrete convolution. <

> Figure 1.2: Computing the output values of a discrete convolution for N = 2, i1 =i2 = 5, k1 =k2 = 3, s1 =s2 = 2, and p1 =p2 = 1. slides across the input feature map. At each location, the product between each element of the kernel and the input element it overlaps is computed and the results are summed up to obtain the output in the current location. The procedure can be repeated using different kernels to for m as many output feature maps as desired (Figure 1.3). The final outputs of this procedure are called output feature maps.2 If there are multiple input feature maps, the kernel will have to be 3-dimensional – or, equivalently each one of the feature maps will be convolved with a distinct kernel – and the resulting feature maps will be summed up elementwise to produce the output feature map. The convolution depicted in Figure 1.1 is an instance of a 2-D convolution, but it can be generalized to N-D convolutions. for instance, in a 3-D convolu- tion, the kernel would be a cuboid and would slide across the height, width and depth of the input feature map. The collection of kernels defining a discrete convolution has a shape corre- sponding to some permutation of(n;m;k 1 ;:::;k N ), where <> The following properties affect the output size oj of a convolutional layer along axis j: <> for instance, Figure 1.2 shows a 3x3 kernel applied to a 5x5 input padded with a 1x1 border of zeros using 2x2 strides. Note that strides constitute a for m of subsampling. As an alternative to being interpreted as a measure of how much the kernel is translated, strides can also be viewed as how much of the output is retained. for instance, moving the kernel by hops of two is equivalent to moving the kernel by hops of one but retaining only odd output elements (Figure 1.4). 1 An example of this is what was referred to earlier as channels for images and sound clips. 2 While there is a distinction between convolution and cross-correlation from a signal pro- cessing perspective, the two become interchangeable when the kernel is learned. for the sake of simplicity and to stay consistent with most of the machine learning literature, the term convolution will be used in this guide. <

> Figure 1.3: A convolution mapping from two input feature maps to three output feature maps using a32 3x3 collection of kernels w. In the left pathway, input feature map 1 is convolved with kernel w1;1 and input feature map 2 is convolved with kernel w1;2 , and the results are summed together elementwise to for m the first output feature map. The same is repeated for the middle and right pathways to for m the second and third feature maps, and all three output feature maps are grouped together to for m the output. <

> Figure 1.4: An alternative way of viewing strides. Instead of translating the 3x3 kernel by increments ofs= 2(left), the kernel is translated by increments of1 and only one ins= 2output elements is retained (right). 1.2 Pooling In addition to discrete convolutions themselves,pooling operations make up another important building block in CNNs. Pooling operations reduce the size of feature maps by using some function to summarize subregions, such as taking the average or the maximum value. Pooling works by sliding a window across the input and feeding the content of the window to a pooling function. In some sense, pooling works very much like a discrete convolution, but replaces the linear combination described by the kernel with some other function. Figure 1.5 provides an example for average pooling, and Figure 1.6 does the same for max pooling. The following properties affect the output size j of a pooling layer along axisj: <> <

> Figure 1.5: Computing the output values of a 3x3 average pooling operation on a 5x5 input using 1x1 strides. <

> Figure 1.6: Computing the output values of a 3x3 max pooling operation on a 5X5 input using 1X1 strides. Convolution arithmetic The analysis of the relationship between convolutional layer properties is eased by the fact that they don’t interact across axes, i.e., the choice of kernel size, stride and zero padding along axis j only affects the output size of axis j. Because of that, this chapter will focus on the following simplified setting: 2-D discrete convolutions (N= 2), square inputs (i1 =i2 =i), square kernel size (k1 =k2 =k), same strides along both axes (s1 =s2 =s), same zero padding along both axes (p1 =p2 =p). This facilitates the analysis and the visualization, but keep in mind that the results outlined here also generalize to the N-D and non-square cases. 2.1 No zero padding, unit strides The simplest case to analyze is when the kernel just slides across every position of the input (i.e.,s= 1 and p= 0). Figure 2.1 provides an example for i= 4 and k= 3. One way of defining the output size in this case is by the number of possible placements of the kernel on the input. Let’s consider the width axis: the kernel starts on the leftmost part of the input feature map and slides by steps of one until it touches the right side of the input. The size of the output will be equal to the number of steps made, plus one, accounting for the initial position of the kernel (Figure 2.8a). The same logic applies for the height axis. More formally, the following relationship can be inferred: Relationship 1.for any i,k and p, and for s= 1, <> 2.2 Zero padding, unit strides To factor in zero padding (i.e., only restricting tos= 1), let’s consider its effect on the effective input size: padding with p zeros changes the effective input size from i to i+ 2p. In the general case, Relationship 1 can then be used to infer the following relationship: Relationship 2.for any i,k and p, and for s= 1, <> Figure 2.2 provides an example for i= 5,k= 4 and p= 2. In practice, two specific instances of zero padding are used quite extensively because of their respective properties. Let’s discuss them in more detail. 2.2.1 Half (same) padding Having the output size be the same as the input size (i.e.,o=i) can be a desirable property: Relationship 3.for any i and for k o d (k= 2n+ 1; n2N), s= 1 and p=b k=2 c=n, <> This is sometimes referred to as half(or same) padding. Figure 2.3 provides an example for i= 5,k= 3 and (therefor e) p= 1. 2.2.2 Full padding While convolving a kernel generally decreases the output size with respect to the input size, sometimes the opposite is required. This can be achieved with proper zero padding: Relationship 4.for any i and k, and for p=kx1 and s= 1, <> <

> Figure 2.1: (No padding, unit strides) Convolving a 3x3 kernel over a 4x4 input using unit strides (i.e.,i= 4,k= 3,s= 1 and p= 0). <

> Figure 2.2: (Arbitrary padding, unit strides) Convolving a 4x4 kernel over a 5x5 input padded with a 2x2 border of zeros using unit strides (i.e.,i= 5, k= 4,s= 1 and p= 2). <

> Figure 2.3: (Half padding, unit strides) Convolving a 3x3 kernel over a 5x5 input using half padding and unit strides (i.e.,i= 5,k= 3,s= 1 and p= 1). <

> Figure 2.4: (Full padding, unit strides) Convolving a 3x3 kernel over a 5x5 input using full padding and unit strides (i.e.,i= 5,k= 3,s= 1 and p= 2). This is sometimes referred to as full padding, because in this setting every possible partial or complete superimposition of the kernel on the input feature map is taken into account. Figure 2.4 provides an example for i= 5,k= 3 and (therefore) p= 2. 2.3 No zero padding, non-unit strides All relationships derived so far only apply for unit-strided convolutions. Incorporating non unitary strides requires another inference leap. To facilitate the analysis, let’s momentarily ignore zero padding (i.e.,s >1 and p= 0). Figure 2.5 provides an example for i= 5,k= 3 and s= 2. Once again, the output size can be defined in terms of the number of possible placements of the kernel on the input. Let’s consider the width axis: the kernel starts as usual on the leftmost part of the input, but this time it slides by steps of sizes until it touches the right side of the input. The size of the output is again equal to the number of steps made, plus one, accounting for the initial position of the kernel (Figure 2.8b). The same logic applies for the height axis. From this, the following relationship can be inferred: Relationship 5.for any i,k and s, and for p= 0, <> The floor function accounts for the fact that sometimes the last possible step does not coincide with the kernel reaching the end of the input, i.e., some input units are left out (see Figure 2.7 for an example of such a case). 2.4 Zero padding, non-unit strides The most general case (convolving over a zero padded input using non-unit strides) can be derived by applying Relationship 5 on an effective input of size i+ 2p, in analogy to what was done for Relationship 2: Relationship 6.for any i,k,p and s, <> As before, the floor function means that in some cases a convolution will produce the same output size for multiple input sizes. More specifically, ifi+ 2pkis a multiple ofs, then any input size j=i+a; a2 f0;:::; sx1 g will produce the same output size. Note that this ambiguity applies only for s >1. <

> Figure 2.6 shows an example with i= 5,k= 3,s= 2 and p= 1, while <

> Figure 2.7 provides an example for i= 6,k= 3,s= 2 and p= 1. Interestingly, despite having different input sizes these convolutions share the same output size. While this doesn’t affect the analysis for convolutions, this will complicate the analysis in the case of transposed convolutions. <

> Figure 2.5: (No zero padding, arbitrary strides) Convolving a 3x3 kernel over a 5x5 input using 2x2 strides (i.e.,i= 5,k= 3,s= 2 and p= 0). <

> Figure 2.6: (Arbitrary padding and strides) Convolving a 3x3 kernel over a 5x5 input padded with a 1x1 border of zeros using 2x2 strides (i.e.,i= 5, k= 3,s= 2 and p= 1). <

> Figure 2.7: (Arbitrary padding and strides) Convolving a 3x3 kernel over a 6x6 input padded with a 1x1 border of zeros using 2x2 strides (i.e.,i= 6, k= 3,s= 2 and p= 1). In this case, the bottom row and right column of the zero padded input are not covered by the kernel. (a) The kernel has to slide two steps (b) The kernel has to slide one step of to the right to touch the right side of size two to the right to touch the right the input ( and equivalently downwards). side of the input ( and equivalently down- Adding one to account for the initial ker- wards). Adding one to account for the nel position, the output size is 3x3. initial kernel position, the output size is 2x2. <

> Figure 2.8: Counting kernel positions. Chapter 3 Pooling arithmetic In a neural network, pooling layers provide invariance to small translations of the input. The most common kind of pooling is max pooling, which consists in splitting the input in (usually non-overlapping) patches and outputting the maximum value of each patch. Other kinds of pooling exist, e.g., mean or average pooling, which all share the same idea of aggregating the input locally by applying a non-linearity to the content of some patches (Boureauet al., 2010a,b, 2011; Saxeet al., 2011). Some readers may have noticed that the treatment of convolution arithmetic only relies on the assumption that some function is repeatedly applied onto subsets of the input. This means that the relationships derived in the previous chapter can be reused in the case of pooling arithmetic. Since pooling does not involve zero padding, the relationship describing the general case is as follows: Relationship 7.for any i,k and s, <> This relationship holds for any type of pooling. Chapter 4 Transposed convolution arithmetic The need for transposed convolutions generally arises from the desire to use a transfor mation going in the opposite direction of a normal convolution, i.e., from something that has the shape of the output of some convolution to something that has the shape of its input while maintaining a connectivity pattern that is compatible with said convolution. for instance, one might use such a trans- for mation as the decoding layer of a convolutional autoencoder or to project feature maps to a higher-dimensional space. Once again, the convolutional case is considerably more complex than the fully-connected case, which only requires to use a weight matrix whose shape has been transposed. However, since every convolution boils down to an efficient im- plementation of a matrix operation, the insights gained from the fully-connected case are useful in solving the convolutional case. Like for convolution arithmetic, the dissertation about transposed convolu- tion arithmetic is simplified by the fact that transposed convolution properties don’t interact across axes. The chapter will focus on the following setting: 2-D transposed convolutions (N= 2), square inputs (i1 =i2 =i), square kernel size (k1 =k2 =k), same strides along both axes (s1 =s2 =s), same zero padding along both axes (p1 =p2 =p). Once again, the results outlined generalize to the N-D and non-square cases. 4.1 Convolution as a matrix operation Take for example the convolution represented in Figure 2.1. If the input and output were to be unrolled into vectors from left to right, top to bottom, the convolution could be represented as a sparse matrix C where the non-zero elements are the elements w i;j of the kernel (with i and j being the row and column of the kernel respectively): <> This linear operation takes the input matrix flattened as a 16-dimensional vector and produces a 4-dimensional vector that is later reshaped as the 2x2 output matrix. Using this representation, the backward pass is easily obtained by trans- posingC; in other words, the error is backpropagated by multiplying the loss withCT . This operation takes a 4-dimensional vector as input and produces a 16-dimensional vector as output, and its connectivity pattern is compatible withCby construction. Notably, the kernel w defines both the matrices C and CT used for the for ward and backward passes. 4.2 Transposed convolution Let’s now consider what would be required to go the other way around, i.e., map from a 4-dimensional space to a 16-dimensional space, while keeping the connectivity pattern of the convolution depicted in Figure 2.1. This operation is known as a transposed convolution. Transposed convolutions – also called fractionally strided convolutions or deconvolutions 1 – work by swapping the for ward and backward passes of a con- volution. One way to put it is to note that the kernel defines a convolution, but whether it’s a direct convolution or a transposed convolution is determined by how the for ward and backward passes are computed. for instance, although the kernel w defines a convolution whose for ward and backward passes are computed by multiplying with C and CT respectively, it also defines a transposed convolution whose for ward and backward passes are computed by multiplying withCT and (CT )T =C respectively. 2 Finally note that it is always possible to emulate a transposed convolution with a direct convolution. The disadvantage is that it usually involves adding 1 The term “deconvolution” is sometimes used in the literature, but we advocate against it on the grounds that a deconvolution is mathematically defined as the inverse of a convolution, which is different from a transposed convolution. 2 The transposed convolution operation can be thought of as the gradient of some convolution with respect to its input, which is usually how transposed convolutions are implemented in practice. many columns and rows of zeros to the input, resulting in a much less efficient implementation. Building on what has been introduced so far, this chapter will proceed some- what backwards with respect to the convolution arithmetic chapter, deriving the properties of each transposed convolution by referring to the direct convolution with which it shares the kernel, and defining the equivalent direct convolution. 4.3 No zero padding, unit strides, transposed The simplest way to think about a transposed convolution on a given input is to imagine such an input as being the result of a direct convolution applied on some initial feature map. The transposed convolution can be then considered as the operation that allows to recover the shape 3 of this initial feature map. Let’s consider the convolution of a 3x3 kernel on a 4x4 input with unitary stride and no padding (i.e.,i= 4,k= 3,s= 1 and p= 0). As depicted in Figure 2.1, this produces a 2x2 output. The transpose of this convolution will then have an output of shape 4x4 when applied on a 2x2 input. Another way to obtain the result of a transposed convolution is to apply an equivalent – but much less efficient – direct convolution. The example described so far could be tackled by convolving a 3x3 kernel over a 2x2 input padded with a 2x2 border of zeros using unit strides (i.e.,i0 = 2,k0 =k,s0 = 1 and p0 = 2), as shown in Figure 4.1. Notably, the kernel’s and stride’s sizes remain the same, but the input of the transposed convolution is now zero padded. 4 One way to understand the logic behind zero padding is to consider the connectivity pattern of the transposed convolution and use it to guide the design of the equivalent convolution. for example, the top left pixel of the input of the direct convolution only contribute to the top left pixel of the output, the top right pixel is only connected to the top right output pixel, and so on. To maintain the same connectivity pattern in the equivalent convolution it is necessary to zero pad the input in such a way that the first (top-left) application of the kernel only touches the top-left pixel, i.e., the padding has to be equal to the size of the kernel minus one. Proceeding in the same fashion it is possible to determine similar observa- tions for the other elements of the image, giving rise to the following relationship: 3 Note that the transposed convolution does not guarantee to recover the input itself, as it is not defined as the inverse of the convolution, but rather just returns a feature map that has the same width and height. 4 Note that although equivalent to applying the transposed matrix, this visualization adds a lot of zero multiplications in the for m of zero padding. This is done here for illustration purposes, but it is inefficient, and software implementations will normally not perfor m the useless zero multiplications. Relationship 8.A convolution described bys= 1,p= 0 and k has an associated transposed convolution described byk0 =k,s0 =s and p0 = kx1 and its output size is <> Interestingly, this corresponds to a fully padded convolution with unit strides. 4.4 Zero padding, unit strides, transposed Knowing that the transpose of a non-padded convolution is equivalent to con- volving a zero padded input, it would be reasonable to suppose that the trans- pose of a zero padded convolution is equivalent to convolving an input padded withlesszeros. It is indeed the case, as shown in Figure 4.2 for i= 5,k= 4 and p= 2. for mally, the following relationship applies for zero padded convolutions: Relationship 9.A convolution described by s= 1,k and phas an associated transposed convolution described by k0 =k,s0 =s and p0 =kp1 and its output size is <> 4.4.1 Half (same) padding, transposed By applying the same inductive reasoning as befor e, it is reasonable to expect that the equivalent convolution of the transpose of a half padded convolution is itself a half padded convolution, given that the output size of a half padded convolution is the same as its input size. Thus the following relation applies: Relationship 10.A convolution described byk= 2n+1; n2N, s= 1 and p=bk=2c=nh as an associated transposed convolution described byk0 =k,s0 =s and p0 =p and its output size is <> <

> Figure 4.3 provides an example for i= 5,k= 3 and (therefor e)p= 1. 4.4.2 Full padding, transposed Knowing that the equivalent convolution of the transpose of a non-padded con- volution involves full padding, it is unsurprising that the equivalent of the trans- pose of a fully padded convolution is a non-padded convolution: <

> Figure 4.1: The transpose of convolving a 3x3 kernel over a 4x4 input using unit strides (i.e.,i= 4,k= 3,s= 1 and p= 0). It is equivalent to convolving a 3x3 kernel over a 2x2 input padded with a 2x2 border of zeros using unit strides (i.e.,i0 = 2,k0 =k,s0 = 1 and p0 = 2). <

> Figure 4.2: The transpose of convolving a 4x4 kernel over a 5x5 input padded with a 2x2 border of zeros using unit strides (i.e.,i= 5,k= 4,s= 1 and p= 2). It is equivalent to convolving a 4x4 kernel over a 6x6 input padded with a 1x1 border of zeros using unit strides (i.e.,i0 = 6,k0 =k,s0 = 1 and p0 = 1). <

> Figure 4.3: The transpose of convolving a 3x3 kernel over a 5x5 input using half padding and unit strides (i.e.,i= 5,k= 3,s= 1 and p= 1). It is equivalent to convolving a 3x3 kernel over a 5x5 input using half padding and unit strides (i.e.,i0 = 5,k0 =k,s0 = 1 and p0 = 1). Relationship 11.A convolution described bys= 1,k and p= kx1 has an associated transposed convolution described byk0 =k,s0 =s and p0 = 0 and its output size is <

> Figure 4.4 provides an example for i= 5,k= 3 and (therefor e)p= 2. 4.5 No zero padding, non-unit strides, transposed Using the same kind of inductive logic as for zero padded convolutions, one might expect that the transpose of a convolution with s >1 involves an equiv- alent convolution with s <1. As will be explained, this is a valid intuition, which is why transposed convolutions are sometimes called fractionally strided convolutions. Figure 4.5 provides an example for i= 5,k= 3 and s= 2which helps understand what fractional strides involve: zeros are inserted between input units, which makes the kernel move around at a slower pace than with unit strides. 5 for the moment, it will be assumed that the convolution is non-padded (p= 0) and that its input size i is such that ixk is a multiple ofs. In that case, the following relationship holds: Relationship 12.A convolution described byp= 0,k and s and whose input size is such that ixk is a multiple ofs, has an associated transposed convolution described by~i0 ,k0 =k,s0 = 1 and p0 = kx1 , where~i0 is the size of the stretched input obtained by adding sx1 zeros between each input unit, and its output size is <> 4.6 Zero padding, non-unit strides, transposed When the convolution’s input sizeiis such thati+ 2pkis a multiple ofs, the analysis can extended to the zero padded case by combining Relationship 9 and Relationship 12: 5 Doing so is inefficient and real-world implementations avoid useless multiplications by zero, but conceptually it is how the transpose of a strided convolution can be thought of. <

> Figure 4.4: The transpose of convolving a 3x3 kernel over a 5x5 input using full padding and unit strides (i.e.,i= 5,k= 3,s= 1 and p= 2). It is equivalent to convolving a 3x3 kernel over a77input using unit strides (i.e.,i0 = 7, k0 =k,s0 = 1 and p0 = 0). <

> Figure 4.5: The transpose of convolving a 3x3 kernel over a 5x5 input using 2x2 strides (i.e.,i= 5,k= 3,s= 2 and p= 0). It is equivalent to convolving a 3x3 kernel over a 2x2 input (with1zero inserted between inputs) padded with a 2x2 border of zeros using unit strides (i.e.,i0 = 2,~i0 = 3,k0 =k,s0 = 1 and p0 = 2). <

> Figure 4.6: The transpose of convolving a 3x3 kernel over a 5x5 input padded with a 1x1 border of zeros using 2x2 strides (i.e.,i= 5,k= 3,s= 2 and p= 1). It is equivalent to convolving a 3x3 kernel over a 3x3 input (with 1zero inserted between inputs) padded with a 1x1 border of zeros using unit strides (i.e.,i0 = 3,~i0 = 5,k0 =k,s0 = 1 and p0 = 1). Relationship 13.A convolution described byk,s and p and whose input sizeiis such tha ti+2pk is a multiple of s has an associated transposed convolution described by~i0 ,k0 =k,s0 = 1 and p0 = kp1, where ~i0 is the size of the stretched input obtained by adding sx1 zeros between each input unit, and its output size is <> <

> Figure 4.6 provides an example for i= 5,k= 3,s= 2 and p= 1. The constraint on the size of the inputican be relaxed by introducing another parametera2 f0;:::; sx1 gthat allows to distinguish between thes different cases that all lead to the samei0 : Relationship 14.A convolution described byk,s and phas an associated transposed convolution described bya,~i0 ,k0 =k,s0 = 1 and p0 =kp1, where~i0 is the size of the stretched input obtained by adding sx1 zeros between each input unit, and a= (i+ 2pk) modsrepresents the number of zeros added to the bottom and right edges of the input, and its output size is <> <

> Figure 4.7 provides an example for i= 6,k= 3,s= 2 and p= 1. <

> Figure 4.7: The transpose of convolving a 3x3 kernel over a 6x6 input padded with a 1x1 border of zeros using 2x2 strides (i.e.,i= 6,k= 3,s= 2 and p= 1). It is equivalent to convolving a 3x3 kernel over a 2x2 input (with 1zero inserted between inputs) padded with a 1x1 border of zeros (with an additional border of size1added to the bottom and right edges) using unit strides (i.e.,i0 = 3,~i0 = 5,a= 1,k0 =k,s0 = 1 and p0 = 1). Chapter 5 Miscellaneous convolutions 5.1 Dilated convolutions Readers familiar with the deep learning literature may have noticed the term “dilated convolutions” (or “atrous convolutions”, from the French expressioncon- volutions à trous) appear in recent papers. Here we attempt to provide an in- tuitive underst and ing of dilated convolutions. for a more in-depth description and to underst and in what contexts they are applied, see Chenet al.(2014); Yu and Koltun (2015). Dilated convolutions “inflate” the kernel by inserting spaces between the ker- nel elements. The dilation “rate” is controlled by an additional hyperparameter d. Implementations may vary, but there are usually dx1 spaces inserted between kernel elements such thatd= 1corresponds to a regular convolution. Dilated convolutions are used to cheaply increase the receptive field of output units without increasing the kernel size, which is especially effective when multi- ple dilated convolutions are stacked one after another. for a concrete example, see Oordet al.(2016), in which the proposed WaveNet model implements an autoregressive generative model for raw audio which uses dilated convolutions to condition new audio frames on a large context of past audio frames. To underst and the relationship tying the dilation rated and the output size o, it is useful to think of the impact ofdon theeffective kernel size. A kernel of sizekdilated by a factordhas an effective size <> This can be combined with Relationship 6 to for m the following relationship for dilated convolutions: Relationship 15.for any i,k,p and s, and for a dilation rated, <> <

> Figure 5.1: (Dilated convolution) Convolving a 3x3 kernel over a77input with a dilation factor of 2 (i.e.,i= 7,k= 3,d= 2,s= 1 and p= 0). Figure 5.1 provides an example for i= 7,k= 3 and d= 2. Bibliography Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin, M.,et al.(2015). Tensorflow: Large- scale machine learning on heterogeneous systems. Software available from tensorflow.org. Bastien, F., Lamblin, P., Pascanu, R., Bergstra, J., Goodfellow, I., Bergeron, A., Bouchard, N., Warde-Farley, D., and Bengio, Y. (2012). Theano: new features and speed improvements.arXiv preprint arXiv:1211.5590. Bergstra, J., Breuleux, O., Bastien, F., Lamblin, P., Pascanu, R., Desjardins, G., Turian, J., Warde-Farley, D., and Bengio, Y. (2010). Theano: A cpu and gpu math compiler in python. InProc. 9th Python in Science Conf, pages 1–7. Boureau, Y., Bach, F., LeCun, Y., and Ponce, J. (2010a). Learning mid-level features for recognition. InProc. International Conference on Computer Vi- sion and Pattern Recognition (CVPR’10). IEEE. Boureau, Y., Ponce, J., and LeCun, Y. (2010b). A theoretical analysis of feature pooling in vision algorithms. InProc. International Conference on Machine learning (ICML’10). Boureau, Y., Le Roux, N., Bach, F., Ponce, J., and LeCun, Y. (2011). Ask the locals: multi-way local pooling for image recognition. InProc. International Conference on Computer Vision (ICCV’11). IEEE. Chen, L.-C., Pap and reou, G., Kokkinos, I., Murphy, K., and Yuille, A. L. (2014). Semantic image segmentation with deep convolutional nets and fully con- nected crfs.arXiv preprint arXiv:1412.7062. Collobert, R., Kavukcuoglu, K., and Farabet, C. (2011). Torch7: A matlab-like environment for machine learning. InBigLearn, NIPS Workshop, number EPFL-CONF-192376. Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep learning. Book in preparation for MIT Press. Im, D. J., Kim, C. D., Jiang, H., and Memisevic, R. (2016). Generating images with recurrent adversarial networks.arXiv preprint arXiv:1602.05110. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadar- rama, S., and Darrell, T. (2014). Caffe: Convolutional architecture for fast feature embedding. InProceedings of the ACM International Conference on Multimedia, pages 675–678. ACM. Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. InAdvances in neural infor mation processing systems, pages 1097–1105. Le Cun, Y., Bottou, L., and Bengio, Y. (1997). Reading checks with multilayer graph transfor mer networks. InAcoustics, Speech, and Signal Processing, 1997. ICASSP-97., 1997 IEEE International Conference on, volume 1, pages 151–154. IEEE. Long, J., Shelhamer, E., and Darrell, T. (2015). Fully convolutional networks for semantic segmentation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431–3440. Oord, A. v. d., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., Kalchbrenner, N., Senior, A., and Kavukcuoglu, K. (2016). Wavenet: A generative model for raw audio.arXiv preprint arXiv:1609.03499. Radfor d, A., Metz, L., and Chintala, S. (2015). Unsupervised representa- tion learning with deep convolutional generative adversarial networks.arXiv preprint arXiv:1511.06434. Saxe, A., Koh, P. W., Chen, Z., Bh and , M., Suresh, B., and Ng, A. (2011). On r and om weights and unsupervised feature learning. In L. Getoor and T. Scheffer, editors,Proceedings of the 28th International Conference on Ma- chine Learning (ICML-11), ICML ’11, pages 1089–1096, New York, NY, USA. ACM. Visin, F., Kastner, K., Courville, A. C., Bengio, Y., Matteucci, M., and Cho, K. (2015). Reseg: A recurrent neural network for object segmentation. Yu, F. and Koltun, V. (2015). Multi-scale context aggregation by dilated con- volutions.arXiv preprint arXiv:1511.07122. Zeiler, M. D. and Fergus, R. (2014). Visualizing and underst and ing convolu- tional networks. InComputer vision–ECCV 2014, pages 818–833. Springer. Zeiler, M. D., Taylor, G. W., and Fergus, R. (2011). Adaptive deconvolutional networks for mid and high level feature learning. InComputer Vision (ICCV), 2011 IEEE International Conference on, pages 2018–2025. IEEE. <> <> <> <> <> <> A Survey of Model Compression and Acceleration for Deep Neural Networks Yu Cheng, Duo Wang, Pan Zhou Member IEEE, and Tao Zhang Senior Member IEEE Abstract—Deep convolutional neural networks (CNNs) have [2], [3]. It is also very time-consuming to train such a model recently achieved great success in many visual recognition tasks. to get reasonable performance. In architectures that rely only However, existing deep neural network models are computation- on fully-connected layers, the number of parameters can grow ally expensive and memory intensive, hindering their deployment in devices with low memory resources or in applications with to billions [4]. strict latency requirements. Therefore, a natural thought is to As larger neural networks with more layers and nodes without significantly decreasing the model performance. During becomes critical, especially for some real-time applications the past few years, tremendous progress has been made in such as online learning and incremental learning. In addi- this area. In this paper, we survey the recent advanced tech- niques for compacting and accelerating CNNs model developed. tion, recent years witnessed significant progress in virtual These techniques are roughly categorized into four schemes: reality, augmented reality, and smart wearable devices, cre- parameter pruning and sharing, low-rank factorization, trans- ating unprecedented opportunities for researchers to tackle ferred/compact convolutional filters, and knowledge distillation. fundamental challenges in deploying deep learning systems to Methods of parameter pruning and sharing will be described at portable devices with limited resources (e.g. memory, CPU, the beginning, after that the other techniques will be introduced. For each scheme, we provide insightful analysis regarding the energy, bandwidth). Efficient deep learning methods can have performance, related applications, advantages, and drawbacks significant impacts on distributed systems, embedded devices, etc. Then we will go through a few very recent additional and FPGA for Artificial Intelligence. For example, the ResNet- successful methods, for example, dynamic capacity networks and 50 [5] with 50 convolutional layers needs over 95MB memory stochastic depths networks. After that, we survey the evaluation for storage and over 3.8 billion floating number multiplications matrix, the main datasets used for evaluating the model per- formance and recent benchmarking efforts. Finally, we conclude when processing an image. After discarding some redundant this paper, discuss remaining challenges and possible directions weights, the network still works as usual but saves more than on this topic. 75% of parameters and 50% computational time. For devices Index Terms—Deep Learning, Convolutional Neural Networks, like cell phones and FPGAs with only several megabyte Model Compression and Acceleration, resources, how to compact the models used on them is also important. Achieving these goal calls for joint solutions from many I. INTRODUCTION disciplines, including but not limited to machine learning, op- In recent years, deep neural networks have recently received timization, computer architecture, data compression, indexing, lots of attention, been applied to different applications and and hardware design. In this paper, we review recent works achieved dramatic accuracy improvements in many tasks. on compressing and accelerating deep neural networks, which These works rely on deep networks with millions or even attracted a lot of attention from the deep learning community billions of parameters, and the availability of GPUs with and already achieved lots of progress in the past years. very high computation capability plays a key role in their We classify these approaches into four categories: pa- success. For example, the work by Krizhevskyet al.[1] rameter pruning and sharing, low-rank factorization, trans- achieved breakthrough results in the 2012 ImageNet Challenge ferred/compact convolutional filters, and knowledge distil- using a network containing 60 million parameters with five lation. The parameter pruning and sharing based methods convolutional layers and three fully-connected layers. Usually, explore the redundancy in the model parameters and try to it takes two to three days to train the whole model on remove the redundant and uncritical ones. Low-rank factor- ImagetNet dataset with a NVIDIA K40 machine. Another ization based techniques use matrix/tensor decomposition to example is the top face verification results on the Labeled estimate the informative parameters of the deep CNNs. The Faces in the Wild (LFW) dataset were obtained with networks approaches based on transferred/compact convolutional filters containing hundreds of millions of parameters, using a mix design special structural convolutional filters to reduce the of convolutional, locally-connected, and fully-connected layers parameter space and save storage/computation. The knowledge distillation methods learn a distilled model and train a more Yu Cheng is a Researcher from Microsoft AI & Research, One Microsoft Way, Redmond, WA 98052, USA. compact neural network to reproduce the output of a larger Duo Wang and Tao Zhang are with the Department of Automation, network. Tsinghua University, Beijing 100084, China. In Table I, we briefly summarize these four types of Pan Zhou is with the School of Electronic Information and Communi- methods. Generally, the parameter pruning & sharing, low- cations, Huazhong University of Science and Technology, Wuhan 430074, China. rank factorization and knowledge distillation approaches can IEEE SIGNAL PROCESSING MAGAZINE, SPECIAL ISSUE ON DEEP LEARNING FOR IMAGE UNDERSTANDING (ARXIV EXTENDED VERSION) 2 TABLE I <

> be used in DNN models with fully connected layers and convolutional layers, achieving comparable performances. On the other hand, methods using transferred/compact filters are designed for models with convolutional layers only. Low-rank factorization and transfered/compact filters based approaches provide an end-to-end pipeline and can be easily implemented in CPU/GPU environment, which is straightforward. while parameter pruning & sharing use different methods such as vector quantization, binary coding and sparse constraints to perform the task. Generally it will take several steps to achieve the goal. <

> Fig. 1. The three-stage compression method proposed in [10]: pruning, Regarding the training protocols, models based on param- quantization and encoding. The input is the original model and the output eter pruning/sharing low-rank factorization can be extracted is the compression model. from pre-trained ones or trained from scratch. While the transferred/compact filter and knowledge distillation models can only support train from scratch. These methods are inde- memory usage and float point operations with little loss in pendently designed and complement each other. For example, classification accuracy. transferred layers and parameter pruning & sharing can be The method proposed in [10] quantized the link weights used together, and model quantization & binarization can be using weight sharing and then applied Huffman coding to the used together with low-rank approximations to achieve further quantized weights as well as the codebook to further reduce speedup. We will describe the details of each theme, their the rate. As shown in Figure 1, it started by learning the con- properties, strengths and drawbacks in the following sections. nectivity via normal network training, followed by pruning the small-weight connections. Finally, the network was retrained to learn the final weights for the remaining sparse connections. II. PARAMETER PRUNING AND SHARING This work achieved the state-of-art performance among allEarly works showed that network pruning is effective in parameter quantization based methods. It was shown in [11] reducing the network complexity and addressing the over- that Hessian weight could be used to measure the importancefitting problem [6]. After that researcher found pruning orig- of network parameters, and proposed to minimize Hessian-inally introduced to reduce the structure in neural networks weighted quantization errors in average for clustering networkand hence improve generalization, it has been widely studied parameters.to compress DNN models, trying to remove parameters which In the extreme case of the 1-bit representation of eachare not crucial to the model performance. These techniques can weight, that is binary weight neural networks. There arebe further classified into three sub-categories: quantization and many works that directly train CNNs with binary weights, forbinarization, parameter sharing, and structural matrix. instance, BinaryConnect [12], BinaryNet [13] and XNORNet- works [14]. The main idea is to directly learn binary weights orA. Quantization and Binarization activation during the model training. The systematic study in Network quantization compresses the original network by [15] showed that networks trained with back propagation could reducing the number of bits required to represent each weight. be resilient to specific weight distortions, including binary Gonget al.[6] and Wu et al. [7] appliedk-means scalar weights. quantization to the parameter values. Vanhouckeet al.[8] Drawbacks: the accuracy of the binary nets is significantly showed that 8-bit quantization of the parameters can result lowered when dealing with large CNNs such as GoogleNet. in significant speed-up with minimal loss of accuracy. The Another drawback of such binary nets is that existing bina- work in [9] used 16-bit fixed-point representation in stochastic rization schemes are based on simple matrix approximations rounding based CNN training, which significantly reduced and ignore the effect of binarization on the accuracy loss. IEEE SIGNAL PROCESSING MAGAZINE, SPECIAL ISSUE ON DEEP LEARNING FOR IMAGE UNDERSTANDING (ARXIV EXTENDED VERSION) 3 To address this issue, the work in [16] proposed a proximal connected layers, which is often the bottleneck in terms of Newton algorithm with diagonal Hessian approximation that memory consumption. These network layers use the nonlinear directly minimizes the loss with respect to the binary weights. transformsf(x;M) =(Mx), where()is an element-wise The work in [17] reduced the time on float point multiplication nonlinear operator,xis the input vector, andMis themn in the training stage by stochastically binarizing weights and matrix of parameters [29]. WhenMis a large general dense converting multiplications in the hidden state computation to matrix, the cost of storingmnparameters and computing significant changes. matrix-vector products inO(mn)time. Thus, an intuitive way to prune parameters is to imposexas a parameterizedB. Pruning and Sharing structural matrix. Anmn matrix that can be described Network pruning and sharing has been used both to reduce using much fewer parameters thanmnis called a structured network complexity and to address the over-fitting issue. An matrix. Typically, the structure should not only reduce the early approach to pruning was the Biased Weight Decay memory cost, but also dramatically accelerate the inference [18]. The Optimal Brain Damage [19] and the Optimal Brain and training stage via fast matrix-vector multiplication and Surgeon [20] methods reduced the number of connections gradient computations. based on the Hessian of the loss function, and their work sug- Following this direction, the work in [30], [31] proposed a gested that such pruning gave higher accuracy than magnitude- simple and efficient approach based on circulant projections, while maintaining competitive error rates. Given a vectorr=based pruning such as the weight decay method. The training procedure of those methods followed the way training from <>, a circulant matrix R^2 R^dxd is defined as: <> scratch manner. A recent trend in this direction is to prune redundant, <> non-informative weights in a pre-trained CNN model. For <> example, Srinivas and Babu [21] explored the redundancy <> among neurons, and proposed a data-free pruning method to remove redundant neurons. Hanet al.[22] proposed to reduce <> the total number of parameters and operations in the entire thus the memory cost becomesO(d)instead of O(d^2) network. Chenet al.[23] proposed a HashedNets model that This circulant structure also enables the use of Fast Fourier used a low-cost hash function to group weights into hash Transform (FFT) to speed up the computation. Given ad-buckets for parameter sharing. The deep compression method dimensional vectorr, the above 1-layer circulant neural net-in [10] removed the redundant connections and quantized the work in Eq. 1 has time complexity ofO(dlogd).weights, and then used Huffman coding to encode the quan- In [32], a novel Adaptive Fastfood transform was introducedtized weights. In [24], a simple regularization method based to reparameterize the matrix-vector multiplication of fully on soft weight-sharing was proposed, which included both connected layers. The Adaptive Fast food transform matrix quantization and pruning in one simple (re-)training procedure. R2Rnd was defined as:The above pruning schemes typically produce connections pruning in CNNs. <> (2) There is also growing interest in training compact CNNs whereS,GandBare random diagonal matrices. 2 with sparsity constraints. Those sparsity constraints are typ- <> is a random permutation matrix, and H denotes ically introduced in the optimization problem asl0 orl1 - the Walsh-Hadamard matrix. Reparameterizing a fully con- norm regularizers. The work in [25] imposed group sparsity nected layer with d inputs and n outputs using the Adaptive constraint on the convolutional filters to achieve structured Fast food transform reduces the storage and the computational brain Damage, i.e., pruning entries of the convolution kernels costs from O(n^d) to O(n) and from O(n^d) to O(n*log(d)), in a group-wise fashion. In [26], a group-sparse regularizer respectively. on neurons was introduced during the training stage to learn The work in [29] showed the effectiveness of the new compact CNNs with reduced filters. Wenet al.[27] added a notion of parsimony in the theory of structured matrices. Their structured sparsity regularizer on each layer to reduce trivial proposed method can be extended to various other structured filters, channels or even layers. In the filter-level pruning, all matrix classes, including block and multi-level Toeplitz-like the above works used l2-norm regularizers. The work in [28] [33] matrices related to multi-dimensional convolution [34]. usedl1 -norm to select and prune unimportant filters. Following this idea, [35] proposed a general structured effi- Drawbacks: there are some potential issues of the pruning cient linear layer for CNNs. and sharing. First, pruning with l1 or l2 regularization requires Drawbacks: one problem of this kind of approaches is that more iterations to converge than general. In addition, all the structural constraint will hurt the performance since the pruning criteria require manual setup of sensitivity for layers, constraint might bring bias to the model. On the other hand, which demands fine-tuning of the parameters and could be how to find a proper structural matrix is difficult. There is no cumbersome for some applications. theoretical way to derive it out. C. Designing Structural Matrix III. LOW-RANK FACTORIZATION AND SPARSITY In architectures that contain fully-connected layers, it is Convolution operations contribute the bulk of most com- critical to explore this redundancy of parameters in fully- putations in deep CNNs, thus reducing the convolution layer IEEE SIGNAL PROCESSING MAGAZINE, SPECIAL ISSUE ON DEEP LEARNING FOR IMAGE UNDERSTANDING (ARXIV EXTENDED VERSION) 4 TABLE II COMPARISONS BETWEEN THE LOW -RANK MODELS AND THEIR BASELINES ON ILSVRC-2012. <

> <

> Fig. 2. A typical framework of the low-rank regularization method. The left is the original convolutional layer and the right is the low-rank constraint convolutional layer with rank-K. would improve the compression rate as well as the overall speedup. For the convolution kernels, it can be viewed as a 4D tensor. Ideas based on tensor decomposition is derived by For instance, Mishaet al.[41] reduced the number of dynamic the intuition that there is a significant amount of redundancy parameters in deep models using the low-rank method. [42] in the 4D tensor, which is a particularly promising way to explored a low-rank matrix factorization of the final weight remove the redundancy. Regarding the fully-connected layer, layer in a DNN for acoustic modeling. In [3], Luet al.adopted it can be view as a 2D matrix and the low-rankness can also truncated SVD (singular value decomposition) to decompsite help. the fully connected layer for designing compact multi-task It has been a long time for using low-rank filters to acceler- deep learning architectures. ate convolution, for example, high dimensional DCT (discrete Drawbacks: low-rank approaches are straightforward for cosine transform) and wavelet systems using tensor products model compression and acceleration. The idea complements to be constructed from 1D DCT transform and 1D wavelets recent advances in deep learning, such as dropout, recti- respectively. Learning separable 1D filters was introduced fied units and maxout. However, the implementation is not by Rigamontiet al.[36], following the dictionary learning that easy since it involves decomposition operation, which idea. Regarding some simple DNN models, a few low-rank is computationally expensive. Another issue is that current approximation and clustering schemes for the convolutional methods perform low-rank approximation layer by layer, and kernels were proposed in [37]. They achieved 2speedup thus cannot perform global parameter compression, which for a single convolutional layer with 1% drop in classification is important as different layers hold different information. accuracy. The work in [38] proposed using different tensor Finally, factorization requires extensive model retraining to decomposition schemes, reporting a 4.5speedup with 1% achieve convergence when compared to the original model. drop in accuracy in text recognition. The low-rank approximation was done layer by layer. The IV. T RANSFERRED /COMPACT CONVOLUTIONAL FILTERS parameters of one layer were fixed after it was done, and the CNNs are parameter efficient due to exploring the trans-layers above were fine-tuned based on a reconstruction error lation invariant property of the representations to the input criterion. These are typical low-rank methods for compressing image, which is the key to the success of training very deep2D convolutional layers, which is described in Figure 2. Fol- models without severe over-fitting. Although a strong theory lowing this direction, Canonical Polyadic (CP) decomposition is currently missing, a large number of empirical evidenceof was proposed for the kernel tensors in [39]. Their work support the notion that both the translation invariant property used nonlinear least squares to compute the CP decomposition. and the convolutional weight sharing are important for good In [40], a new algorithm for computing the low-rank tensor predictive performance. The idea of using transferred convolu- decomposition for training low-rank constrained CNNs from tional filters to compress CNN models is motivated by recent scratch were proposed. It used Batch Normalization (BN) to works in [43], which introduced the equivariant group theory.transform the activation of the internal hidden units. In general, Letxbe an input,()be a network or layer and T() be the both the CP and the BN decomposition schemes in [40] (BN transform matrix. The concept of equivalence is defined as:Low-rank) can be used to train CNNs from scratch. However, there are few differences between them. For example, finding <> (3) the best low-rank approximation in CP decomposition is an ill- posed problem, and the best rank-K (K is the rank number) indicating that transforming the input x by the transform T() approximation may not exist sometimes. While for the BN and then passing it through the network or layer () should scheme, the decomposition always exists. We perform a simple give the same result as first mapping x through the network comparison of both methods shown in Table II. The actual and then transforming the representation. Note that in Eq. speedup and the compression rates are used to measure their (10), the transforms <> and <> are not necessarily the performances. same as they operate on different objects. According to this As we mentioned before, the fully connected layers can theory, it is reasonable applying transform to layers or filters be viewed as a 2D matrix and thus the above mentioned () to compress the whole network models. From empirical methods can also be applied there. There are several classical observation, deep CNNs also benefit from using a large set of works on exploiting low-rankness in fully connected layers. convolutional filters by applying certain transformT()to a IEEE SIGNAL PROCESSING MAGAZINE, SPECIAL ISSUE ON DEEP LEARNING FOR IMAGE UNDERSTANDING (ARXIV EXTENDED VERSION) 5 small set of base filters since it acts as a regularizer for the TABLE III model. A SIMPLE COMPARISON OF DIFFERENT APPROACHES ON CIFAR-10 AND Following this direction, there are many recent reworks proposed to build a convolutional layer from a set of base <

> filters [43]–[46]. What they have in common is that the transform T() lies in the family of functions that only operate in the spatial domain of the convolutional filters. For example, the work in [45] found that the lower convolution layers of CNNs learned redundant filters to extract both positive and negative phase information of an input signal, and definedT() Drawbacks: there are few issues to be addressed for ap-to be the simple negation function: proaches that apply transform constraints to convolutional fil- <> (4) ters. First, these methods can achieve competitive performance x for wide/flat architectures (like VGGNet) but not thin/deepwhereWx is the basis convolutional filter andW is the filter x ones (like GoogleNet, Residual Net). Secondly, the transferconsisting of the shifts whose activation is opposite to that assumptions sometimes are too strong to guide the learning,ofWx and selected after max-pooling operation. By doing making the results unstable in some cases.this, the work in [45] can easily achieve 2compression Using a compact filter for convolution can directly reducerate on all the convolutional layers. It is also shown that the the computation cost. The key idea is to replace the loosenegation transform acts as a strong regularizer to improve and over-parametric filters with compact blocks to improve the classification accuracy. The intuition is that the learning the speed, which significantly accelerate CNNs on severalalgorithm with pair-wise positive-negative constraint can lead benchmarks. Decomposing33convolution into two11to useful convolutional filters instead of redundant ones. convolutions was used in [48], which achieved significantIn [46], it was observed that magnitudes of the responses acceleration on object recognition. SqueezeNet [49] was pro-from convolutional kernels had a wide diversity of pattern posed to replace33convolution with11convolu-representations in the network, and it was not proper to discard tion, which created a compact neural network with about 50weaker signals with a single threshold. Thus a multi-bias non- fewer parameters and comparable accuracy when compared tolinearity activation function was proposed to generates more AlexNet.patterns in the feature space at low computational cost. The transformT()was define as: <> (5) V. KNOWLEDGE DISTILLATION To the best of our knowledge, exploiting knowledge transfer where were the multi-bias factors. The work in [47] con- (KT) to compress model was first proposed by Caruanaet side red a combination of rotation by a multiple of 90 and al.[50]. They trained a compressed/ensemble model of strong horizontal/vertical flipping with: classifiers with pseudo-data labeled, and reproduced the output of the original larger network. But the work is limited to <> (6) shallow models. The idea has been recently adopted in [51] whereWT was the transformation matrix which rotated the as knowledge distillation (KD) to compress deep and wide original filters with angle2 f90;180;270g. In [43], the networks into shallower ones, where the compressed model transform was generalized to any angle learned from data, and mimicked the function learned by the complex model. The was directly obtained from data. Both works [47] and [43] main idea of KD based approaches is to shift knowledge from can achieve good classification performance. a large teacher model into a small one by learning the class The work in [44] definedT()as the set of translation distributions output via softmax. functions applied to 2D filters: The work in [52] introduced a KD compression framework, which eased the training of deep networks by following a <> (7) student-teacher paradigm, in which the student was penalized whereT(;x;y)denoted the translation of the first operand by according to a softened version of the teacher’s output. The (x;y)along its spatial dimensions, with proper zero padding framework compressed an ensemble of teacher networks into at borders to maintain the shape. The proposed framework a student network of similar depth. The student was trained can be used to 1) improve the classification accuracy as a to predict the output and the classification labels. Despite regularized version of maxout networks, and 2) to achieve its simplicity, KD demonstrates promising results in various parameter efficiency by flexibly varying their architectures to image classification tasks. The work in [53] aimed to address compress networks. the network compression problem by taking advantage of Table III briefly compares the performance of different depth neural networks. It proposed an approach to train thin methods with transferred convolutional filters, using VGGNet but deep networks, called FitNets, to compress wide and (16 layers) as the baseline model. The results are reported shallower (but still deep) networks. The method was extended on CIFAR-10 and CIFAR-100 datasets with Top-5 error. It is the idea to allow for thinner and deeper student models. In observed that they can achieve reduction in parameters with order to learn from the intermediate representations of teacher little or no drop in classification accuracy. network, FitNet made the student mimic the full feature maps of the teacher. However, such assumptions are too strict since layer with global average pooling [44], [62]. Network architec- the capacities of teacher and student may differ greatly. ture such as GoogleNet or Network in Network, can achieve All the above approaches are validated on MNIST, CIFAR- state-of-the-art results on several benchmarks by adopting 10, CIFAR-100, SVHN and AFLW benchmark datasets, and this idea. However, these architectures have not been fully experimental results show that these methods match or outper- optimized the utilization of the computing resources inside form the teacher’s performance, while requiring notably fewer the network. This problem was noted by Szegedyet al.[62] parameters and multiplications. and motivated them to increase the depth and width of the There are several extension along this direction of dis- network while keeping the computational budget constant. tillation knowledge. The work in [54] trained a parametric The work in [63] targeted the Residual Network based student model to approximate a Monte Carlo teacher. The model with a spatially varying computation time, called proposed framework used online training, and used deep stochastic depth, which enabled the seemingly contradictory neural networks for the student model. Different from previous setup to train short networks and used deep networks at test works which represented the knowledge using the soften label time. It started with very deep networks, while during training, probabilities, [55] represented the knowledge by using the for each mini-batch, randomly dropped a subset of layers neurons in the higher hidden layer, which preserved as much and bypassed them with the identity function. Following this information as the label probabilities, but are more compact. direction, thew work in [64] proposed a pyramidal residual The work in [56] accelerated the experimentation process by networks with stochastic depth. In [65], Wuet al.proposed instantaneously transferring the knowledge from a previous an approach that learns to dynamically choose which layers network to each new deeper or wider network. The techniques of a deep network to execute during inference so as to best are based on the concept of function-preserving transfor- reduce total computation. Veitet al.exploited convolutional mations between neural network specifications. Zagoruyko networks with adaptive inference graphs to adaptively define et al.[57] proposed Attention Transfer (AT) to relax the their network topology conditioned on the input image [66]. assumption of FitNet. They transferred the attention maps that Other approaches to reduce the convolutional overheads in-are summaries of the full activations. clude using FFT based convolutions [67] and fast convolutionDrawbacks: KD-based approaches can make deeper models using the Winograd algorithm [68]. Zhaiet al.[69] proposed athinner and help significantly reduce the computational cost. strategy call stochastic spatial sampling pooling, which speed-However, there are a few disadvantages. One of those is that up the pooling operations by a more general stochastic version.KD can only be applied to classification tasks with softmax Saeedanet al.presented a novel pooling layer for convolu-loss function, which hinders its usage. Another drawback is tional neural networks termed detail-preserving pooling (DPP),the model assumptions sometimes are too strict to make the based on the idea of inverse bilateral filters [70]. Those worksperformance competitive with other type of approaches. only aim to speed up the computation but not reduce the memory storage. VI. OTHER TYPES OF APPROACHES We first summarize the works utilizing attention-based methods. Note that attention-based mechanism [58] can reduce VII. BENCHMARKS , EVALUATION AND DATABASES computations significantly by learning to selectively focus or In the past five years the deep learning community had“attend” to a few, task-relevant input regions. The work in made great efforts in benchmark models. One of the most[59] introduced the dynamic capacity network (DCN) that well-known model used in compression and acceleration forcombined two types of modules: the small sub-networks with CNNs is Alexnet [1], which has been occasionally usedlow capacity, and the large ones with high capacity. The low- for assessing the performance of compression. Other popularcapacity sub-networks were active on the whole input to first standard models include LeNets [71], All-CNN-nets [72] andfind the task-relevant areas, and then the attention mechanism many others. LeNet-300-100 is a fully connected networkwas used to direct the high-capacity sub-networks to focus on with two hidden layers, with 300 and 100 neurons each.the task-relevant regions. By dong this, the size of the CNNs LeNet-5 is a convolutional network that has two convolutionalmodel has been significantly reduced. layers and two fully connected layers. Recently, more andFollowing this direction, the work in [60] introduced the more state-of-the-art architectures are used as baseline modelsconditional computation idea, which only computes the gra- in many works, including network in networks (NIN) [73],dient for some important neurons. It proposed a sparsely- VGG nets [74] and residual networks (ResNet) [75]. Table IVgated mixture-of-experts Layer (MoE). The MoE module summarizes the baseline models commonly used in severalconsisted of a number of experts, each a simple feed-forward typical compression methods.neural network, and a trainable gating network that selected a sparse combination of the experts to process each input. In The standard criteria to measure the quality of model [61], dynamic deep neural networks (D2NN) were introduced, compression and acceleration are the compression and the which were a type of feed-forward deep neural network that speedup rates. Assume thatais the number of the parameters selected and executed a subset of D2NN neurons based on the in the original model Manda is that of the compressed input. model M , then the compression rate (M;M ) of M over There have been other attempts to reduce the number of Mis aparameters of neural networks by replacing the fully connected (M;M ) = : (8)a IEEE SIGNAL PROCESSING MAGAZINE, SPECIAL ISSUE ON DEEP LEARNING FOR IMAGE UNDERSTANDING (ARXIV EXTENDED VERSION) 7 TABLE IV or low rank factorization based methods. If you need SUMMARIZATION OF BASELINE MODELS USED IN DIFFERENT end-to-end solutions for your problem, the low rank REPRESENTATIVE WORKS OF NETWORK COMPRESSION . and transferred convolutional filters approaches could be considered. For applications in some specific domains, methods with low-rank factorization [40] human prior (like the transferred convolutional filters, Network in network [73] low-rank factorization [40] <> structural matrix) sometimes have benefits. For example, when doing medical images classification, transferred Residual networks [75] compact filters [49], stochastic depth [63] convolutional filters could work well as medical images parameter sharing [24] (like organ) do have the rotation transformation property. Usually the approaches of pruning & sharing could give parameter pruning [20], [22] reasonable compression rate while not hurt the accuracy. Thus for applications which requires stable model accu- Another widely used measurement is the index space saving racy, it is better to utilize pruning & sharing. defined in several papers [30], [35] as If your problem involves small/medium size datasets, you can try the knowledge distillation approaches. The com-aa <> (9) pressed student model can take the benefit of transferring a knowledge from teacher model, making it robust datasets where a and a are the number of the dimension of the index which are not large. space in the original model and that of the compressed model, As we mentioned before, techniques of the four groups respectively. are orthogonal. It is reasonable to combine two or three Similarly, given the running timesofMands ofM , of them to maximize the performance. For some spe- the speedup rate <> is defined as: cific applications, like object detection, which requires s both convolutional and fully connected layers, you can <> (10) compress the convolutional layers with low rank based Most work used the average training time per epoch to measure method and the fully connected layers with a pruning the running time, while in [30], [35], the average testing time technique. was used. Generally, the compression rate and speedup rate B. Technique Challengesare highly correlated, as smaller models often results in faster computation for both the training and the testing stages. Techniques for deep model compression and acceleration Good compression methods are expected to achieve almost are still in the early stage and the following challenges still the same performance as the original model with much smaller need to be addressed. parameters and less computational time. However, for different Most of the current state-of-the-art approaches are built applications with different CNN designs, the relation between on well-designed CNN models, which have limited free- parameter size and computational time may be different. dom to change the configuration (e.g., network structural, For example, it is observed that for deep CNNs with fully hyper-parameters). To handle more complicated tasks, connected layers, most of the parameters are in the fully it should provide more plausible ways to configure the connected layers; while for image classification tasks, float compressed models. point operations are mainly in the first few convolutional layers Pruning is an effective way to compress and acceler- since each filter is convolved with the whole image, which is ate CNNs. The current pruning techniques are mostly usually very large at the beginning. Thus compression and designed to eliminate connections between neurons. On acceleration of the network should focus on different type of the other hand, pruning channel can directly reduce the layers for different applications. feature map width and shrink the model into a thinner one. It is efficient but also challenging because removing VIII. D ISCUSSION AND CHALLENGES channels might dramatically change the input of the following layer.In this paper, we summarized recent efforts on compressing and accelerating deep neural networks (DNNs). Here we dis- As we mentioned before, methods of structural matrix and transferred convolutional filters impose prior humancuss more details about how to choose different compression knowledge to the model, which could significantly affectapproaches, and possible challenges/solutions on this area. the performance and stability. It is critical to investigate how to control the impact of those prior knowledge.A. General Suggestions The methods of knowledge distillation provide many ben- There is no golden rule to measure which approach is the efits such as directly accelerating model without special best. How to choose the proper method is really depending hardware or implementations. It is still worthy developing on the applications and requirements. Here are some general KD-based approaches and exploring how to improve their guidance we can provide: performances. If the applications need compacted models from pre- Hardware constraints in various of small platforms (e.g., trained models, you can choose either pruning & sharing mobile, robotic, self-driving car) are still a major problem IEEE SIGNAL PROCESSING MAGAZINE, SPECIAL ISSUE ON DEEP LEARNING FOR IMAGE UNDERSTANDING (ARXIV EXTENDED VERSION) 8 to hinder the extension of deep CNNs. How to make full see more work for applications with larger deep nets (e.g., use of the limited computational source and how to design video and image frames [88], [89]). special compression methods for such platforms are still challenges that need to be addressed. IX. ACKNOWLEDGMENTS Despite the great achievements of these compression ap- proaches, the black box mechanism is still the key barrier The authors would like to thank the reviewers and broader to the adoption. Exploring the knowledge interpret-ability community for their feedback on this survey. In particular, is still an important problem. we would like to thank Hong Zhao from the Department of Automation of Tsinghua University for her help on modifying C. Possible Solutions the paper. This research is supported by National Science Foundation of China with Grant number 61401169.To solve the hyper-parameters configuration problem, we can rely on the recent learning-to-learn strategies [76], [77]. This framework provides a mechanism allowing the algorithm REFERENCES to automatically learn how to exploit structure in the problem [1]A. Krizhevsky, I. Sutskever, and G. Hinton, “Imagenet classification with of interest. Very recently, leveraging reinforcement learning deep convolutional neural networks,” inNIPS, 2012. to efficiently sample the design space and improve the model [2]Y. Taigman, M. Yang, M. Ranzato, and L. Wolf, “Deepface: Closing the compression has also been tried [78]. gap to human-level performance in face verification,” inCVPR, 2014. [3]Y. Lu, A. Kumar, S. Zhai, Y. Cheng, T. Javidi, and R. S. Feris, “Fully- Channel pruning provides the efficiency benefit on both adaptive feature sharing in multi-task networks with applications in CPU and GPU because no special implementation is required. person attribute classification,”CoRR, vol. abs/1611.05377, 2016. But it is also challenging to handle the input configuration. [4]J. Dean, G. Corrado, R. Monga, K. Chen, M. Devin, Q. Le, M. Mao, M. Ranzato, A. Senior, P. Tucker, K. Yang, and A. Ng, “Large scale One possible solution is to use the training-based channel distributed deep networks,” inNIPS, 2012. pruning methods [79], which focus on imposing sparse con- [5]K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image straints on weights during training. However, training from recognition,”CoRR, vol. abs/1512.03385, 2015. [6]Y. Gong, L. Liu, M. Yang, and L. D. Bourdev, “Compressing scratch for such method is costly for very deep CNNs. In deep convolutional networks using vector quantization,”CoRR, vol. [80], the authors provided an iterative two-step algorithm to abs/1412.6115, 2014. effectively prune channels in each layer. [7]Y. W. Q. H. Jiaxiang Wu, Cong Leng and J. Cheng, “Quantized convolutional neural networks for mobile devices,” inIEEE Conference Exploring new types of knowledge in the teacher models on Computer Vision and Pattern Recognition (CVPR), 2016. and transferring it to the student models is useful for the [8]V. Vanhoucke, A. Senior, and M. Z. Mao, “Improving the speed of knowledge distillation (KD) approaches. Instead of directly re- neural networks on cpus,” inDeep Learning and Unsupervised Feature Learning Workshop, NIPS 2011, 2011. ducing and transferring parameters, passing selectivity knowl- [9]S. Gupta, A. Agrawal, K. Gopalakrishnan, and P. Narayanan, “Deep edge of neurons could be helpful. One can derive a way to learning with limited numerical precision,” inProceedings of the select essential neurons related to the task [81], [82]. The 32Nd International Conference on International Conference on Machine Learning - Volume 37, ser. ICML’15, 2015, pp. 1737–1746. intuition is that if a neuron is activated in certain regions [10]S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing or samples, that implies these regions or samples share some deep neural networks with pruning, trained quantization and huffman common properties that may relate to the task. coding,”International Conference on Learning Representations (ICLR), 2016. For methods with the convolutional filters and the structural [11]Y. Choi, M. El-Khamy, and J. Lee, “Towards the limit of network matrix, we can conclude that the transformation lies in the quantization,”CoRR, vol. abs/1612.01543, 2016. family of functions that only operations on the spatial dimen- [12]M. Courbariaux, Y. Bengio, and J. David, “Binaryconnect: Training deep neural networks with binary weights during propagations,” inAdvances sions. Hence to address the imposed prior issue, one solution is in Neural Information Processing Systems 28: Annual Conference on to provide a generalization of the aforementioned approaches Neural Information Processing Systems 2015, December 7-12, 2015, in two aspects: 1) instead of limiting the transformation to Montreal, Quebec, Canada, 2015, pp. 3123–3131. [13]M. Courbariaux and Y. Bengio, “Binarynet: Training deep neural net- belong to a set of predefined transformations, let it be the works with weights and activations constrained to +1 or -1,”CoRR, vol. whole family of spatial transformations applied on 2D filters abs/1602.02830, 2016. or matrix, and 2) learn the transformation jointly with all the [14]M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi, “Xnor-net: Imagenet classification using binary convolutional neural networks,” in model parameters. ECCV, 2016. Regarding the use of CNNs in small platforms, proposing [15]P. Merolla, R. Appuswamy, J. V. Arthur, S. K. Esser, and D. S. Modha, some general/unified approaches is one direction. Wanget al. “Deep neural networks are robust to weight binarization and other non- [83] presented a feature map dimensionality reduction method linear distortions,”CoRR, vol. abs/1606.01981, 2016. [16]L. Hou, Q. Yao, and J. T. Kwok, “Loss-aware binarization of deep by excavating and removing redundancy in feature maps gen- networks,”CoRR, vol. abs/1611.01600, 2016. erated from different filters, which could also preserve intrinsic [17]Z. Lin, M. Courbariaux, R. Memisevic, and Y. Bengio, “Neural networks information of the original network. The idea can be applied with few multiplications,”CoRR, vol. abs/1510.03009, 2015. [18]S. J. Hanson and L. Y. Pratt, “Comparing biases for minimal network to make CNNs more applicable for different platforms. The construction with back-propagation,” inAdvances in Neural Information work in [84] proposed a one-shot whole network compression Processing Systems 1, D. S. Touretzky, Ed., 1989, pp. 177–185. scheme consisting of three components: rank selection, low- [19]Y. L. Cun, J. S. Denker, and S. A. Solla, “Advances in neural information processing systems 2,” D. S. Touretzky, Ed., 1990, ch. Optimal Brain rank tensor decomposition, and fine-tuning to make deep Damage, pp. 598–605. CNNs work in mobile devices. [20]B. Hassibi, D. G. Stork, and S. C. R. Com, “Second order derivatives Despite the classification task, people are also adapting the for network pruning: Optimal brain surgeon,” inAdvances in Neural Information Processing Systems 5. Morgan Kaufmann, 1993, pp. 164– compacted models in other tasks [85]–[87]. We would like to 171. IEEE SIGNAL PROCESSING MAGAZINE, SPECIAL ISSUE ON DEEP LEARNING FOR IMAGE UNDERSTANDING (ARXIV EXTENDED VERSION) 9 [21]S. Srinivas and R. V. Babu, “Data-free parameter pruning for deep neural [43]T. S. Cohen and M. Welling, “Group equivariant convolutional net- networks,” inProceedings of the British Machine Vision Conference works,”arXiv preprint arXiv:1602.07576, 2016. 2015, BMVC 2015, Swansea, UK, September 7-10, 2015, 2015, pp. [44]S. Zhai, Y. Cheng, and Z. M. Zhang, “Doubly convolutional neural 31.1–31.12. networks,” inAdvances In Neural Information Processing Systems, 2016, [22]S. Han, J. Pool, J. Tran, and W. J. Dally, “Learning both weights and pp. 1082–1090. connections for efficient neural networks,” inProceedings of the 28th [45]W. Shang, K. Sohn, D. Almeida, and H. Lee, “Understanding and International Conference on Neural Information Processing Systems, ser. improving convolutional neural networks via concatenated rectified NIPS’15, 2015. linear units,”arXiv preprint arXiv:1603.05201, 2016. [23]W. Chen, J. Wilson, S. Tyree, K. Q. Weinberger, and Y. Chen, “Com- [46]H. Li, W. Ouyang, and X. Wang, “Multi-bias non-linear activation in pressing neural networks with the hashing trick.” JMLR Workshop and deep neural networks,”arXiv preprint arXiv:1604.00676, 2016. Conference Proceedings, 2015. [47]S. Dieleman, J. De Fauw, and K. Kavukcuoglu, “Exploiting cyclic [24]K. Ullrich, E. Meeds, and M. Welling, “Soft weight-sharing for neural symmetry in convolutional neural networks,” inProceedings of the network compression,”CoRR, vol. abs/1702.04008, 2017. 33rd International Conference on International Conference on Machine [25]V. Lebedev and V. S. Lempitsky, “Fast convnets using group-wise brain Learning - Volume 48, ser. ICML’16, 2016. damage,” in2016 IEEE Conference on Computer Vision and Pattern [48]C. Szegedy, S. Ioffe, and V. Vanhoucke, “Inception-v4, inception- Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016, resnet and the impact of residual connections on learning.”CoRR, vol. pp. 2554–2564. abs/1602.07261, 2016. [26]H. Zhou, J. M. Alvarez, and F. Porikli, “Less is more: Towards compact [49]B. Wu, F. N. Iandola, P. H. Jin, and K. Keutzer, “Squeezedet: Unified, cnns,” inEuropean Conference on Computer Vision, Amsterdam, the small, low power fully convolutional neural networks for real-time object Netherlands, October 2016, pp. 662–677. detection for autonomous driving,”CoRR, vol. abs/1612.01051, 2016. [27]W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li, “Learning structured [50]C. Bucilua, R. Caruana, and A. Niculescu-Mizil, “Model compression,”ˇ sparsity in deep neural networks,” inAdvances in Neural Information inProceedings of the 12th ACM SIGKDD International Conference on Processing Systems 29, D. D. Lee, M. Sugiyama, U. V. Luxburg, Knowledge Discovery and Data Mining, ser. KDD ’06, 2006, pp. 535– I. Guyon, and R. Garnett, Eds., 2016, pp. 2074–2082. 541. [28]H. Li, A. Kadav, I. Durdanovic, H. Samet, and H. P. Graf, “Pruning [51]J. Ba and R. Caruana, “Do deep nets really need to be deep?” in filters for efficient convnets,”CoRR, vol. abs/1608.08710, 2016. Advances in Neural Information Processing Systems 27: Annual Confer- [29]V. Sindhwani, T. Sainath, and S. Kumar, “Structured transforms for ence on Neural Information Processing Systems 2014, December 8-13 small-footprint deep learning,” inAdvances in Neural Information Pro- 2014, Montreal, Quebec, Canada, 2014, pp. 2654–2662. cessing Systems 28, C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, [52]G. E. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a and R. Garnett, Eds., 2015, pp. 3088–3096. neural network,”CoRR, vol. abs/1503.02531, 2015. [30]Y. Cheng, F. X. Yu, R. Feris, S. Kumar, A. Choudhary, and S.-F. [53]A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Chang, “An exploration of parameter redundancy in deep networks with Y. Bengio, “Fitnets: Hints for thin deep nets,”CoRR, vol. abs/1412.6550, circulant projections,” inInternational Conference on Computer Vision 2014. (ICCV), 2015. [54]A. Korattikara Balan, V. Rathod, K. P. Murphy, and M. Welling, [31]Y. Cheng, F. X. Yu, R. S. Feris, S. Kumar, A. N. Choudhary, and “Bayesian dark knowledge,” inAdvances in Neural Information Process- S. Chang, “Fast neural networks with circulant projections,”CoRR, vol. ing Systems 28, C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, abs/1502.03436, 2015. and R. Garnett, Eds., 2015, pp. 3420–3428. [32]Z. Yang, M. Moczulski, M. Denil, N. de Freitas, A. Smola, L. Song, [55]P. Luo, Z. Zhu, Z. Liu, X. Wang, and X. Tang, “Face model compression and Z. Wang, “Deep fried convnets,” inInternational Conference on by distilling knowledge from neurons,” inProceedings of the Thirtieth Computer Vision (ICCV), 2015. AAAI Conference on Artificial Intelligence, February 12-17, 2016, [33]J. Chun and T. Kailath,Generalized Displacement Structure for Block- Phoenix, Arizona, USA., 2016, pp. 3560–3566. Toeplitz, Toeplitz-block, and Toeplitz-derived Matrices. Berlin, Heidel- [56]T. Chen, I. J. Goodfellow, and J. Shlens, “Net2net: Accelerating learning berg: Springer Berlin Heidelberg, 1991, pp. 215–236. via knowledge transfer,”CoRR, vol. abs/1511.05641, 2015. [34]M. V. Rakhuba and I. V. Oseledets, “Fast multidimensional convolution [57]S. Zagoruyko and N. Komodakis, “Paying more attention to attention: in low-rank tensor formats via cross approximation,”SIAM J. Scientific Improving the performance of convolutional neural networks via atten- Computing, vol. 37, no. 2, 2015. tion transfer,”CoRR, vol. abs/1612.03928, 2016. [35]M. Moczulski, M. Denil, J. Appleyard, and N. de Freitas, “Acdc: [58]D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by A structured efficient linear layer,” inInternational Conference on jointly learning to align and translate,”CoRR, vol. abs/1409.0473, 2014. Learning Representations (ICLR), 2016. [59]A. Almahairi, N. Ballas, T. Cooijmans, Y. Zheng, H. Larochelle, and [36]R. Rigamonti, A. Sironi, V. Lepetit, and P. Fua, “Learning separable A. C. Courville, “Dynamic capacity networks,” inProceedings of the filters,” in2013 IEEE Conference on Computer Vision and Pattern 33nd International Conference on Machine Learning, ICML 2016, New Recognition, Portland, OR, USA, June 23-28, 2013, 2013, pp. 2754– York City, NY, USA, June 19-24, 2016, 2016, pp. 2549–2558. 2761. [60]N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, [37]E. L. Denton, W. Zaremba, J. Bruna, Y. LeCun, and R. Fergus, and J. Dean, “Outrageously large neural networks: The sparsely-gated “Exploiting linear structure within convolutional networks for efficient mixture-of-experts layer,” 2017. evaluation,” inAdvances in Neural Information Processing Systems 27, [61]D. Wu, L. Pigou, P. Kindermans, N. D. Le, L. Shao, J. Dambre, and Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. J. Odobez, “Deep dynamic neural networks for multimodal gesture Weinberger, Eds., 2014, pp. 1269–1277. segmentation and recognition,”IEEE Trans. Pattern Anal. Mach. Intell., [38]M. Jaderberg, A. Vedaldi, and A. Zisserman, “Speeding up convolutional vol. 38, no. 8, pp. 1583–1597, 2016. neural networks with low rank expansions,” inProceedings of the British [62]C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, Machine Vision Conference. BMVA Press, 2014. V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” [39]V. Lebedev, Y. Ganin, M. Rakhuba, I. V. Oseledets, and V. S. Lempit- inComputer Vision and Pattern Recognition (CVPR), 2015. sky, “Speeding-up convolutional neural networks using fine-tuned cp- [63]G. Huang, Y. Sun, Z. Liu, D. Sedra, and K. Q. Weinberger,Deep decomposition,”CoRR, vol. abs/1412.6553, 2014. Networks with Stochastic Depth, 2016. [40]C. Tai, T. Xiao, X. Wang, and W. E, “Convolutional neural networks [64]Y. Yamada, M. Iwamura, and K. Kise, “Deep pyramidal residual with low-rank regularization,” vol. abs/1511.06067, 2015. networks with separated stochastic depth,”CoRR, vol. abs/1612.01230, [41]M. Denil, B. Shakibi, L. Dinh, M. Ranzato, and N. D. Freitas, 2016. “Predicting parameters in deep learning,” in Advances in Neural [65]Z. Wu, T. Nagarajan, A. Kumar, S. Rennie, L. S. Davis, K. Grauman, and Information Processing Systems 26, C. Burges, L. Bottou, M. Welling, R. Feris, “Blockdrop: Dynamic inference paths in residual networks,” Z. Ghahramani, and K. Weinberger, Eds., 2013, pp. 2148–2156. inCVPR, 2018. [Online]. Available: http://media.nips.cc/nipsbooks/nipspapers/paper [66]A. Veit and S. Belongie, “Convolutional networks with adaptive infer- files/nips26/1053.pdf ence graphs,” 2018. [42]T. N. Sainath, B. Kingsbury, V. Sindhwani, E. Arisoy, and B. Ramab- [67]M. Mathieu, M. Henaff, and Y. Lecun,Fast training of convolutional hadran, “Low-rank matrix factorization for deep neural network training networks through FFTs, 2014. with high-dimensional output targets,” inin Proc. IEEE Int. Conf. on [68]A. Lavin and S. Gray, “Fast algorithms for convolutional neural net- Acoustics, Speech and Signal Processing, 2013. works,” in2016 IEEE Conference on Computer Vision and Pattern IEEE SIGNAL PROCESSING MAGAZINE, SPECIAL ISSUE ON DEEP LEARNING FOR IMAGE UNDERSTANDING (ARXIV EXTENDED VERSION) 10 Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016, [89]L. Cao, S.-F. Chang, N. Codella, C. V. Cotton, D. Ellis, L. Gong, pp. 4013–4021. M. Hill, G. Hua, J. Kender, M. Merler, Y. Mu, J. R. Smith, and F. X. [69]S. Zhai, H. Wu, A. Kumar, Y. Cheng, Y. Lu, Z. Zhang, and R. S. Yu, “Ibm research and columbia university trecvid-2012 multimedia Feris, “S3pool: Pooling with stochastic spatial sampling,”CoRR, vol. event detection (med), multimedia event recounting (mer), and semantic abs/1611.05138, 2016. indexing (sin) systems,” 2012. [70]F. Saeedan, N. Weber, M. Goesele, and S. Roth, “Detail-preserving pooling in deep networks,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. Yu Cheng(yu.cheng@microsoft.com) currently is a [71]Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning Researcher at Microsoft. Before that, he was a Re- applied to document recognition,” inProceedings of the IEEE, 1998, pp. search Staff Member at IBM T.J. Watson Research 2278–2324. Center. Yu got his Ph.D. from Northwestern Univer- [72]J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. A. Ried- sity in 2015 and bachelor from Tsinghua University miller, “Striving for simplicity: The all convolutional net,”CoRR, vol. in 2010. His research is about deep learning in abs/1412.6806, 2014. general, with specific interests in the deep generative [73]M. Lin, Q. Chen, and S. Yan, “Network in network,” inICLR, 2014. model, model compression, and transfer learning. [74]K. Simonyan and A. Zisserman, “Very deep convolutional networks for He regularly serves on the program committees of large-scale image recognition,”CoRR, vol. abs/1409.1556, 2014. top-tier AI conferences such as NIPS, ICML, ICLR, [75]K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image CVPR and ACL. recognition,”arXiv preprint arXiv:1512.03385, 2015. [76]M. Andrychowicz, M. Denil, S. G. Colmenarejo, M. W. Hoffman, D. Pfau, T. Schaul, and N. de Freitas, “Learning to learn by gradient descent by gradient descent,” inNeural Information Processing Systems (NIPS), 2016. Duo Wang (d-wang15@mail.tsinghua.edu.cn) re-[77]D. Ha, A. Dai, and Q. Le, “Hypernetworks,” inInternational Conference ceived the B.S. degree in automation from theon Learning Representations 2016, 2016. Harbin Institute of Technology, China, in 2015.[78]Y. He, J. Lin, Z. Liu, H. Wang, L.-J. Li, and S. Han, “Amc: Automl Currently he is purchasing his Ph.D. degree at thefor model compression and acceleration on mobile devices,” inThe Department of Automation, Tsinghua University,European Conference on Computer Vision (ECCV), September 2018. Beijing, P.R. China. Currently his research interests[79]J. M. Alvarez and M. Salzmann, “Learning the number of neurons in are about deep learning, particularly in few-shotdeep networks,” pp. 2270–2278, 2016. learning and deep generative models. He also works[80]Y. He, X. Zhang, and J. Sun, “Channel pruning for accelerating on a lot of applications in computer vision andvery deep neural networks,” inThe IEEE International Conference on robotics vision.Computer Vision (ICCV), Oct 2017. [81]Z. Huang and N. Wang, “Data-driven sparse structure selection for deep neural networks,”ECCV, 2018. [82]Y. Chen, N. Wang, and Z. Zhang, “Darkrank: Accelerating deep metric learning via cross sample similarities transfer,” inProceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, 2018, pp. 2852– Pan Zhou(panzhou@hust.edu.cn) is currently an 2859. associate professor with School of Electronic In- [83]Y. Wang, C. Xu, C. Xu, and D. Tao, “Beyond filters: Compact feature formation and Communications, Wuhan, China. He map for portable deep model,” inProceedings of the 34th International received his Ph.D. in the School of Electrical and Conference on Machine Learning, ser. Proceedings of Machine Learning Computer Engineering at the Georgia Institute of Research, D. Precup and Y. W. Teh, Eds., vol. 70. International Technology in 2011. Before that, he received his Convention Centre, Sydney, Australia: PMLR, 06–11 Aug 2017, pp. B.S. degree in theAdvanced Classof HUST, and 3703–3711. a M.S. degree in the Department of Electronics [84]Y.-D. Kim, E. Park, S. Yoo, T. Choi, L. Yang, and D. Shin, “Compression and Information Engineering from HUST, Wuhan, of deep convolutional neural networks for fast and low power mobile China, in 2006 and 2008, respectively. His current applications,”CoRR, vol. abs/1511.06530, 2015. research interest includes big data analytics and [85]G. Chen, W. Choi, X. Yu, T. Han, and M. Chandraker, “Learning efficient machine learning, security and privacy, and information networks. object detection models with knowledge distillation,” inAdvances in Neural Information Processing Systems 30, I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, Eds., 2017, pp. 742–751. [86]M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, Tao Zhang (taozhang@mail.tsinghua.edu.cn) ob- “Mobilenetv2: Inverted residuals and linear bottlenecks,” inThe IEEE tained his B.S., M.S., and Ph.D. degrees from Ts- Conference on Computer Vision and Pattern Recognition (CVPR), June inghua University, Beijing, China, in 1993, 1995, 2018. and 1999, respectively, and another Ph.D. degree [87]J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara, A. Fathi, I. Fischer, from Saga University, Saga, Japan, in 2002, all in Z. Wojna, Y. Song, S. Guadarrama, and K. Murphy, “Speed/accuracy control engineering. He is currently a Professor with trade-offs for modern convolutional object detectors,” in2017 IEEE the Department of Automation, Tsinghua University. Conference on Computer Vision and Pattern Recognition, CVPR 2017, He serves the Associate Dean, School of Information Honolulu, HI, USA, July 21-26, 2017, 2017, pp. 3296–3297. Science and Technology and Head of the Department [88]Y. Cheng, Q. Fan, S. Pankanti, and A. Choudhary, “Temporal sequence of Automation. His current research interests include modeling for video event detection,” in The IEEE Conference on artificial intelligence, robotics, image processing, Computer Vision and Pattern Recognition (CVPR), June 2014. control theory, and control of spacecraft. <> <> <> <> <> <> Analysis and Design of Echo State Networks Mustafa C. Ozturk can@cnel.ufl.edu Dongming Xu dmxu@cnel.ufl.edu Jose C. Principe principe@cnel.ufl.edu Computational NeuroEngineering Laboratory, Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL 32611, U.S.A. The design of echo state network (ESN) parameters relies on the selec- tion of the maximum eigenvalue of the linearized system around zero (spectral radius). However, this procedure does not quantify in a sys- tematic manner the performance of the ESN in terms of approximation error. This article presents a functional space approximation framework to better understand the operation of ESNs and proposes an information- theoretic metric, the average entropy of echo states, to assess the richness of the ESN dynamics. Furthermore, it provides an interpretation of the ESN dynamics rooted in system theory as families of coupled linearized systems whose poles move according to the input signal dynamics. With this interpretation, a design methodology for functional approximation is put forward where ESNs are designed with uniform pole distributions covering the frequency spectrum to abide by the richness metric, irre- spective of the spectral radius. A single bias parameter at the ESN input, adapted with the modeling error, configures the ESN spectral radius to the input-output joint space. Function approximation examples compare the proposed design methodology versus the conventional design. 1 Introduction Dynamic computational models require the ability to store and access the time history of their inputs and outputs. The most common dynamic neural architecture is the time-delay neural network (TDNN) that couples delay lines with a nonlinear static architecture where all the parameters (weights) are adapted with the backpropagation algorithm. The conventional delay line utilizes ideal delay operators, but delay lines with local first-order re- cursive filters have been proposed by Werbos (1992) and extensively stud- ied in the gamma model (de Vries, 1991; Principe, de Vries, & de Oliviera, 1993). Chains of first-order integrators are interesting because they effec- tively decrease the number of delays necessary to create time embeddings (Principe, 2001). Recurrent neural networks (RNNs) implement a differ- ent type of embedding that is largely unexplored. RNNs are perhaps the most biologically plausible of the artificial neural network (ANN) models (Anderson, Silverstein, Ritz, & Jones, 1977; Hopfield, 1984; Elman, 1990), but are not well understood theoretically (Siegelmann & Sontag, 1991; Siegelmann, 1993; Kremer, 1995). One of the main practical problems with RNNs is the difficulty to adapt the system weights. Various algorithms, such as backpropagation through time (Werbos, 1990) and real-time recur- rent learning (Williams & Zipser, 1989), have been proposed to train RNNs; however, these algorithms suffer from computational complexity, resulting in slow training, complex performance surfaces, the possibility of instabil- ity, and the decay of gradients through the topology and time (Haykin, 1998). The problem of decaying gradients has been addressed with spe- cial processing elements (PEs) (Hochreiter & Schmidhuber, 1997). Alter- native second-order training methods based on extended Kalman filtering (Singhal & Wu, 1989; Puskorius & Feldkamp, 1994; Feldkamp, Prokhorov, Eagen, & Yuan, 1998) and the multistreaming training approach (Feldkamp et al., 1998) provide more reliable performance and have enabled practical applications in identification and control of dynamical systems (Kechri- otis, Zervas, & Monolakos, 1994; Puskorius & Feldkamp, 1994; Delgado, Kambhampati, & Warwick, 1995). Recently,twonewrecurrentnetworktopologieshavebeenproposed:the echo state network (ESN) by Jaeger (2001, 2002a; Jaeger & Hass, 2004) and the liquid state machine (LSM) by Maass (Maass, Natschlager, & Markram,¨ 2002). ESNs possess a highly interconnected and recurrent topology of nonlinear PEs that constitutes a “reservoir of rich dynamics” (Jaeger, 2001) and contain information about the history of input and output patterns. The outputs of these internal PEs (echo states) are fed to a memoryless but adaptive readout network (generally linear) that produces the network out- put. The interesting property of ESN is that only the memoryless readout is trained, whereas the recurrent topology has fixed connection weights. This reduces the complexity of RNN training to simple linear regression while preserving a recurrent topology, but obviously places important constraints in the overall architecture that have not yet been fully studied. Similar ideas have been explored independently by Maass and formalized in the LSM architecture. LSMs, although formulated quite generally, are mostly im- plemented as neural microcircuits of spiking neurons (Maass et al., 2002), whereas ESNs are dynamical ANN models. Both attempt to model biolog- ical information processing using similar principles. We focus on the ESN formulation in this letter. The echo state condition is defined in terms of the spectral radius (the largest among the absolute values of the eigenvalues of a matrix, denoted by·) of the reservoir’s weight matrix (W<1). This condition states that the dynamics of the ESN is uniquely controlled by the input, and the effect of the initial states vanishes. The current design of ESN parameters relies on the selection of spectral radius. However, there are many possible weight matrices with the same spectral radius, and unfortunately they do not all perform at the same level of mean square error (MSE) for functional approximation. A similar problem exists in the design of the LSM. LSMs have been shown to possess universal approximation given the separation property (SP) for the liquid (reservoir in ESNs) and the approximation property (AP) for the readout (Maass et al., 2002). SP is quantified by a kernel-quality measure proposed in Maass, Legenstein, and Bertschinger (2005) that is based on the rank of a matrix formed by the system states corresponding to different input signals. The kernel quality is a measure for the complexity and diversity of nonlinear operations carried out by the liquid on its input stream in order to boost the classification power of a subsequent linear decision hyperplane (Maass et al., 2005). A variation of SP has been proposed in Bertschinger and Natschlager (2004), and it has¨ been argued that complex calculations can be best carried out by networks on the boundary between ordered and chaotic dynamics. In this letter,we are interested in studying the ESN for functional approx- imation (filters that map input function su(·) of time on output function sy(·) of time). We see two major shortcomings with the current ESN approach that uses echo state condition as a design principle. First, the impact of fixed reservoir parameters for function approximation means that the informa- tion about the desired response is conveyed only to the output projection. This is not optimal, and strategies to select different reservoirs for different applications have not been devised. Second, imposing a constraint only on the spectral radius is a weak condition to properly set the parameters of the reservoir, as experiments show (different randomizations with the same spectral radius perform differently for the same problem; see Figure 2). This letter aims to address these two problems by proposing a frame- work, a metric, and a design principle for ESNs. The framework is a signal processing interpretation of basis and projections in functional spaces to describe and understand the ESN architecture. According to this interpre- tation, the ESN states implement a set of basis functionals (representation space) constructed dynamically by the input, while the readout simply projects the desired response onto this representation space. The metric to describe the richness of the ESN dynamics is an information-theoretic quantity, the average state entropy (ASE). Entropy measures the amount of information contained in a given random variable (Shannon, 1948). Here, the random variable is the instantaneous echo state from which the en- tropy for the overall state (vector) is estimated. The probability density function (pdf) in a differential geometric framework should be thought of as a volume form; that is, in our case, the pdf of the state vector describes the metric of the state space manifold (Amari, 1990). Moreover, Cox (1946) established information as a coordinate free metric in the state manifold. Therefore, entropy becomes a global descriptor of information that quanti- fies the volume of the manifold defined by the random variable. Due to the time dependency of the states, the state entropy averaged over time (ASE) is an appropriate estimate of the volume of the state manifold. The design principle specifies that one should consider independently thecorrelationamongthebasisandthespectralradius.In the absence of any information about the desired response, the ESN states should be designed with the highest ASE, independent of the spectral radius. We interpret the ESN dynamics as a combination of time-varying linear systems obtained from the linearization of the ESN nonlinear PE in a small, local neighbor- hood of the current state. The design principle means that the poles of the linearized ESN reservoir should have uniform pole distributions to gener- ate echo states with the most diverse pole locations (which correspond to the uniformity of time constants). Effectively, this will create the least cor- related bases for a given spectral radius, which corresponds to the largest volume spanned by the basis set. When the designer has no other informa- tion about the desired response to set the basis, this principle distributes the system’s degrees of freedom uniformly in space. It approximates for ESNs the well-known property of orthogonal basis. The unresolved issue that ASE does not quantify is how to set the spectral radius, which depends again on the desired mapping. The concept of memory depth as explained in Principe et al. (1993) and Jaeger (2002a) is helpful in understanding the issues associated with the spectral radius. The correlation time of the de- sired response (as estimated by the first zero of the autocorrelation function) gives an indication of the type of spectral radius required (long correlation time requires high spectral radius). Alternatively, a simple adaptive bias is added at the ESN input to control the spectral radius integrating the infor- mation from the input-output joint space in the ESN bases. For sigmoidal PEs, the bias adjusts the operating points of the reservoir PEs, which has the net effect of adjusting the volume of the state manifold as required to approximate the desired response with a small error. This letter shows that ESNs designed with this strategy obtain systematically better results in a set of experiments when compared with the conventional ESN design. 2 Analysis of Echo State Networks 2.1 Echo States as Bases and Projections.Let us consider the ar- chitecture and recursive update equation of a typical ESN more closely. Consider the recurrent discrete-time neural network given in Figure 1 with M input units, N internal PEs, and L output units. The value of the input unit at time n is <> , of internal units are <> , and of output units are <> . The connection weights are given in anN×M weight matrixWin =(win ) for connections between the input and the inter- ij nalPEs,in an N×N matrix W=(wij ) for connections between the internal PEs, in an L×N matrix <> for connections from PEs to the ij Input Layer Dynamical Reservoir Read-out <

> Figure 1: An echo state network (ESN). ESN is composed of two parts: a fixed- weight (W<1) recurrent network and a linear readout. The recurrent net- work is a reservoir of highly interconnected dynamical components, states of which are called echo states. The memoryless linear readout is trained to pro- duce the output. output units, and in an N× L matrix <> for the connections ij that project back from the output to the internal PEs (Jaeger, 2001). The activation of the internal PEs (echo state) is updated according to <>, (2.1) where f=(f1 ,f2 ,...,fN ) are the internal PEs’ activation functions.Here, all i ’s are hyperbolic tangent functions ( ex − ). The output from the readout ex +e−x network is computed according to <>, (2.2) where <> are the output unit’s nonlinear functions <> (Jaeger, 2001, 2002a). Generally, the readout is linear so f_out is identity. ESNs resemble the RNN architecture proposed in Puskorius and Feldkamp (1996) and also used by Sanchez (2004) in brain-machine interfaces. The critical difference is the dimensionality of the hidden re- current PE layer and the adaptation of the recurrent weights. We submit that the ideas of approximation theory in functional spaces (bases and pro- jections), so useful in adaptive signal processing (Principe, 2001), should be utilized to understand the ESN architecture. Let h(u(t)) be a real-valued function of a real-valued vector <>. In functional approximation, the goal is to estimate the behavior ofh(u(t)) as a combination of simpler functions ϕi (t), called the basis functionals, such that its approximant,hˆ(u(t)), is given by <>. Here,ai ’s are the projections ofh(u(t)) onto each basis function. One of the central questions in practical functional approximation is how to choose the set of bases to approximate a given desired signal. In signal processing, thechoicenormallygoesforacompletesetoforthogonalbasis,independent of the input. When the basis set is complete and can be made as large as required, fixed bases work wonders (e.g., Fourier decompositions). In neural computing, the basic idea is to derive the set of bases from the input signal through a multilayered architecture. For instance, consider a single hidden layer TDNN with NPEs and a linear output. The hidden- layer PE outputs can be considered a set of nonorthogonal basis functionals dependent on the input, <> bij ’s are the input layer weights, andgis the PE nonlinearity. The approxi- mation produced by the TDNN is then <>, (2.3) whereai ’s are the weights of the output layer. Notice that thebij ’s adapt the bases and theai ’s adapt the projection in the projection space. Here the goal is to restrict the number of bases (number of hidden layer PEs) because their number is coupled with the number of parameters to adapt, which has an impact on generalization and training set size, for example. Usually, since all of the parameters of the network are adapted, the best basis in the joint (input and desired signals) space as well as the best projection can be achieved and represents the optimal solution. The output of the TDNN is a linear combination of its internal representations, but to achieve a basis set (even if nonorthogonal), linear independence among theϕi (u(t))’s must be enforced. Ito, Shah and Pon, and others have shown that this is indeed the case (Ito, 1996; Shah & Poon, 1999), but a thorough discussion is outside the scope of this article. The ESN (and the RNN) architecture can also be studied in this frame- work. The states of equation 2.1 correspond to the basis set, which are recursively computed from the input, output, and previous states through Win ,W,andWback . Notice, however, that none of these weight matrices is adapted, that is, the functional bases in the ESN are uniquely defined by the input and the initial selection of weights. In a sense, ESNs are trading the adaptive connections in the RNN hidden layer by a brute force approach of creating fixed diversified dynamics in the hidden layer. For an ESN with a linear readout network, the output equation (y(n+ 1)=Wout x(n+1)) has the same form of equation 2.3, where theϕi ’s and ai ’s are replaced by the echo states and the readout weights, respectively. The readout weights are adapted in the training data, which means that the ESN is able to find the optimal projection in the projection space, just like the RNN or the TDNN. A similar perspective of basis and projections for information processing in biological networks has been proposed by Pouget and Sejnowski (1997). They explored the possibility that the response of neurons in parietal cortex serves as basis functions for the transformations from the sensory input to the motor responses. They proposed that “the role of spatial represen- tations is to code the sensory inputs and posture signals in a format that simplifies subsequent computation, particularly in the generation of motor commands”. The central issue in ESN design is exactly the nonadaptive nature of the basis set. Parameter sets in the reservoir that provide linearly inde- pendent states and possess a given spectral radius may define drastically different projection spaces because the correlation among the bases is not constrained. A simple experiment was designed to demonstrate that the se- lection of the ESN parameters by constraining the spectral radius is not the most suitable for function approximation. Consider a 100-unit ESN where the input signal is sin(2πn/10π). Mimicking Jaeger (2001), the goal is to let the ESN generate the seventh power of the input signal. Different realiza- tions of a randomly connected 100-unit ESN were constructed where the entries ofWare set to 0.4,−0.4, and 0 with probabilities of 0.025, 0.025, and 0.95, respectively. This corresponds to a spectral radius of 0.88. Input weights are set to+1or,−1 with equal probabilities, andWback is set to zero. Input is applied for 300 time steps, and the echo states are calculated using equation 2.1. The next step is to train the linear readout. One method <

> Figure 2: Performances of ESNs for different realizations ofWwith the same weight distribution. The weight values are set to 0.4,−0.4, and 0 with proba- bilities of 0.025, 0.025, and 0.95. All realizations have the same spectral radius of 0.88. In the 50 realizations, MSEs vary from 5.9×10 −9 to 8.9×10 −5 . Results show that for each set of random weights that provide the same spectral ra- dius, the correlation or degree of redundancy among the bases will change, and different performances are encountered in practice. to determine the optimal output weight matrix,Wout , in the mean square error (MSE) sense (where MSE is defined by <>) is to use 2 the Wiener solution given by Haykin (2001): <> Here,E[.] denotes the expected value operator, andddenotes the desired signal. Figure 2 depicts the MSE values for 50 different realizations of the ESNs. As observed, even though each ESN has the same sparseness and spectral radius, the MSE values obtained vary greatly among differ- ent realizations. The minimum MSE value obtained among the 50 realiza- tions is 5.9x10 −9 , whereas the maximum MSE is 8.9x10 −5 . This experiment demonstrates that a design strategy that is based solely on the spectral radius is not sufficient to specify the system architecture for function ap- proximation. This shows that for each set of random weights that provide thesamespectralradius,thecorrelationordegreeofredundancyamongthe bases will change, and different performances are encountered in practice. 2.2 ESN Dynamics as a Combination of Linear Systems. It is well known that the dynamics of a nonlinear system can be approximated by that of a linear system in a small neighborhood of an equilibrium point (Kuznetsov, Kuznetsov, & Marsden, 1998). Here, we perform the analysis with hyperbolic tangent nonlinearities and approximate the ESN dynam- ics by the dynamics of the linearized system in the neighborhood of the current system state. Hence, when the system operating point varies over time, the linear system approximating the ESN dynamics changes. We are particularly interested in the movement of the poles of the linearized ESN. Consider the update equation for the ESN without output feedback given by <>. Linearizing the system around the current statex(n), one obtains the Jacobian matrix, <>, defined by <> Here,net i(n) is the ith entry of the vector <<(W_in u(n+1)+Wx(n))>>, and w_ij denotes the (i,j)th entry of W. The poles of the linearized system at time n+1 are given by the eigenvalues of the Jacobian matrixJ(n+1). 1 As the amplitude of each PE changes, the local slope changes, and so the poles of A. The transfer function of a linear system <> is <> Adjoint <<(zI−A)>>. The poles of the transfer function can be obtained by solving <>. The solution corresponds to the eigenvalues of A. the linearized system are time varying, although the parameters of ESN are fixed. In order to visualize the movement of the poles, consider an ESN with 100 states. The entries of the internal weight matrix are chosen to be 0, 0.4 and −0.4 with probabilities 0.9, 0.05, and 0.05.W is scaled such that a spectral radius of 0.95 is obtained. Input weights are set to +1 or −1 with equal probabilities. A sinusoidal signal with a period of 100 is fed to the system, and the echo states are computed according to equation 2.1. Then the Jacobian matrix and the eigenvalues are calculated using equation 2.5. Figure 3 shows the pole tracks of the linearized ESN for different input values. A single ESN with fixed parameters implements a combination of many linear systems with varying pole locations, hence many different time constants that modulate the richness of the reservoir of dynamics as a function of input amplitude. Higher-amplitude portions of the signal tend to saturate the nonlinear function and cause the poles to shrink toward the origin of thez-plane (decreases the spectral radius), which results in a system with a large stability margin. When the input is close to zero, the poles of the linearized ESN are close to the maximal spectral radius chosen, decreasing the stability margin. When compared to their linear counterpart, an ESN with the same number of states results in a detailed coverage of thez-plane dynamics, which illustrates the power of nonlinear systems. Similar results can be obtained using signals of different shapes at the ESN input. A key corollary of the above analysis is that the spectral radius of an ESN can be adjusted using a constant bias signal at the ESN input without changing the recurrent connection matrix,W. The application of a nonzero constant bias will move the operating point to regions of the sigmoid func- tion closer to saturation and always decrease the spectral radius due to the shape of the nonlinearity. 2 The relevance of bias in terms of overall system performance has also been discussed in Jaeger (2002b) and Bertschinger and Natschlager (2004), but here we approach it from a system theory per-¨ spective and explain its effect on reservoir dynamics. 3 Average State Entropy as a Measure of the Richness of ESN Reservoir Previous research was aware of the influence of diversity of the recurrent layer outputs on the overall performance of ESNs and LSMs. Several met- rics to quantify the diversity have been proposed (Jaeger, 2001; Maass, et al., 2 Assume W has nondegenerate eigenvalues and corresponding linearly independent eigenvectors. Then consider the eigendecomposition of W, where <>,Pis the eigenvectormatrixandDisthediagonalmatrixofeigenvalues <> of W.SinceF(n)andD are diagonal, <> is the eigendecomposition of <>. Here, each entry of <>, is an eigenvalue of J. Therefore, <> since <>. <

> Figure 3: The pole tracks of the linearized ESN with 100 PEs when the input goes through a cycle. An ESN with fixed parameters implements a combination of linear systems with varying pole locations. (A) One cycle of sinusoidal signal with a period of 100. (B–E) The positions of poles of the linearized systems when the input values are at B, C, D, and E in Figure 5A. (F) The cumulative pole locations show the movement of the poles as the input changes. Due to the varying pole locations, different time constants modulate the richness of the reservoir of dynamics as a function of input amplitude. Higher-amplitude signals tend to saturate the nonlinear function and cause the poles to shrink toward the origin of thez-plane (decreases the spectral radius), which results in a system with a large stability margin. When the input is close to zero, the poles ofthelinearizedESNareclosetothemaximalspectralradiuschosen,decreasing the stability margin. An ESN with more states results in a detailed coverage of thez-plane dynamics, which illustrates the power of nonlinear systems, when compared to their linear counterpart. Here, our approach of bases and projections leads to a new metric. We propose the instantaneous state entropy to quantify the distribution of instantaneous amplitudes across the ESN states. Entropy of the instanta- neous ESN states is appropriate to quantify performance in function ap- proximation because the ESN output is a mere weighted combination of the instantaneous value of the ESN states. If the echo state’s instantaneous amplitudes are concentrated on only a few values across the ESN state dy- namic range, the ability to approximate an arbitrary desired response by weighting the states is limited (and wasteful due to redundancy between the different states), and performance will suffer. On the other hand, if the ESN states provide a diversity of instantaneous amplitudes, it is much eas- ier to achieve the desired mapping. Hence, the instantaneous entropy of the states appears as a good measure to quantify the richness of dynamics with instantaneous mappers. Due to the time structure of signals, the average state entropy (ASE), defined as the state entropy averaged over time, will be the parameter used to quantify the diversity in the dynamical reservoir of the ESN. Moreover, entropy has been proposed as an appropriate measure of the volume of the signal manifold (Cox, 1946; Amari, 1990). Here, ASE measures the volume of the echo state manifold spanned by trajectories. Renyi’squadraticentropyisemployedherebecauseitisaglobalmeasure of information. In addition, an efficient nonparametric estimator of Renyi’s entropy,whichavoidsexplicitpdfestimation,hasbeendeveloped(Principe, Xu, & Fisher, 2000). Renyi’s entropy with parameterγfor a random variable X with a <> is given by Renyi (1970): <> Renyi’s quadratic entropy is obtained forγ=2 (forγ→1, Shannon’s en- tropy is obtained). GivenNsamples{x1 ,x2 ,...,xN }drawn from the un- known pdf to be estimated, Parzen windowing approximates the underly- ing pdf by <> whereKσ is the kernel function with the kernel sizeσ. Then the Renyi’s quadratic entropy can be estimated by (Principe et al., 2000) <> The instantaneous state entropy is estimated using equation 3.1 where thesamplesaretheentriesofthestatevectorx(n)=[x1 (n),x2 (n),...,xN (n)] T of an ESN withNinternal PEs. Results will be shown with a gaussian kernel with kernel size chosen to be 0.3 of the standard deviation of the entries of the state vector. We will show that ASE is a more sensitive parameter to quantify the approximation properties of ESNs by experimentally demon- strating that ESNs with different spectral radius and even with the same spectral radius display different ASEs. Let us consider the same 100-unit ESN that we used in the previous section built with three different spectral radii 0.2, 0.5, 0.8 with an input signal of sin(2πn/20). Figure 4A depicts the echo states over 200 time ticks. The instantaneous state entropy is also calculated at each time step using equation 3.1 and plotted in Figure 4B. First, note that the instantaneous state entropy changes over time with the distribution of the echo states as we would expect, since state entropy is dependent on the input signal that also changes in this case. Second, as the spectral radius increases in the simulation, the diversity in the echo states increases. For the spectral radius of 0.2, echo state’s instantaneous amplitudes are concentrated on only a few values, which is wasteful due to redundancy between different states. In practice, to quantify the overall representation ability over time, we will use ASE, which takes values−0.735,−0.007, and 0.335 for the spectral radii of 0.2, 0.5, and 0.8, respectively. Moreover, even for the same spectral radius, several ASEs are possible. Figure 4C shows ASEs from 50 different realizations of ESNs with the same spectral radius of 0.5, which means that ASE is a finer descriptor of the dynamics of the reservoir. Although we have presented an experiment with sinusoidal signal, similar results are obtained for other inputs as long as the input dynamic range is properly selected. Maximizing ASE means that the diversity of the states over time is the largest and should provide a basis set that is as uncorrelated as possible. This condition is unfortunately not a guarantee that the ESN so designed will perform the best, because the basis set in ESNs is created independent of the desired response and the application may require a small spectral radius. However, we maintain that when the desired response is not ac- cessible for the design of the ESN bases or when the same reservoir is to be used for a number of problems, the default strategy should be to maximize the ASE of the state vector. The following section addresses the design of ESNs with high ASE values and a simple mechanism to adjust the reservoir dynamics without changing the recurrent connection weights. 4 Designing Echo State Networks 4.1 Design of the Echo State Recurrent Connections.According to the interpretation of ESNs as coupled linear systems, the design of the internal connection matrix, W, will be based on the distribution of the poles of the linearized system around zero state. Our proposal is to design the ESN such that the linearized system has uniform pole distribution inside the unit circle of thez-plane. With this design scenario, the system dynamics will include uniform coverage of time constants arising from the uniform distribution of the poles, which also decorrelates as much as possible the basis functionals. This principle was chosen by analogy to the identification oflinearsystemsusingKautzfilters(Kautz,1954),whichshowsthatthebest approximation of a given transfer function by a linear system with finite order is achieved when poles are placed in the neighborhood of the spectral resonances. When no information is available about the desired response, we should uniformly spread the poles to anticipate good approximation to arbitrary mappings. We again use a maximum entropy principle to distribute the poles inside the unit circle uniformly. The constraints of a circle as boundary conditions for discrete linear systems and complex conjugate locations are easy to include for the pole distribution (Thogula, 2003). The poles are first initial- ized at random locations; the quadratic Renyi’s entropy is calculated by equation 3.1, and poles are moved such that the entropy of the new dis- tribution is increased over iterations (Erdogmus & Principe, 2002). This method is efficient to find uniform coverage of the unit circle with an arbi- trary number of poles. The system with the uniform pole locations can be interpreted using linear system theory. The poles that are close to the unit circle correspond to many sharp bandpass filters specializing in different frequency regions, whereas the inner poles realize filters of larger frequency support. Moreover, different orientations (angles) of the poles create filters of different center frequencies. Now the problem is to construct an internal weight matrix from the pole locations (eigenvalues ofW). In principle, we would like to create a sparse <

> Figure 4: Examples of echo states and instantaneous state entropy. (A) Outputs ofechostates(100PEs)producedbyESNswithspectralradiusof0.2,0.5,and0.8, from top to bottom, respectively. The diversity of echo states increases when the spectral radius increases. Within the dynamic range of the echo states, systems with smaller spectral radius can generate only uneven representations, while forW=0.8, outputs of echo states almost uniformly distribute within their dynamic range. (B) Instantaneous state entropy is calculated using equation 3.1. Information contained in the echo states is changing over time according to the input amplitude. Therefore, the richness of representation is controlled by the input amplitude. Moreover, the value of ASE increases with spectral radius. (C) ASEs from 50 different realizations of ESNs with the same spectral radius of 0.5. The plot shows that ASE is a finer descriptor of the dynamics of the reservoir than the spectral radius. matrix, so we started with the sparsest matrix (with an inverse), which is the direct canonical structure given by (Kailath, 1980) <> The characteristic polynomial of W_i's <>, (4.2) wherepi ’s are the eigenvalues andai ’s are the coefficients of the character- istic polynomial ofW. Here, we know the pole locations of the linear system obtained from the linearization of the ESN, so using equation 4.2, we can obtain the characteristic polynomial and constructWmatrix in the canon- ical form using equation 4.1. We will call the ESN constructed based on the uniform pole principle ASE-ESN. All other possible solutions with the same eigenvalues can be obtained byQ−1 WQ,whereQis any nonsingular matrix. To corroborate our hypothesis, we would like to show that the linearized ESN designed with the recurrent weight matrix having the eigenvalues uniformly distributed inside the unit circle creates higher ASE values for a given spectral radius compared to other ESNs with random internal con- nection weight matrices. We will consider an ESN with 30 states and use our procedure to create theWmatrix for ASE-ESN for different spectral radii between <<[0.1, 0.95]>>. Similarly, we constructed ESNs with sparse randomW matrices with different sparseness constraints. This corresponds to a weight distribution having the values 0, c and −c with probabilities <> ,<<(1−p_1)/2>>, and <<(1−p_1)/2>>, wherep1 defines the sparseness ofWandcis a constant that takes a specific value depending on the spectral radius. We also created Wmatrices with values uniformly distributed between−1 and 1 (U-ESN) and scaled to obtain a given spectral radius (Jaeger & Hass, 2004). Then, for differentWin matrices, we run the ASE-ESNs with the sinusoidal input given in section 3 and calculate ASE. Figure 5 compares the ASE values averaged over 1000 realizations. As observed from the figure, the ASE-ESN with uniform pole distribution generates higher ASE on average for all spectral radii compared to ESNs with sparse and uniform random connec- tions. This approach is indeed conceptually similar to Jeffreys’ maximum entropy prior (Jeffreys, 1946): it will provide a consistently good response for the largest class of problems. Concentrating the poles of the linearized <

> Figure 5: Comparison of ASE values obtained for ASE-ESN havingWwith uniform eigenvalue distribution, ESNs with randomWmatrix, and U-ESN with uniformly distributed weights between−1 and 1. Randomly generated weights have sparseness of 0.07, 0.1, and 0.2. ASE values are calculated for the networks with spectral radius from 0.1 to 0.95. The ASE-ESN with uniform pole distribution generates a higher ASE on average for all spectral radii compared to ESNs with random connections. system in certain regions of the space provides good performance only if the desired response has energy in this part of the space, as is well known from the theory of Kautz filters (Kautz, 1954). 4.2 Design of the Adaptive Bias. In conventional ESNs, only the output weights are trained, optimizing the projections of the desired response onto the basis functions (echo states). Since the dynamical reservoir is fixed, the basis functions are only input dependent. However, since function ap- proximation is a problem in the joint space of the input and desired signals, a penalty in performance will be incurred. From the linearization analysis that shows the crucial importance of the operating point of the PE non- linearity in defining the echo state dynamics, we propose to use a single external adaptive bias to adjust the effective spectral radius of an ESN. No- tice that according to linearization analysis, bias can reduce only spectral radius. The information for adaptation of bias is the MSE in training, which modulates the spectral radius of the system with the information derived from the approximation error. With this simple mechanism, some informa- tionfromtheinput-outputjointspaceisincorporatedinthedefinitionofthe projection space of the ESN. The beauty of this method is that the spectral radius can be adjusted by a single parameter that is external to the system without changing reservoir weights. The training of bias can be easily accomplished. Indeed, since the pa- rameter space is only one-dimensional, a simple line search method can be efficiently employed to optimize the bias. Among different line search al- gorithms, we will use a search that uses Fibonacci numbers in the selection of points to be evaluated (Wilde, 1964). The Fibonacci search method min- imizes the maximum number of evaluations needed to reduce the interval of uncertainty to within the prescribed length. In our problem, a bias value is picked according to Fibonacci search. For each value of bias, training data are applied to the ESN, and the echo states are calculated. Then the corresponding optimal output weights and the objective function (MSE) are evaluated to pick the next bias value. Alternatively, gradient-based methods can be utilized to optimize the bias, due to simplicity and low computational cost. System update equation with an external bias signal,b,isgivenby <>. The update equation forbis given by <> Here,Ois the MSE defined previously. This algorithm may suffer from similar problems observed in gradient-based methods in recurrent net- works training. However, we observed that the performance surface is rather simple. Moreover, since the search parameter is one-dimensional, the gradient vector can assume only one of the two directions. Hence, im- precision in the gradient estimation should affect the speed of convergence but normally not change the correct gradient direction. 5 Experiments This section presents a variety of experiments in order to test the validity of the ESN design scheme proposed in the previous section. 5.1 Short-Term Memory Capacity. This experiment compares the shortterm memory (STM) capacity of ESNs with the same spectral radius using the framework presented in Jaeger (2002a). Consider an ESN with a sin- gle input signal, <>, optimally trained with the desired signal <>, for a given delayk. Denoting the optimal output signalyk (n), thek-delay STM capacity of a network,MC k , is defined as a squared correlation coef- ficient betweenu <<(n−k)>> and <> (Jaeger, 2002a). The STM capacity, MC, of the network is defined as <>. STM capacity measures how accu- rately the delayed versions of the input signal are recovered with optimally trained output units. Jaeger (2002a) has shown that the memory capacity for recalling an independent and identically distributed (i.i.d.) input by an Nunit RNN with linear output units is bounded by N. We use ESNs with 20 PEs and a single input unit. ESNs are driven by an i.i.d. random input signal,<>, that is uniformly distributed over [−0.5, 0.5]. The goal is to train the ESN to generate the delayed versions of the input, <>. We used four different ESNs: R-ESN, U-ESN, ASE-ESN, and BASE-ESN. R-ESN is a randomly connected ESN used in Jaeger (2002a) where the entries ofWmatrix are set to 0, 0.47, −0.47 with probabilities 0.8, 0.1, 0.1, respectively. This corresponds to a sparse connectivity of 20% and a spectral radius of 0.9. The entries ofWof U-ESN are uniformly distributed over [−1, 1] and scaled to obtain the spec- tral radius of 0.9. ASE-ESN also has a spectral radius of 0.9 and is designed with uniform poles. BASE-ESN has the same recurrent weight matrix as ASE-ESN and an adaptive bias at its input. In each ESN, the input weights are set to 0.1 or−0.1 with equal probability, and direct connections from the input to the output are allowed, whereasWback is set to 0 (Jaeger, 2002a). The echo states are calculated using equation 2.1 for 200 samples of the input signal, and the first 100 samples corresponding to initial transient are eliminated. Then the output weight matrix is calculated using equation 2.4. For the BASE-ESN, the bias is trained for each task. All networks are run with a test input signal, and the corresponding output andMC k are calculated. Figure 6 shows thek-delay STM capacity (averaged over 100 trials) of each ESN for delays 1,...,40 for the test signal. The STM capac- ities of R-ESN, U-ESN, ASE-ESN, and BASE-ESN are 13.09, 13.55, 16.70, and 16.90, respectively. First, ESNs with uniform pole distribution (ASE- ESN and BASE-ESN) haveMCs that are much longer than the randomly generated ESN given in Jaeger (2002a) in spite of all having the same spec- tral radius. In fact, the STM capacity of ASE-ESN is close to the theoretical maximumvalueofN=20.AcloserlookatthefigureshowsthatR-ESNper- forms slightly better than ASE-ESN for delays less than 9. In fact, for small k, large ASE degrades the performance because the tasks do not need long memory depth. However, the drawback of high ASE for smallkis recov- ered in BASE-ESN, which reduces the ASE to the appropriate level required for the task. Overall, the addition of the bias to the ASE-ESN increases the STM capacity from 16.70 to 16.90. On the other hand, U-ESN has slightly better STM compared to R-ESN with only three different weight values, although it has more distinct weight values compared to R-ESN. It is also significant to note that theMCwill be very poor for an ESN with smaller spectral radius even with an adaptive bias, since the problem requires large ASE and bias can only reduce ASE. This experiment demonstrates the <

> Figure 6: Thek-delay STM capacity of each ESN for delays 1,...,40 computed using the test signal. The results are averaged over 100 different realizations of each ESN type with the specifications given in the text for differentWandWin matrices. The STM capacities of R-ESN, U-ESN, ASE-ESN, and BASE-ESN are 13.09, 13.55, 16.70, and 16.90, respectively. suitability of maximizing ASE in tasks that require a substantial memory length. 5.2 Binary Parity Check. The effect of the adaptive bias was marginal in the previous experiment since the nature of the problem required large ASE values. However, there are tasks in which the optimal solutions re- quire smaller ASE values and smaller spectral radius. Those are the tasks where the adaptive bias becomes a crucial design parameter in our design methodology. Consider an ESN with 100 internal units and a single input unit. ESN is drivenbyabinaryinputsignal,u(n),thatassumesthevalues0or1.Thegoal is to train an ESN to generate them-bit parity corresponding to lastmbits received, wheremis 3,...,8. Similar to the previous experiments, we used the R-ESN, ASE-ESN, and BASE-ESN topologies. R-ESN is a randomly connected ESN where the entries ofWmatrix are set to 0, 0.06,−0.06 with probabilities 0.8, 0.1, 0.1, respectively. This corresponds to a sparse connectivity of 20% and a spectral radius of 0.3. ASE-ESN and BASE-ESN are designed with a spectral radius of 0.9. The input weights are set to 1 or -1 with equal probability, and direct connections from the input to the output are allowed whereasWback is set to 0. The echo states are calculated using equation 2.1 for 1000 samples of the input signal, and the first 100 samples corresponding to the initial transient are eliminated.Then the output weight <

> Figure 7: The number of wrong decisions made by each ESN form=3,...,8 in the binary parity check problem. The results are averaged over 100 differ- ent realizations of R-ESN, ASE-ESN, and BASE-ESN for differentWandWin matrices with the specifications given in the text. The total numbers of wrong decisions form=3,...,8 of R-ESN, ASE-ESN, and BASE-ESN are 722, 862, and 699. matrix is calculated using equation 2.4. For ESN with adaptive bias, the bias is trained for each task. The binary decision is made by a threshold detector that compares the output of the ESN to 0.5. Figure 7 shows the number of wrong decisions (averaged over 100 different realizations) made by each ESN for <>. The total numbers of wrong decisions for <> of R-ESN, ASE- ESN, and BASE-ESN are 722, 862, and 699, respectively. ASE-ESN performs poorly since the nature of the problem requires a short time constant for fast response, but ASE-ESN has a large spectral radius. For 5-bit parity, the R-ESN has no wrong decisions, whereas ASE-ESN has 53 wrong decisions. BASE-ESN performs a lot better than ASE-ESN and slightly better than the R-ESN since the adaptive bias reduces the spectral radius effectively. Note that form=7 and 8, the ASE-ESN performs similar to the R-ESN, since the task requires access to longer input history, which compromises the need for fast response. Indeed, the bias in the BASE-ESN takes effect when there are errors (m>4) and when the task benefits from smaller spectral radius. The optimal bias values are approximately 3.2, 2.8, 2.6, and 2.7 form=3, 4, 5, and 6, respectively. Form=7 or 8, there is a wide range of bias values that result in similar MSE values (between 0 and 3). In summary, this experiment clearly demonstrates the power of the bias signal to configure the ESN reservoir according to the mapping task. 5.3 System Identification. This section presents a function approxima- tion task where the aim is to identify a nonlinear dynamical system. The unknown system is defined by the difference equation <>, where <>. The input to the system is chosen to be <>. We used three different ESNs—R-ESN, ASE-ESN, and BASE-ESN—with 30 internal units and a single input unit. TheWmatrix of each ESN is scaled suchthatithasaspectralradiusof0.95.R-ESNisarandomlyconnectedESN where the entries ofWmatrix are set to 0, 0.35,−0.35 with probabilities 0.8, 0.1, 0.1, respectively. In each ESN, the input weights are set to 1 or−1 with equal probability, and direct connections from the input to the output are allowed,whereasWback issetto0.Theoptimaloutputweightsarecalculated using equation 2.4. The MSE values (averaged over 100 realizations) for R- ESN and ASE-ESN are 1.23x10 −5 and 1.83x10 −6 , respectively. The addition of the adaptive bias to the ASE-ESN reduces the MSE value from 1.83x10^−6 to 3.27x10^−9 . 6 Discussion The great appeal of echo state networks (ESNs) and liquid state machine (LSM) is their ability to construct arbitrary mappings of signals with rich and time-varying temporal structures without requiring adaptation of the free parameters of the recurrent layer. The echo state condition allows the recurrent connections to be fixed with training limited to the linear output layer. However, the literature did not elucidate on how to properly choose the recurrent parameters for system identification applications. Here, we provide an alternate framework that interprets the echo states as a set of functional bases formed by fixed nonlinear combinations of the input. The linear readout at the output stage simply computes the projection of the desired output space onto this representation space. We further in- troduce an information-theoretic criterion, ASE, to better understand and evaluate the capability of a given ESN to construct such a representation layer. The average entropy of the distribution of the echo states quantifies thevolumespannedbythebases.Assuch,thisvolumeshouldbethelargest to achieve the smallest correlation among the bases and be able to cope with arbitrary mappings. However, not all function approximation problems re- quire the same memory depth, which is coupled to the spectral radius. The effective spectral radius of an ESN can be optimized for the given problem with the help of an external bias signal that is adapted using the joint input- output space information. The interesting property of this method when applied to ESN built from sigmoidal nonlinearities is that it allows the fine tuning of the system dynamics for a given problem with a single external adaptive bias input and without changing internal system parameters. In our opinion, the combination of the largest possible ASE and the adapta- tion of the spectral radius by the bias produces the most parsimonious pole location of the linearized ESN when no knowledge about the mapping is available to optimally locate the bass functionals. Moreover, the bias can be easily trained with either a line search method or a gradient-based method since it is one-dimensional. We have illustrated experimentally that the de- sign of the ESN using the maximization of ASE with the adaptation of the spectral radius by the bias has provided consistently better performance across tasks that require different memory depths. This means that these two parameters’ design methodology is preferred to the spectral radius criterion proposed by Jaeger, and it is still easily incorporated in the ESN design. Experiments demonstrate that the ASE for ESN with uniform linearized poles is maximized when the spectral radius of the recurrent weight matrix approaches one (instability). It is interesting to relate this observation with the computational properties found in dynamical systems “at the edge of chaos” (Packard, 1988; Langton, 1990; Mitchell, Hraber, & Crutchfield, 1993; Bertschinger & Natschlager, 2004). Langton stated that when cellular au-¨ tomata rules are evolved to perform a complex computation, evolution will tend to select rules with “critical” parameter values, which correlate with a phase transition between ordered and chaotic regimes. Recently, similar conclusions were suggested for LSMs (Bertschinger & Natschlager, 2004).¨ Langton’s interpretation of edge of chaos was questioned by Mitchell et al. (1993). Here, we provide a system-theoretic view and explain the computa- tional behavior with the diversity of dynamics achieved with linearizations that have poles close to the unit circle. According to our results, the spectral radiusoftheoptimalESNinfunctionapproximationisproblemdependent, and in general it is impossible to forecast the computational performance as the system approaches instability (the spectral radius of the recurrent weight matrix approaches one). However, allowing the system to modu- late the spectral radius by either the output or internal biasing may allow a system close to instability to solve various problems requiring different spectral radii. Our emphasis here is mostly on ESNs without output feedback connec- tions. However, the proposed design methodology can also be applied to ESNs with output feedback. Both feedforward and feedback connections contribute to specify the bases to create the projection space. At the same time, there are applications where the output feedback contributes to the system dynamics in a different fashion. For example, it has been shown that a fixed weight (fully trained) RNN with output feedback can implement a family of functions (meta-learners) (Prokhorov, Feldkamp, & Tyukin, 1992). In meta-learning, the role of output feedback in the network is to bias the system to different regions of dynamics, providing multiple input-output mappings required (Santiago & Lendaris, 2004). However, results could not be replicated with ESNs (Prokhorov, 2005). We believe that more work has to be done on output feedback in the context of ESNs but also suspect that the echo state condition may be a restriction on the system dynamics for this type of problem. There are many interesting issues to be researched in this exciting new area. Besides an evaluation tool, ASE may also be utilized to train the ESN’s representation layer in an unsupervised fashion. In fact, we can easily adapt withtheSIG(stochasticinformationgradient)describedinErdogmus,Hild, and Principe (2003): extra weights linking the outputs of recurrent states to maximize output entropy. Output entropy maximization is a well-known metric to create independent components (Bell & Sejnowski, 1995), and here it means that the echo states will become as independent as possible. This would circumvent the linearization of the dynamical system to set the recurrent weights and would fine-tune continuously in an unsupervised manner the parameters of the ESN among different inputs. However, it goes against the idea of a fixed ESN reservoir. The reservoir of recurrent PEs can be thought of as a new form of a time- to-space mapping. Unlike the delay line that forms an embedding (Takens, 1981), this mapping may have the advantage of filtering noise and produce representations with better SNRs to the peaks of the input, which is very appealing for signal processing and seems to be used in biology. However, further theoretical work is necessary in order to understand the embedding capabilities of ESNs. One of the disadvantages of the ESN correlated basis is in the design of the readout. Gradient-based algorithms will be very slow to converge (due to the large eigenvalue spread of modes), and even if recursive methods are used, their stability may be compromised by the condition number of the matrix. However, our recent results incorporating anL1 norm penalty in the LMS (Rao et al., 2005) show great promise of solving this problem. Finally we would like to briefly comment on the implications of these models to neurobiology and computational neuroscience. The work by Pouget and Sejnowski (1997) has shown that the available physiological data are consistent with the hypothesis that the response of a single neuron in the parietal cortex serves as a basis function generated by the sensory input in a nonlinear fashion. In other words, the neurons transform the sensory input into a format (representation space) such that the subsequent computation is simplified. Then, whenever a motor command (output of the biological system) needs to be generated, this simple computation to read out the neuronal activity is done. There is an intriguing similarity betweentheinterpretationoftheneuronalactivitybyPougetandSejnowski and our interpretation of echo states in ESN. We believe that similar ideas can be applied to improve the design of microcircuit implementations of LSMs. First, the framework of functional space interpretation (bases and projections) is also applicable to microcircuits. Second, the ASE measure may be directly utilized for LSM states because the states are normally low- pass-filtered before the readout. However, the control of ASE by changing the liquid dynamics is unclear. Perhaps global control of thresholds or bias current will be able to accomplish bias control as in ESN with sigmoid PEs. Acknowledgments This work was partially supported by NSFECS-0422718, NSFCNS-0540304, and ONR N00014-1-1-0405. References Amari, S.-I. (1990).Differential-geometrical methods in statistics.NewYork:Springer. Anderson, J., Silverstein, J., Ritz, S., & Jones, R. (1977). Distinctive features, categor- ical perception, and probability learning: Some applications of a neural model. Psychological Review, 84, 413–451. Bell, A. J., & Sejnowski, T. J. (1995). An information-maximization approach to blind separation and blind deconvolution.Neural Computation, 7(6), 1129– 1159. Bertschinger,N.,&Natschlager,T.(2004).Real-timecomputationattheedgeofchaos¨ in recurrent neural networks.Neural Computation, 16(7), 1413–1436. Cox,R.T.(1946).Probability,frequency,andreasonableexpectation.AmericanJournal of Physics, 14(1), 1–13. de Vries, B. (1991).Temporal processing with neural networks—the development of the gamma model. Unpublished doctoral dissertation, University of Florida. Delgado, A., Kambhampati, C., & Warwick, K. (1995). Dynamic recurrent neural network for system identification and control.IEEE Proceedings of Control Theory and Applications, 142(4), 307–314. Elman, J. L. (1990). Finding structure in time.Cognitive Science, 14(2), 179–211. Erdogmus, D., Hild, K. E., & Principe, J. (2003). Online entropy manipulation: Stochastic information gradient.Signal Processing Letters, 10(8), 242–245. Erdogmus, D., & Principe, J. (2002). Generalized information potential criterion for adaptive system training.IEEE Transactions on Neural Networks, 13(5), 1035–1044. Feldkamp,L.A.,Prokhorov,D.V.,Eagen,C.,&Yuan,F.(1998).Enhancedmultistream Kalman filter training for recurrent networks. In J. Suykens, & J. Vandewalle (Eds.),Nonlinear modeling: Advanced black-box techniques(pp. 29–53). Dordrecht, Netherlands: Kluwer. 136 M. Ozturk, D. Xu, and J. Pr´ıncipe Haykin,S.(1998).Neuralnetworks:Acomprehensivefoundation(2nded.).UpperSaddle River, NJ. Prentice Hall. Haykin, S. (2001).Adaptive filter theory(4th ed.). Upper Saddle River, NJ: Prentice Hall. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory.Neural Computa- tion, 9(8), 1735–1780. Hopfield, J. (1984). Neurons with graded response have collective computational properties like those of two-state neurons.Proceedings of the National Academy of Sciences, 81, 3088–3092. Ito, Y. (1996). Nonlinearity creates linear independence.Advances in Computer Math- ematics, 5(1), 189–203. Jaeger, H. (2001).The echo state approach to analyzing and training recurrent neural networks(Tech. Rep. No. 148). Bremen: German National Research Center for Information Technology. Jaeger, H. (2002a).Short term memory in echo state networks(Tech. Rep. No. 152). Bremen: German National Research Center for Information Technology. Jaeger, H. (2002b).Tutorial on training recurrent neural networks, covering BPPT, RTRL, EKF and the “echo state network” approach(Tech. Rep. No. 159). Bremen: German National Research Center for Information Technology. Jaeger, H., & Hass, H. (2004). Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication.Science, 304(5667), 78–80. Jeffreys,H.(1946).Aninvariantformforthepriorprobabilityinestimationproblems. Proceedings of the Royal Society of London, A 196, 453–461. Kailath, T. (1980).Linear systems. Upper Saddle River, NJ: Prentice Hall. Kautz, W. (1954). Transient synthesis in time domain.IRE Transactions on Circuit Theory, 1(3), 29–39. Kechriotis,G.,Zervas,E.,&Manolakos,E.S.(1994). Usingrecurrentneuralnetworks for adaptive communication channel equalization.IEEE Transactions on Neural Networks, 5(2), 267–278. Kremer,S.C.(1995).OnthecomputationalpowerofElman-stylerecurrentnetworks. IEEE Transactions on Neural Networks, 6(5), 1000–1004. Kuznetsov, Y., Kuznetsov, L., & Marsden, J. (1998).Elements of applied bifurcation theory(2nd ed.). New York: Springer-Verlag. Langton, C. G. (1990). Computation at the edge of chaos.Physica D, 42, 12–37. Maass, W., Legenstein, R. A., & Bertschinger, N. (2005). Methods for estimating the computational power and generalization capability of neural microcircuits. In L. K. Saul, Y. Weiss, L. Bottou (Eds.),Advances in neural information processing systems, no. 17 (pp. 865–872). Cambridge, MA: MIT Press. Maass, W., Natschlager, T., & Markram, H. (2002). Real-time computing without¨ stable states: A new framework for neural computation based on perturbations. Neural Computation, 14(11), 2531–2560. Mitchell, M., Hraber, P., & Crutchfield, J. (1993). Revisiting the edge of chaos: Evolving cellular automata to perform computations.Complex Systems, 7, 89– 130. Packard, N. (1988). Adaptation towards the edge of chaos. In J. A. S. Kelso, A. J. Mandell, & M. F. Shlesinger (Eds.),Dynamic patterns in complex systems(pp. 293– 301). Singapore: World Scientific. Analysis and Design of Echo State Networks 137 Pouget, A., & Sejnowski, T. J. (1997). Spatial transformations in the parietal cortex using basis functions.Journal of Cognitive Neuroscience, 9(2), 222–237. Principe, J. (2001). Dynamic neural networks and optimal signal processing. In Y. Hu & J. Hwang (Eds.),Neural networks for signal processing(Vol. 6-1, pp. 6– 28). Boca Raton, FL: CRC Press. Principe, J. C., de Vries, B., & de Oliviera, P. G. (1993). The gamma filter—a new class of adaptive IIR filters with restricted feedback.IEEE Transactions on Signal Processing, 41(2), 649–656. Principe, J., Xu, D., & Fisher, J. (2000). Information theoretic learning. In S. Haykin (Ed.),Unsupervised adaptive filtering(pp. 265–319). Hoboken, NJ: Wiley. Prokhorov, D. (2005). Echo state networks: Appeal and challenges. InProc. of Inter- national Joint Conference on Neural Networks(pp. 1463–1466). Montreal, Canada. Prokhorov, D., Feldkamp, L., & Tyukin, I. (1992). Adaptive behavior with fixed weights in recurrent neural networks: An overview. InProc. of International Joint Conference on Neural Networks(pp. 2018–2022). Honolulu, Hawaii. Puskorius,G.V.,&Feldkamp,L.A.(1994).Neurocontrolofnonlineardynamicalsys- tems with Kalman filter trained recurrent networks.IEEE Transactions on Neural Networks, 5(2), 279–297. Puskorius, G. V., & Feldkamp, L. A. (1996). Dynamic neural network methods ap- plied to on-vehicle idle speed control.Proceedings of IEEE, 84(10), 1407–1420. Rao, Y., Kim, S., Sanchez, J., Erdogmus, D., Principe, J. C., Carmena, J., Lebedev, M., & Nicolelis, M. (2005). Learning mappings in brain machine interfaces with echo state networks. In2005 IEEE International Conference on Acoustics, Speech, and Signal Processing. Philadelphia. Renyi, A. (1970).Probability theory. New York: Elsevier. Sanchez, J. C. (2004).From cortical neural spike trains to behavior: Modeling and analysis. Unpublished doctoral dissertation, University of Florida. Santiago, R. A., & Lendaris, G. G. (2004). Context discerning multifunction net- works: Reformulating fixed weight neural networks. InProc. of International Joint Conference on Neural Networks(pp. 189–194). Budapest, Hungary. Shah, J. V., & Poon, C.-S. (1999). Linear independence of internal representations in multilayer perceptrons.IEEE Transactions on Neural Networks, 10(1), 10–18. Shannon,C.E.(1948).Amathematicaltheoryofcommunication.BellSystemTechnical Journal, 27, 623–656. Siegelmann, H. T. (1993).Foundations of recurrent neural networks. Unpublished doc- toral dissertation, Rutgers University. Siegelmann,H.T.,&Sontag,E.(1991).Turingcomputabilitywithneuralnets.Applied Mathematics Letters, 4(6), 77–80. Singhal, S., & Wu, L. (1989). Training multilayer perceptrons with the extended Kalman algorithm. In D. S. Touretzky (Ed.),Advances in neural information process- ing systems, 1(pp. 133–140). San Mateo, CA: Morgan Kaufmann. Takens, F. (1981). Detecting strange attractors in turbulence. In D. A. Rand & L.-S. Young (Eds.),Dynamical systems and turbulence(pp. 366–381). Berlin: Springer. Thogula, R. (2003).Information theoretic self-organization of multiple agents.Unpub- lished master’s thesis, University of Florida. Werbos, P. (1990). Backpropagation through time: What it does and how to do it. Proceedings of IEEE, 78(10), 1550–1560. 138 M. Ozturk, D. Xu, and J. Pr´ıncipe Werbos, P. (1992). Neurocontrol and supervised learning: An overview and evalua- tion. In D. White & D. Sofge (Eds.),Handbook of intelligent control(pp. 65–89). New York: Van Nostrand Reinhold. Wilde, D. J. (1964).Optimum seeking methods. Upper Saddle River, NJ: Prentice Hall. Williams, R. J., & Zipser, D. (1989). A learning algorithm for continually running fully recurrent neural networks.Neural Computation, 1, 270–280. <> <> <> <> <> <> Bayesian Compression for Deep Learning Christos Louizos Karen Ullrich Max Welling University of Amsterdam University of Amsterdam University of Amsterdam TNO Intelligent Imaging k.ullrich@uva.nl CIFAR c.louizos@uva.nl m.welling@uva.nl Abstract Compression and computational efficiency in deep learning have become a problem of great significance. In this work, we argue that the most principled and effective way to attack this problem is by adopting a Bayesian point of view, where through sparsity inducing priors we prune large parts of the network. We introduce two novelties in this paper: 1) we use hierarchical priors to prune nodes instead of individual weights, and 2) we use the posterior uncertainties to determine the optimal fixed point precision to encode the weights. Both factors significantly contribute to achieving the state of the art in terms of compression rates, while still staying competitive with methods designed to optimize for speed or energy efficiency. 1 Introduction While deep neural networks have become extremely successful in in a wide range of applications, often exceeding human performance, they remain difficult to apply in many real world scenarios. For instance, making billions of predictions per day comes with substantial energy costs given the energy consumption of common Graphical Processing Units (GPUs). Also, real-time predictions are often about a factor100away in terms of speed from what deep NNs can deliver, and sending NNs with millions of parameters through band limited channels is still impractical. As a result, running them on hardware limited devices such as smart phones, robots or cars requires substantial improvements on all of these issues. For all those reasons, compression and efficiency have become a topic of interest in the deep learning community. While all of these issues are certainly related, compression and performance optimizing procedures might not always be aligned. As an illustration, consider the convolutional layers of Alexnet, which account for only 4% of the parameters but 91% of the computation [68]. Compressing these layers will not contribute much to the overall memory footprint. There is a variety of approaches to address these problem settings. However, most methods have the common strategy of reducing both the neural network structure and the effective fixed point precision for each weight. A justification for the former is the finding that NNs suffer from significant parameter redundancy [14]. Methods in this line of thought are network pruning, where unnecessary connections are being removed [40,24,21], or student-teacher learning where a large network is used to train a significantly smaller network [5, 27]. From a Bayesian perspective network pruning and reducing bit precision for the weights is aligned with achieving high accuracy, because Bayesian methods search for the optimal model structure (which leads to pruning with sparsity inducing priors), and reward uncertain posteriors over parameters through the bits back argument [28] (which leads to removing insignificant bits). This relation is made explicit in the MDL principle [20] which is known to be related to Bayesian inference. In this paper we will use the variational Bayesian approximation for Bayesian inference which has also been explicitly interpreted in terms of model compression [28]. By employing sparsity inducing priors for hidden units (and not individual weights) we can prune neurons including all their ingoing and outgoing weights. This avoids more complicated and inefficient coding schemes needed for pruning or vector quantizing individual weights. As an additional Bayesian bonus we can use the variational posterior uncertainty to assess which bits are significant and remove the ones which fluctuate too much under approximate posterior sampling. From this we derive the optimal fixed point precision per layer, which is still practical on chip. 2 Variational Bayes and Minimum Description Length A fundamental theorem in information theory is the minimum description length (MDL) principle [20]. It relates to compression directly in that it defines the best hypothesis to be the one that communicates the sum of the model (complexity costLC ) and the data misfit (error costLE ) with the minimum number of bits [59,60]. It is well understood that variational inference can be reinterpreted from an MDL point of view [56,72,28,30,19]. More specifically, assume that we are presented with a dataset QD that consists from N input-output pairs <>. Let <> be a parametric model, e.g. a deep neural network, that maps inputs x to their corresponding outputs y using parameters w governed by a prior distribution <>. In this scenario, we wish to approximate the intractable posterior distribution <> with a fixed form approximate posterior <> by optimizing the variational parameters according to: <> where <> denotes the entropy and <> is known as the evidence-lower-bound (ELBO) or negative variational free energy. As indicated in eq.1, <> naturally decomposes into a minimum cost for communicating the targets <> under the assumption that the sender and receiver agreed on a n=1 prior <> and that the receiver knows the inputs <> and form of the parametric model. n=1 By using sparsity inducing priors for groups of weights that feed into a neuron the Bayesian mecha- nism will start pruning hidden units that are not strictly necessary for prediction and thus achieving compression. But there is also a second mechanism by which Bayes can help us compress. By explicitly entertaining noisy weight encodings through <> we can benefit from the bits-back argument [28,30] due to the entropy term; this is in contrast to infinitely precise weights that lead to <>. Nevertheless in practice, the data misfit termLE is intractable for neural network models under a noisy weight encoding, so as a solution Monte Carlo integration is usually employed. Continuous q(w) allow for the reparametrization trick [36,58]. Here, we replace sampling from q(w) by a deterministic function of the variational parameters and random samples from some noise variables: <>; (2) where <>. By applying this trick, we obtain unbiased stochastic gradients of the ELBO with respect to the variational parameters, thus resulting in a standard optimization problem that is fit for stochastic gradient ascent. The efficiency of the gradient estimator resulting from eq. 2 can be further improved for neural networks by utilizing local reparametrizations [37] (which we will use in our experiments); they provide variance reduction in an efficient way by locally marginalizing the weights at each layer and instead sampling the distribution of the pre-activations. 3 Related Work One of the earliest ideas and most direct approaches to tackle efficiency is pruning. Originally introduced by [40], pruning has recently been demonstrated to be applicable to modern architectures [25,21]. It had been demonstrated that an overwhelming amount of up to 99,5% of parameters can be pruned in common architectures. There have been quite a few encouraging results obtained by (empirical) Bayesian approaches that employ weight pruning [19,7,52,70,51]. Nevertheless, 2 In practice this term is a large constant determined by the weight precision. weight pruning is in general inefficient for compression since the matrix format of the weights is not taken into consideration, therefore the Compressed Sparse Column (CSC) format has to be employed. Moreover, note that in conventional CNNs most flops are used by the convolution operation. Inspired by this observation, several authors proposed pruning schemes that take these considerations into account [73, 74] or even go as far as efficiency aware architectures to begin with [32, 15, 31]. From the Bayesian viewpoint, similar pruning schemes have been explored at [47, 53, 39, 34]. Given optimal architecture, NNs can further be compressed by quantization. More precisely, there are two common techniques. First, the set of accessible weights can be reduced drastically. As an extreme example, [13,48,57,76] and [11] trained NN to use only binary or tertiary weights with floating point gradients. This approach however is in need of significantly more parameters than their ordinary counterparts. Work by [18] explores various techniques beyond binary quantization: k-means quantization, product quantization and residual quantization. Later studies extent this set to optimal fixed point [44] and hashing quantization [10]. [25] apply k-means clustering and consequent center training. From a practical point of view, however, all these are fairly unpractical during test time. For the computation of each feature map in a net, the original weight matrix must be reconstructed from the indexes in the matrix and a codebook that contains all the original weights. This is an expensive operation and this is why some studies propose a different approach than set quantization. Precision quantization simply reduces the bit size per weight. This has a great advantage over set quantization at inference time since feature maps can simply be computed with less precision weights. Several studies show that this has little to no effect on network accuracy when using 16bit weights [49,22,12,71,9]. Somewhat orthogonal to the above discussion but certainly relevant are approaches that customize the implementation of CNNs for hardware limited devices[31, 4, 62]. 4 Bayesian compression with scale mixtures of normals Consider the following prior over a parameter w where its scale z is governed by a distribution <>: <>; (3) with z2 serving as the variance of the zero-mean normal distribution over w. By treating the scales of w as random variables we can recover marginal prior distributions over the parameters that have heavier tails and more mass at zero; this subsequently biases the posterior distribution over w to be sparse. This family of distributions is known as scale-mixtures of normals [6,2] and it is quite general, as a lot of well known sparsity inducing distributions are special cases. One example of the aforementioned framework is the spike-and-slab distribution [50], the golden standard for sparse Bayesian inference. Under the spike-and-slab, the mixing density of the scales is a Bernoulli distribution, thus the marginal <> has a delta “spike” at zero and a continuous “slab” over the real line. Unfortunately, this prior leads to a computationally expensive inference since we have to explore a space of2M models, whereMis the number of the model parameters. Dropout [29,67], one of the most popular regularization techniques for neural networks, can be interpreted as positing a spike and slab distribution over the weights where the variance of the “slab” is zero [17,45]. Another example is the Laplace distribution which arises by considering <>. The mode of the posterior distribution under a Laplace prior is known as the Lasso [69] estimator and has been previously used for sparsifying neural networks at [73,61]. While computationally simple, the Lasso estimator is prone to “shrinking" large signals [8] and only provides point estimates about the parameters. As a result it does not provide uncertainty estimates, it can potentially overfit and, according to the bits-back argument, is inefficient for compression. For these reasons, in this paper we will tackle the problem of compression and efficiency in neural networks by adopting a Bayesian treatment and inferring an approximate posterior distribution over the parameters under a scale mixture prior. We will consider two choices for the prior over the scales p(z); the hyperparameter free log-uniform prior [16,37] and the half-Cauchy prior, which results into a horseshoe [8] distribution. Both of these distributions correspond to a continuous relaxation of the spike-and-slab prior and we provide a brief discussion on their shrinkage properties at Appendix C. 4.1 Reparametrizing variational dropout for group sparsity One potential choice for p(z) is the improper log-uniform prior [37] <>. It turns out that we can recover the log-uniform prior over the weightswif we marginalize over the scales z: <> (4) This alternative parametrization of the log uniform prior is known in the statistics literature as the normal-Jeffreys prior and has been introduced by [16]. This formulation allows to “couple" the scales of weights that belong to the same group (e.g. neuron or feature map), by simply sharing the corresponding scale variablezin the joint prior 3 : <>; (5) where W is the weight matrix of a fully connected neural network layer with A being the dimen- sionality of the input and B the dimensionality of the output. Now consider performing variational inference with a joint approximate posterior parametrized as follows: <>; (6) where _i is the dropout rate [67,37,51] of the given group. As explained at [37,51], the multiplicative parametrization of the approximate posterior over z suffers from high variance gradients; therefore we will follow [51] and re-parametrize it in terms of <>, hence optimize w.r.t._2 . The <> lower bound under this prior and approximate posterior becomes: <> (7) Under this particular variational posterior parametrization the negative KL-divergence from the conditional prior <> to the approximate posterior <> is independent of z: <> (8) This independence can be better understood if we consider a non-centered parametrization of the prior [55]. More specifically, consider reparametrizing the weights asw~ij =wij ; this will then result zi into <>, where <>. Now if <> and <> we perform variational inference under the p(W~)p(z)prior with an approximate posterior that has Q the form of <>, with <>, then we see that we ij arrive at the same expressions for the negative KL-divergence from the prior to the approximate posterior. Finally, the negative KL-divergence from the normal-Jeffreys scale prior p(z) to the Gaussian variational posterior q depends only on the “implied” dropout rate, <>, and zi z takes the following form [51]: <>; (9) where <> are the sigmoid and softplus functions respectively 4 and k1 = 0:63576,k2 = 1:87320,k3 = 1:48695. We can now prune entire groups of parameters by simply specifying a thresh- old for the variational dropout rate of the corresponding group, e.g.<>. It should be mentioned that this prior parametrization readily allows for a more flexible marginal pos- terior over the weights as we now have a compound distribution, <>; this is in contrast to the original parametrization and the Gaussian approximations employed by [37,51]. Strictly speaking the result of eq. 4 only holds when each weight has its own scale and not when that scale is shared across multiple weights. Nevertheless, in practice we obtain a prior that behaves in a similar way, i.e. it biases the variational posterior to be sparse. <> Furthermore, this approach generalizes the low variance additive parametrization of variational dropout proposed for weight sparsity at [51] to group sparsity (which was left as an open question at [51]) in a principled way. At test time, in order to have a single feedforward pass we replace the distribution overWat each layer with a single weight matrix, the masked variational posterior mean: <>; (10) where m is a binary mask determined according to the group variational dropout rate andMW are the means ofq (W~). We further use the variational posterior marginal variances 5 for this particular posterior approximation: <>; (11) to assess the bit precision of each weight in the weight matrix. More specifically, we employed the mean variance across the weight matrixW^ to compute the unit round off necessary to represent the weights. This method will give us the amount significant bits, and by adding 3 exponent and 1 sign bits we arrive at the final bit precision for the entire weight matrixW^6 . We provide more details at Appendix B. 4.2 Group horseshoe with half-Cauchy scale priors Another choice for p(z) is a proper half-Cauchy distribution: <>; it induces a horseshoe prior [8] distribution over the weights, which is a well known sparsity inducing prior in the statistics literature. More formally, the prior hierarchy over the weights is expressed as (in a non-centered parametrization): <>; (12) where0 is the free parameter that can be tuned for specific desiderata. The idea behind the horseshoe is that of the “global-local" shrinkage; the global scale variablespulls all of the variables towards zero whereas the heavy tailed local variableszi can compensate and allow for some weights to escape. Instead of directly working with the half-Cauchy priors we will employ a decomposition of the half-Cauchy that relies upon (inverse) gamma distributions [54] as this will allow us to compute the negative KL-divergence from the scale priorp(z)to an approximate log-normal scale posterior q (z)in closed form (the derivation is given in Appendix D). More specifically, we have that the half-Cauchy prior can be expressed in a non-centered parametrization as: <>; (13) where <> correspond to the inverse Gamma and Gamma distributions in the scale parametrization, and z follows a half-Cauchy distribution with scale k. Therefore we will re-express the whole hierarchy as: <>; (14) It should be mentioned that the improper log-uniform prior is the limiting case of the horseshoe prior when the shapes of the (inverse) Gamma hyperpriors on <> go to zero [8]. In fact, several well known shrinkage priors can be expressed in this form by altering the shapes of the (inverse) Gamma hyperpriors [3]. For the variational posterior we will employ the following mean field approximation: <>. Notice that the fact that we are using mean-field variational approximations (which we chose for simplicity) can potentially underestimate the variance, thus lead to higher bit precisions for the weights. We leave the exploration of more involved posteriors for future work. Where <> is a log-normal distribution. It should be mentioned that a similar form of non- centered variational inference for the horseshoe has been also successfully employed for undirected models at [q 33]. Notice that we can also apply local reparametrizations [37] when we are sampling <> i i and sa sb by exploiting properties of the log-normal distribution 7 and thus forming the implied: <> (17) As a threshold rule for group pruning we will use the negative log-mode 8 of the local log-normal r.v. <> , i.e. prune when <>, with <>. This ignores <> and <>, but nonetheless we found <> dependencies among the zi elements induced by the common scale that it works well in practice. Similarly with the group normal-Jeffreys prior, we will replace the distribution overWat each layer with the masked variational posterior mean during test time: <>; (19) wheremis a binary mask determined according to the aforementioned threshold,MW are the means ofq(W~)and;2 are the means and variances of the local log-normals over <>. Furthermore, similarly to the group normal-Jeffreys approach, we will use the variational posterior marginal variances: <>; (20) to compute the final bit precision for the entire weight matrix W. 5 Experiments We validated the compression and speed-up capabilities of our models on the well-known architectures of LeNet-300-100 [41], LeNet-5-Caffe 9 on MNIST [42] and, similarly with [51], VGG [63]10 on CIFAR 10 [38]. The groups of parameters were constructed by coupling the scale variables for each filter for the convolutional layers and for each input neuron for the fully connected layers. We provide the algorithms that describe the forward pass using local reparametrizations for fully connected and convolutional layers with each of the employed approximate posteriors at appendix F. For the horseshoe prior we set the scale 0 of the global half-Cauchy prior to a reasonably small value, e.g. 0 = 1e5. This further increases the prior mass at zero, which is essential for sparse estimation and compression. We also found that constraining the standard deviations as described at [46] and “warm-up" [65] helps in avoiding bad local optima of the variational objective. Further details about the experimental setup can be found at Appendix A. Determining the threshold for pruning can be easily done with manual inspection as usually there are two well separated clusters (signal and noise). We provide a sample visualization at Appendix E. 5.1 Architecture learning & bit precisions We will first demonstrate the group sparsity capabilities of our methods by illustrating the learned architectures at Table 1, along with the inferred bit precision per layer. As we can observe, our methods infer significantly smaller architectures for the LeNet-300-100 and LeNet-5-Caffe, compared to Sparse Variational Dropout, Generalized Dropout and Group Lasso. Interestingly, we observe that for the VGG network almost all of big 512 feature map layers are drastically reduced to around 10 feature maps whereas the initial layers are mostly kept intact. Furthermore, all of the Bayesian methods considered require far fewer than the standard 32 bits per-layer to represent the weights, sometimes even allowing for 5 bit precisions. The product of log-normal r.v.s is another log-normal and a power of a log-normal r.v. is another log-normal. Empirically, it slightly better separates the scales compared to the negative log-mean <>. https://github.com/BVLC/caffe/tree/master/examples/mnist The adapted CIFAR 10 version described athttp://torch.ch/blog/2015/07/30/cifar.html. Table 1: Learned architectures with Sparse VD [51], Generalized Dropout (GD) [66] and Group Lasso (GL) [73]. Bayesian Compression (BC) with group normal-Jeffreys (BC-GNJ) and group horseshoe (BC-GHS) priors correspond to the proposed models. We show the amount of neurons left after pruning along with the average bit precisions for the weights at each layer. <

> 5.2 Compression Rates For the actual compression task we compare our method to current work in three different scenarios: (i) compression achieved only by pruning, here, for non-group methods we use the CSC format to store parameters; (ii) compression based on the former but with reduced bit precision per layer (only for the weights); and (iii) the maximum compression rate as proposed by [25]. We believe Table 2: Compression results for our methods. “DC” corresponds to Deep Compression method introduced at [25], “DNS” to the method of [21] and “SWS” to the Soft-Weight Sharing of [70]. Numbers marked with * are best case guesses. <> these to be relevant scenarios because (i) can be applied with already existing frameworks such as Tensorflow [1], (ii) is a practical scheme given upcoming GPUs and frameworks will be designed to work with low and mixed precision arithmetics [43,23]. For (iii), we perform k-means clustering on the weights with k=32 and consequently store a weight index that points to a codebook of available weights. Note that the latter achieves highest compression rate but it is however fairly unpractical at test time since the original matrix needs to be restored for each layer. As we can observe at Table 2, our methods are competitive with the state-of-the art for LeNet-300-100 while offering significantly better compression rates on the LeNet-5-Caffe architecture, without any loss in accuracy. Do note that group sparsity and weight sparsity can be combined so as to further prune some weights when a particular group is not removed, thus we can potentially further boost compression performance at e.g. LeNet-300-100. For the VGG network we observe that training from a random initialization yielded consistently less accuracy (around 1%-2% less) compared to initializing the means of the approximate posterior from a pretrained network, similarly with [51], thus we only report the latter results 11 . After initialization we trained the VGG network regularly for 200 epochs using Adam with the default hyperparameters. We observe a small drop in accuracy for the final models when using the deterministic version of the network for prediction, but nevertheless averaging across multiple samples restores the original accuracy. Note, that in general we can maintain the original accuracy on VGG without sampling by simply finetuning with a small learning rate, as done at [51]. This will still induce (less) sparsity but unfortunately it does not lead to good compression as the bit precision remains very high due to not appropriately increasing the marginal variances of the weights. 5.3 Speed and energy consumption We demonstrate that our method is competitive with [73], denoted as GL, a method that explicitly prunes convolutional kernels to reduce compute time. We measure the time and energy consumption of one forward pass of a mini-batch with batch size 8192 through LeNet-5-Caffe. We average over10 4 forward passes and all experiments were run with Tensorflow 1.0.1, cuda 8.0 and respective cuDNN. We apply 16 CPUs run in parallel (CPU) or a Titan X (GPU). Note that we only use the pruned architecture as lower bit precision would further increase the speed-up but is not implementable in any common framework. Further, all methods we compare to in the latter experiments would barely show an improvement at all since they do not learn to prune groups but only parameters. In figure 1 we present our results. As to be expected the largest effect on the speed up is caused by GPU usage. However, both our models and best competing models reach a speed up factor of around 8x. We can further save about 3x energy costs by applying our architecture instead of the original one on a GPU. For larger networks the speed-up is even higher: for the VGG experiments with batch size 256 we have a speed-up factor of 51x. <

> Figure 1:Left:Avg. Time a batch of 8192 samples takes to pass through LeNet-5-Caffe. Numbers on top of the bars represent speed-up factor relative to the CPU implementation of the original network. Right:Energy consumption of the GPU of the same process (when run on GPU). 6 Conclusion We introduced Bayesian compression, a way to tackle efficiency and compression in deep neural networks in a unified and principled way. Our proposed methods allow for theoretically principled compression of neural networks, improved energy efficiency with reduced computation while naturally learning the bit precisions for each weight. This serves as a strong argument in favor of Bayesian methods for neural networks, when we are concerned with compression and speed up. 11 We also tried to finetune the same network with Sparse VD, but unfortunately it increased the error considerably (around 3% extra error), therefore we do not report those results. 8 Acknowledgments We would like to thank Dmitry Molchanov, Dmitry Vetrov, Klamer Schutte and Dennis Koelma for valuable discussions and feedback. This research was supported by TNO, NWO and Google. References [1]M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, et al. Tensorflow: Large-scale machine learning on heterogeneous distributed systems.arXiv preprint arXiv:1603.04467, 2016. [2]D. F. Andrews and C. L. Mallows. Scale mixtures of normal distributions.Journal of the Royal Statistical Society. Series B (Methodological), pages 99–102, 1974. [3]A. Armagan, M. Clyde, and D. B. Dunson. Generalized beta mixtures of gaussians. InAdvances in neural information processing systems, pages 523–531, 2011. [4]E. Azarkhish, D. Rossi, I. Loi, and L. Benini. Neurostream: Scalable and energy efficient deep learning with smart memory cubes.arXiv preprint arXiv:1701.06420, 2017. [5]J. Ba and R. Caruana. Do deep nets really need to be deep? InAdvances in neural information processing systems, pages 2654–2662, 2014. [6] E. Beale, C. Mallows, et al. Scale mixing of symmetric distributions with zero means.The Annals of Mathematical Statistics, 30(4):1145–1151, 1959. [7]C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra. Weight uncertainty in neural networks. Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, 2015. [8]C. M. Carvalho, N. G. Polson, and J. G. Scott. The horseshoe estimator for sparse signals.Biometrika, 97 (2):465–480, 2010. [9]S. Chai, A. Raghavan, D. Zhang, M. Amer, and T. Shields. Low precision neural networks using subband decomposition.arXiv preprint arXiv:1703.08595, 2017. [10]W. Chen, J. T. Wilson, S. Tyree, K. Q. Weinberger, and Y. Chen. Compressing convolutional neural networks.arXiv preprint arXiv:1506.04449, 2015. [11]M. Courbariaux and Y. Bengio. Binarynet: Training deep neural networks with weights and activations constrained to+1or1.arXiv preprint arXiv:1602.02830, 2016. [12]M. Courbariaux, J.-P. David, and Y. Bengio. Training deep neural networks with low precision multiplica- tions.arXiv preprint arXiv:1412.7024, 2014. [13]M. Courbariaux, Y. Bengio, and J.-P. David. Binaryconnect: Training deep neural networks with binary weights during propagations. InAdvances in Neural Information Processing Systems, pages 3105–3113, 2015. [14]M. Denil, B. Shakibi, L. Dinh, N. de Freitas, et al. Predicting parameters in deep learning. InAdvances in Neural Information Processing Systems, pages 2148–2156, 2013. [15]X. Dong, J. Huang, Y. Yang, and S. Yan. More is less: A more complicated network with less inference complexity.arXiv preprint arXiv:1703.08651, 2017. [16]M. A. Figueiredo. Adaptive sparseness using jeffreys’ prior.Advances in neural information processing systems, 1:697–704, 2002. [17]Y. Gal and Z. Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning.ICML, 2016. [18]Y. Gong, L. Liu, M. Yang, and L. Bourdev. Compressing deep convolutional networks using vector quantization.ICLR, 2015. [19]A. Graves. Practical variational inference for neural networks. InAdvances in Neural Information Processing Systems, pages 2348–2356, 2011. [20]P. D. Grünwald.The minimum description length principle. MIT press, 2007. [21]Y. Guo, A. Yao, and Y. Chen. Dynamic network surgery for efficient dnns. InAdvances In Neural Information Processing Systems, pages 1379–1387, 2016. [22]S. Gupta, A. Agrawal, K. Gopalakrishnan, and P. Narayanan. Deep learning with limited numerical precision.CoRR, abs/1502.02551, 392, 2015. [23]P. Gysel. Ristretto: Hardware-oriented approximation of convolutional neural networks.Master’s thesis, University of California, 2016. [24]S. Han, J. Pool, J. Tran, and W. Dally. Learning both weights and connections for efficient neural networks. InAdvances in Neural Information Processing Systems, pages 1135–1143, 2015. [25]S. Han, H. Mao, and W. J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding.ICLR, 2016. [26]K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. InProceedings of the IEEE International Conference on Computer Vision, pages 1026–1034, 2015. [27]G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. [28]G. E. Hinton and D. Van Camp. Keeping the neural networks simple by minimizing the description length of the weights. InProceedings of the sixth annual conference on Computational learning theory, pages 5–13. ACM, 1993. [29]G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors.arXiv preprint arXiv:1207.0580, 2012. [30]A. Honkela and H. Valpola. Variational learning and bits-back coding: an information-theoretic view to bayesian learning.IEEE Transactions on Neural Networks, 15(4):800–810, 2004. [31]A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. [32]F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. Keutzer. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and< 0.5 mb model size.ICLR, 2017. [33]J. B. Ingraham and D. S. Marks. Bayesian sparsity for intractable distributions. arXiv preprint arXiv:1602.03807, 2016. [34]T. Karaletsos and G. Rätsch. Automatic relevance determination for deep generative models.arXiv preprint arXiv:1505.07765, 2015. [35]D. Kingma and J. Ba. Adam: A method for stochastic optimization.International Conference on Learning Representations (ICLR), San Diego, 2015. [36]D. P. Kingma and M. Welling. Auto-encoding variational bayes.International Conference on Learning Representations (ICLR), 2014. [37]D. P. Kingma, T. Salimans, and M. Welling. Variational dropout and the local reparametrization trick. Advances in Neural Information Processing Systems, 2015. [38]A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images, 2009. [39]N. D. Lawrence. Note relevance determination. InNeural Nets WIRN Vietri-01, pages 128–133. Springer, 2002. [40]Y. LeCun, J. S. Denker, S. A. Solla, R. E. Howard, and L. D. Jackel. Optimal brain damage. InNIPs, volume 2, pages 598–605, 1989. [41]Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. [42]Y. LeCun, C. Cortes, and C. J. Burges. The mnist database of handwritten digits, 1998. [43]D. D. Lin and S. S. Talathi. Overcoming challenges in fixed point training of deep convolutional networks. Workshop ICML, 2016. [44]D. D. Lin, S. S. Talathi, and V. S. Annapureddy. Fixed point quantization of deep convolutional networks. arXiv preprint arXiv:1511.06393, 2015. [45]C. Louizos. Smart regularization of deep architectures.Master’s thesis, University of Amsterdam, 2015. [46]C. Louizos and M. Welling. Multiplicative Normalizing Flows for Variational Bayesian Neural Networks. ArXiv e-prints, Mar. 2017. [47]D. J. MacKay. Probable networks and plausible predictions—a review of practical bayesian methods for supervised neural networks.Network: Computation in Neural Systems, 6(3):469–505, 1995. [48]N. Mellempudi, A. Kundu, D. Mudigere, D. Das, B. Kaul, and P. Dubey. Ternary neural networks with fine-grained quantization.arXiv preprint arXiv:1705.01462, 2017. [49]P. Merolla, R. Appuswamy, J. Arthur, S. K. Esser, and D. Modha. Deep neural networks are robust to weight binarization and other non-linear distortions.arXiv preprint arXiv:1606.01981, 2016. [50]T. J. Mitchell and J. J. Beauchamp. Bayesian variable selection in linear regression. Journal of the American Statistical Association, 83(404):1023–1032, 1988. [51]D. Molchanov, A. Ashukha, and D. Vetrov. Variational dropout sparsifies deep neural networks.arXiv preprint arXiv:1701.05369, 2017. [52]E. Nalisnick, A. Anandkumar, and P. Smyth. A scale mixture perspective of multiplicative noise in neural networks.arXiv preprint arXiv:1506.03208, 2015. [53]R. M. Neal.Bayesian learning for neural networks. PhD thesis, Citeseer, 1995. [54]S. E. Neville, J. T. Ormerod, M. Wand, et al. Mean field variational bayes for continuous sparse signal shrinkage: pitfalls and remedies.Electronic Journal of Statistics, 8(1):1113–1151, 2014. [55]O. Papaspiliopoulos, G. O. Roberts, and M. Sköld. A general framework for the parametrization of hierarchical models.Statistical Science, pages 59–73, 2007. [56]C. Peterson. A mean field theory learning algorithm for neural networks.Complex systems, 1:995–1019, 1987. [57]M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. InEuropean Conference on Computer Vision, pages 525–542. Springer, 2016. [58]D. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic backpropagation and approximate inference in deep generative models. InProceedings of the 31th International Conference on Machine Learning, ICML 2014, Beijing, China, 21-26 June 2014, pages 1278–1286, 2014. [59]J. Rissanen. Modeling by shortest data description.Automatica, 14(5):465–471, 1978. [60]J. Rissanen. Stochastic complexity and modeling.The annals of statistics, pages 1080–1100, 1986. [61]S. Scardapane, D. Comminiello, A. Hussain, and A. Uncini. Group sparse regularization for deep neural networks.arXiv preprint arXiv:1607.00485, 2016. [62]S. Shi and X. Chu. Speeding up convolutional neural networks by exploiting the sparsity of rectifier units. arXiv preprint arXiv:1704.07724, 2017. [63]K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. ICLR, 2015. [64]M. Sites. Ieee standard for floating-point arithmetic. 2008. [65]C. K. Sønderby, T. Raiko, L. Maaløe, S. K. Sønderby, and O. Winther. Ladder variational autoencoders. arXiv preprint arXiv:1602.02282, 2016. [66]S. Srinivas and R. V. Babu. Generalized dropout.arXiv preprint arXiv:1611.06791, 2016. [67]N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting.The Journal of Machine Learning Research, 15(1):1929–1958, 2014. [68]V. Sze, Y.-H. Chen, T.-J. Yang, and J. Emer. Efficient processing of deep neural networks: A tutorial and survey.arXiv preprint arXiv:1703.09039, 2017. [69]R. Tibshirani. Regression shrinkage and selection via the lasso.Journal of the Royal Statistical Society. Series B (Methodological), pages 267–288, 1996. [70]K. Ullrich, E. Meeds, and M. Welling. Soft weight-sharing for neural network compression.ICLR, 2017. [71]G. Venkatesh, E. Nurvitadhi, and D. Marr. Accelerating deep convolutional networks using low-precision and sparsity.arXiv preprint arXiv:1610.00324, 2016. [72]C. S. Wallace. Classification by minimum-message-length inference. InInternational Conference on Computing and Information, pages 72–81. Springer, 1990. [73]W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li. Learning structured sparsity in deep neural networks. In Advances In Neural Information Processing Systems, pages 2074–2082, 2016. [74]T.-J. Yang, Y.-H. Chen, and V. Sze. Designing energy-efficient convolutional neural networks using energy-aware pruning.CVPR, 2017. [75]S. Zagoruyko and N. Komodakis. Wide residual networks.arXiv preprint arXiv:1605.07146, 2016. [76]C. Zhu, S. Han, H. Mao, and W. J. Dally. Trained ternary quantization.ICLR, 2017. Appendix A. Detailed experimental setup We implemented our methods in Tensorflow [1] and optimized the variational parameters using Adam [35] with the default hyperparameters. The means of the conditional Gaussian <> Table 3: Floating point formats Bits per Exponent <

> were initialized with the scheme proposed at [26], whereas the log of the standard deviations were initialized by sampling from N(9;1e4). The parameters of q(z) were initialized such that the overall mean of zise 1 and the overall variance is very low (1e^8); this ensures that all of the groups are active during the initial training iterations. As for the standard deviation constraints; for the LeNet-300-100 architecture we constrained the standard deviation of the first layer to be 0:2 whereas for the LeNet-5-Caffe we constrained the standard deviation of the first layer to be 0:5. The remaining standard deviations were left unconstrained. For the VGG network we constrained the standard deviations of the 64 and 128 feature map layers to be 0:1, the standard deviations of the 256 feature map layers to be0:2 and left the rest of the standard deviations unconstrained. We also found beneficial the incorporation of “warm-up” [65], i.e we annealed the negative KL-divergence from the prior to the approximate posterior with a linear schedule for the first 100 epochs. We initialized the means of the approximate posterior by the weights and biases obtained from a VGG network trained with batch normalization and dropout on CIFAR 10. For our method we disabled batch-normalization during training. As for preprocessing the data; for MNIST the only preprocessing we did was to rescale the digits to lie at the [-1,1] range and for CIFAR 10 we used the preprocessed dataset provided by [75]. Furthermore, do note that by pruning a given filter at a particular convolutional layer we can also prune the parameters corresponding to that feature map for the next layer. This similarly holds for fully connected layers; if we drop a given input neuron then the weights corresponding to that node from the previous layer can also be pruned. B. Standards for Floating-Point Arithmetic Floating points values eventually need to be represented in a binary basis in a computer. The most common standard today is the IEEE 754-2008 convention [64]. It definesx-bit base-2 formats, officially referred to as binaryx, withx2 f16;32;64;128g. The formats are also widely known as half, single, double and quadruple precision floats, respectively and used in almost all programming languages as a standard. The format considers 3 kinds of bits: one sign bit,wexponent bits andp precision bits. <

> Figure 2: A symbolic representation of the binaryxformat [64]. The Sign bit determines the sign of the number to be represented. The exponentEis anw-bit signed integer, e.g. for single precisionw= 8and thusE2[127;128]. In practice, exponents range from is smaller since the first and the last number are reserved for special numbers. The true significand or mantissa includes t bits on the right of the binary point. There is an implicit leading bit with value one. A values is consequently decomposed as follows <> (21) <> (22) In table 3, we summarize common and less common floating point formats. There is however the possibility to design a self defined format. There are 3 important quantities when choosing the right specification: overflow, underflow and unit round off also known as machine precision. Each one can be computed knowing the number of exponent and significant bits. in our work for example we consider a format that uses significantly less exponent bits since network parameters usually vary between [-10,10]. We set the unit round off equal to the precision and thus can compute the significant bits necessary to represent a specific weight. Beyond designing a tailored floating point format for deep learning, recent work also explored the possibility of deep learning with mixed formats [43,23]. For example, imagine the activations having high precision while weights can be low precision. C. Shrinkage properties of the normal-Jeffreys and horseshoe priors <

> Figure 3: Comparison of the behavior of the log-uniform / normal-Jeffreys (NJ) prior and the horseshoe (HS) prior (wheres= 1). Both priors behave similarly at zero but the normal-Jeffreys has an extremely heavy tail (thus making it non-normalizable). In this section we will provide some insights about the behavior of each of the priors we employ by following the excellent analysis of [8]; we can perform a change of variables and express the scale mixture distribution of eq.3 in the main paper in terms of a shrinkage coefficient, <> (23) It is easy to observe that eq. 23 corresponds to a continuous relaxation of the spike-and-slab prior: when <<= 0>> we have that <>, i.e. no shrinkage/regularization forw, when <<= 1>> we have that <>, i.e.wis exactly zero, and when <<=1>> we have that <>. Now by examining the implied prior on the shrinkage coefficient for both the log-uniform and the horseshoe priors we can better study their behavior. As it is explained at the half-Cauchy prior onzcorresponds to a beta prior on the shrinkage coefficient, <>, whereas the normal-Jeffreys / log-uniform prior onzcorresponds <> with <>. The densities of both of these distributions can be seen at Figure 3b. As we can observe, the log- uniform prior posits a distribution that concentrates almost all of its mass at either0or1, essentially either pruning the parameter or keeping it close to the maximum likelihood estimate due <>. In contrast the horseshoe prior maintains enough probability mass for the in-between values of and thus can, potentially, offer better regularization and generalization. D. Negative KL-divergences for log-normal approximating posteriors Le <> be a log-normal approximating posterior. Here we will derive the negative KL-divergences toq(z)from inverse gamma, gamma and half-normal distributions. Letp(z)be an inverse gamma distribution, i.e. <>. The negative KL-divergence can be expressed as follows: <> (24) The second term is the entropy of the log-normal distribution which has the following form: <> (25) The first term is the negative cross-entropy of the log-normal approximate posterior from the inverse- Gamma prior: <> (26) <> (27) Since the natural logarithm of a log-normal distribution <> follows a normal distribution <> we have that <>. Furthermore we have that <> then <>, therefore <>. Putting everything together we have that: <> (28) Therefore the negative KL-divergence is: <> (29) Now let p(z) be a Gamma prior, i.e. <>. We have that the negative cross-entropy changes to: <> (30) <> (31) <> (32)2 Therefore the negative KL-divergence is: <> (33) Now, by employing the aforementioned we can express the negative KL-divergence from <> to <> as follows: <> with the KL-divergence for the weight distribution <> given by eq.8 in the main paper. E. Visualizations <

> Figure 4: Distribution of the thresholds for the Sparse Variational Dropout 4a, Bayesian Compression with group normal-Jeffreys (BC-GNJ) 4b and group Horseshoe (BC-GHS) 4c priors for the three layer LeNet-300-100 architecture. It is easily observed that there are usually two well separable groups with BC-GNJ and BC-GHS, thus making the choice for the threshold easy. Smaller values indicate signal whereas larger values indicate noise (i.e. useless groups). <

> Figure 5: Distribution of the bit precisions for the Sparse Variational Dropout 5a, Bayesian Com- pression with group normal-Jeffreys (BC-GNJ) 5b and group Horseshoe (BC-GHS) 5c priors for the three layer LeNet-300-100 architecture. All of the methods usually require far fewer than 32bits for the weights. F. Algorithms for the feedforward pass Algorithms 1, 2, 3, 4 describe the forward pass using local reparametrizations for fully connected and convolutional layers with the approximate posteriors for the Bayesian Compression (BC) with group normal-Jeffreys (BC-GNJ) and group Horseshoe (BC-GHS) priors employed at the experiments. For the fully connected layers we coupled the scales for each input neuron whereas for the convolutional we couple the scales for each output feature map.Mw ;w are the means and variances of each layer, His a minibatch of activations of sizeK. For the first layer we have thatH=XwhereXis the minibatch of inputs. For the convolutional layersNf are the number of convolutional filters,is the convolution operator and we assume the [batch, height, width, feature maps] convention. Algorithm 1 Fully connected BC-GNJ layer h. <> Algorithm 2Convolutional BC-GNJ layerh. <> Algorithm 3 Fully connected BC-GHS layerh. <> Algorithm 4Convolutional BC-GHS layerh. <> <> <> <> <> <> <> Channel Pruning for Accelerating Very Deep Neural Networks Yihui He* Xiangyu Zhang Jian Sun Xifian Jiaotong University Megvii Inc. Megvii Inc. Xifian, 710049, China Beijing, 100190, China Beijing, 100190, China heyihui@stu.xjtu.edu.cn zhangxiangyu@megvii.com sunjian@megvii.com Abstract In this paper, we introduce a new channel pruning method to accelerate very deep convolutional neural net.works. Given a trained CNN model, we propose an it.erative two-step algorithm to effectively prune each layer, by a LASSO regression based channel selection and least square reconstruction. We further generalize this algorithm to multi-layer and multi-branch cases. Our method re.duces the accumulated error and enhance the compatibility with various architectures. Our pruned VGG-16 achieves the state-of-the-art results by 5. speed-up along with only 0.3% increase of error. More importantly, our method is able to accelerate modern networks like ResNet, exception and suffers only 1.4%, 1.0% accuracy loss under 2. speed.up respectively, which is significant. 1. Introduction Recent CNN acceleration works fall into three categories: optimized implementation (e.g., FFT [47]), quantization (e.g., BinaryNet [8]), and structured simplification that convert a CNN into compact one [22]. This work focuses on the last one. Structured simplification mainly involves: tensor factorization [22], sparse connection [17], and channel pruning [48]. Tensor factorization factorizes a convolutional layer into several efficient ones (Fig. 1(c)). However, feature map width (number of channels) could not be reduced, which makes it difficult to decompose 1 . 1 convolutional layer favored by modern networks (e.g., GoogleNet [45], ResNet [18], Xception [7]). This type of method also intro.duces extra computation overhead. Sparse connection deactivates connections between neurons or channels (Fig. 1(b)). Though it is able to achieves high theoretical speed-up ratio, the sparse convolutional layers have an fiirregularfi shape which is not implementation friendly. In contrast, channel pruning directly reduces feature map width, which shrinks <

> Figure 1. Structured simplification methods that accelerate CNNs: (a) a network with 3 conv layers. (b) sparse connection deactivates some connections between channels. (c) tensor factorization factorizes a convolutional layer into several pieces. (d) channel pruning reduces number of channels in each layer (focus of this paper). a network into thinner one, as shown in Fig. 1(d). It is efficient on both CPU and GPU because no special implementation is required. Pruning channels is simple but challenging because re.moving channels in one layer might dramatically change the input of the following layer. Recently, training-based channel pruning works [1, 48] have focused on imposing sparse constrain on weights during training, which could adaptively determine hyper-parameters. However, training from scratch is very costly and results for very deep CNNs on ImageNet have been rarely reported. Inference-time at.tempts [31, 3] have focused on analysis of the importance of individual weight. The reported speed-up ratio is very limited. In this paper, we propose a new inference-time approach for channel pruning, utilizing redundancy inter channels. Inspired by tensor factorization improvement by feature maps reconstruction [52], instead of analyzing filter weights [22, 31], we fully exploits redundancy within feature maps. Specifically, given a trained CNN model, pruning each layer is achieved by minimizing reconstruction error on its output feature maps, as showed in Fig. 2. We solve this mini. <

> Figure 2. Channel pruning for accelerating a convolutional layer. We aim to reduce the width of feature map B, while minimizing the reconstruction error on feature map C. Our optimization algorithm (Sec. 3.1) performs within the dotted box, which does not involve nonlinearity. This figure illustrates the situation that two channels are pruned for feature map B. Thus corresponding channels of filters W can be removed. Furthermore, even though not directly optimized by our algorithm, the corresponding filters in the previous layer can also be removed (marked by dotted filters). c, n: number of channels for feature maps B and C, kh . kw : kernel size. minimization problem by two alternative steps: channels selection and feature map reconstruction. In one step, we figure out the most representative channels, and prune redundant ones, based on LASSO regression. In the other step, we reconstruct the outputs with remaining channels with linear least squares. We alternatively take two steps. Further, we approximate the network layer-by-layer, with accumulated error accounted. We also discuss methodologies to prune multi-branch networks (e.g., ResNet [18], exception [7]). For VGG-16, we achieve 4. acceleration, with only 1.0% increase of top-5 error. Combined with tensor factorization, we reach 5. acceleration but merely suffer 0.3% increase of error, which outperforms previous state-of-the.arts. We further speed up ResNet-50 and Xception-50 by 2. with only 1.4%, 1.0% accuracy loss respectively. 2. Related Work There has been a significant amount of work on accelerating CNNs. Many of them fall into three categories: optimized implementation [4], quantization [40], and structured simplification [22]. Optimized implementation based methods [35, 47, 27, 4] accelerate convolution, with special convolution algorithms like FFT [47]. Quantization [8, 40] reduces floating point computational complexity. Sparse connection eliminates connections between neurons [17, 32, 29, 15, 14]. [51] prunes connections based on weights magnitude. [16] could accelerate fully connected layers up to 50.. However, in practice, the actual speed-up maybe very related to implementation. Tensor factorization [22, 28, 13, 24] decompose weights into several pieces. [50, 10, 12] accelerate fully connected layers with truncated SVD. [52] factorize a layer into 3 . 3 and 1 . 1 combination, driven by feature map redundancy. Channel pruning removes redundant channels on feature maps. There are several training-based approaches. [1, 48] regularize networks to improve accuracy. Channel-wise SSL [48] reaches high compression ratio for first few conv layers of LeNet [30] and AlexNet [26]. However, training-based approaches are more costly, and the effectiveness for very deep networks on large datasets is rarely exploited. Inference-time channel pruning is challenging, as re.ported by previous works [2, 39]. Some works [44, 34, 19] focus on model size compression, which mainly operate the fully connected layers. Data-free approaches [31, 3] results for speed-up ratio (e.g., 5.) have not been reported, and requires long retraining procedure. [3] select channels via over 100 random trials, however it need long time to eval.ate each trial on a deep network, which makes it infeasible to work on very deep models and large datasets. [31] is even worse than naive solution from our observation sometimes (Sec. 4.1.1). 3. Approach In this section, we first propose a channel pruning algorithm for a single layer, then generalize this approach to multiple layers or the whole model. Furthermore, we dis.cuss variants of our approach for multi-branch networks. 3.1. Formulation Fig. 2 illustrates our channel pruning algorithm for a sin.gle convolutional layer. We aim to reduce the width of feature map B, while maintaining outputs in feature map C. Once channels are pruned, we can remove correspond.ing channels of the filters that take these channels as in.put. Also, filters that produce these channels can also be removed. It is clear that channel pruning involves two key points. The first is channel selection, since we need to select most representative channels to maintain as much information. The second is reconstruction. We need to reconstruct the following feature maps using the selected channels. Motivated by this, we propose an iterative two-step algorithm. In one step, we aim to select most representative channels. Since an exhaustive search is infeasible even for tiny networks, we come up with a LASSO regression based method to figure out representative channels and prune redundant ones. In the other step, we reconstruct the outputs with remaining channels with linear least squares. We alter.natively take two steps. Formally, to prune a feature map with c channels, we consider applying n.c.kh .kw convolutional filters W on <> input volumes X sampled from this feature map, which produces N . n output matrix Y. Here, N is the number of samples, n is the number of output channels, and kh,kw are the kernel size. For simple representation, bias term is not included in our formulation. To prune the .. input channels from c to desired <>, while minimizing reconstruction error, we formulate our problem as follow: <> (1) F is Frobenius norm. <> matrix sliced from ith channel of input volumes X_i, i =1, ..., c. W_i is n . filter weights sliced from ith channel of W. is coefficient vector of length c for channel selection, and .i is ith entry of . Notice that, if .i =0, X_i will be no longer useful, which could be safely pruned from feature map. W_i could also be removed. Optimization Solving this minimization problem in Eqn. 1 is NP-hard. Therefore, we relax the l_0 to l_1 regularization: <> (2) . is a penalty coefficient. By increasing l, there will be more zero terms in and one can get higher speed-up ratio. We also add a constrain .i WiF =1 to this formulation, which avoids trivial solution. Now we solve this problem in two folds. First, we fix W, solve for channel selection. Second, we fix , solve W to reconstruct error. (i) The subproblem of . In this case, W is fixed. We solve for channel selection. This problem can be solved by LASSO regression [46, 5], which is widely used for model selection. <> (3) . Here Zi =XiWi (size N .n). We will ignore ith channels if .i =0. (ii) The subproblem of W. In this case, is fixed. We utilize the selected channels to minimize reconstruction error. We can find optimized solution by least squares: <>. (4) Here <> (size N.). W is n reshaped W, <>. After obtained result W, it is reshaped back to W. Then we assign <>. Constrain <> satisfies. We alternatively optimize (i) and (ii). In the beginning, W is initialized from the trained model, <>, namely no penalty, and <>. We gradually increase <> For each change of <>, we iterate these two steps until k is stable. After <> satisfies, we obtain the final solution W from <> In practice, we found that the two steps iteration is time consuming. So we apply (i) multiple times, <> until <> satisfies. Then apply (ii) just once, to obtain <> the final result. From our observation, this result is comparable with two steps iterations. Therefore, in the following experiments, we adopt this approach for efficiency. Discussion: Some recent works [48, 1, 17] (though train. ing based) also introduce .1-norm or LASSO. However, we must emphasis that we use different formulations. Many of them introduced sparsity regularization into training loss, instead of explicitly solving LASSO. Other work [1] solved LASSO, while feature maps or data were not considered during optimization. Because of these differences, our ap.proach could be applied at inference time. 3.2. Whole Model Pruning Inspired by [52], we apply our approach layer by layer sequentially. For each layer, we obtain input volumes from the current input feature map, and output volumes from the output feature map of the un-pruned model. This could be formalized as: <> (5) Different from Eqn. 1, Y is replaced by Y . , which is from feature map of the original model. Therefore, the accumulated error could be accounted during sequential pruning. 3.3. Pruning Multi.Branch Networks The whole model pruning discussed above is enough for single-branch networks like LeNet [30], AlexNet [26] and VGG Nets [43]. However, it is insufficient for multi-branch networks like GoogLeNet [45] and ResNet [18]. We mainly focus on pruning the widely used residual structure (e.g., ResNet [18], Xception [7]). Given a residual block shown in Fig. 3 (left), the input bifurcates into shortcut and residual branch. On the residual branch, there are several convolutional layers (e.g., 3 convolutional layers which have spatial size of 1 . 1, 3 . 3, 1 . 1, Fig. 3, left). Other layers except the first and last layer can be pruned as is described previously. For the first layer, the challenge is that the large input feature map width (for ResNet, 4 times of its output) can it be easily pruned, since it is shared with shortcut. For the last layer, accumulated error from the shortcut is hard to be recovered, since there is no parameter on the shortcut. To address these challenges, we propose several variants of our approach as follows. <

> Figure 3. Illustration of multi-branch enhancement for residual block. Left: original residual block. Right: pruned residual block with enhancement, cx denotes the feature map width. Input channels of the first convolutional layer are sampled, so that the large input feature map width could be reduced. As for the last layer, rather than approximate Y2 , we try to approximate <> directly (Sec. 3.3 Last layer of residual branch). Last layer of residual branch: Shown in Fig. 3, the output layer of a residual block consists of two inputs: feature map Y1 and Y2 from the shortcut and residual branch. We aim to recover Y1 +Y2 for this block. Here, Y1, Y2 are the original feature maps before pruning. Y2 could be approximated as in Eqn. 1. However, shortcut branch is parameter-free, then Y1 could not be recovered directly. To compensate this error, the optimization goal of the last layer is changed from Y2 to Y1 .Y . +Y2, which does not change <> our optimization. Here, Y . is the current feature map after <> previous layers pruned. When pruning, volumes should be sampled correspondingly from these two branches. First layer of residual branch: Illustrated in Fig. 3(left), the input feature map of the residual block could not be pruned, since it is also shared with the short.cut branch. In this condition, we could perform feature map sampling before the first convolution to save computation. We still apply our algorithm as Eqn. 1. Differently, we sample the selected channels on the shared feature maps to construct a new input for the later convolution, shown in Fig. 3(right). Computational cost for this operation could be ignored. More importantly, after introducing feature map sampling, the convolution is still irregular. Filter-wise pruning is another option for the first con.volution on the residual branch. Since the input channels of parameter-free shortcut branch could not be pruned, we apply our Eqn. 1 to each filter independently (each fil.ter chooses its own representative input channels). Under single layer acceleration, filter-wise pruning is more accurate than our original one. From our experiments, it im.proves 0.5% top-5 accuracy for 2. ResNet-50 (applied on the first layer of each residual branch) without fine-tuning. However, after fine-tuning, there is no noticeable improvement. In addition, it outputs irregular convolutional layers, which need special library implementation support. We do not adopt it in the following experiments. 4. Experiment We evaluation our approach for the popular VGG Nets [43], ResNet [18], Xception [7] on ImageNet [9], CIFAR.10 [25] and PASCAL VOC 2007 [11]. For Batch Normalization [21], we first merge it into convolutional weights, which do not affect the outputs of the networks. So that each convolutional layer is followed by ReLU [36]. We use Caffe [23] for deep network evaluation, and scikit-learn [38] for solvers implementation. For channel pruning, we found that it is enough to extract 5000 images, and 10 samples per image. On ImageNet, we evaluate the top-5 accuracy with single view. Images are re.sized such that the shorter side is 256. The testing is on center crop of 224 . 224 pixels. We could gain more per.formance with fine-tuning. We use a batch size of 128 and .5 learning rate 1e^-4. We fine-tune our pruned models for 10 epochs. The augmentation for fine-tuning is random crop of 224 . 224 and mirror. 4.1. Experiments with VGG.16 VGG-16 [43] is a 16 layers single path convolutional neural network, with 13 convolutional layers. It is widely used in recognition, detection and segmentation, etc. Single view top-5 accuracy for VGG-16 is 89.9%1. 4.1.1 Single Layer Pruning In this subsection, we evaluate single layer acceleration performance using our algorithm in Sec. 3.1. For better under.standing, we compare our algorithm with two naive chan.nel selection strategies. first k selects the first k channels. max response selects channels based on corresponding filters that have high absolute weights sum [31]. For fair com.parison, we obtain the feature map indexes selected by each of them, then perform reconstruction (Sec. 3.1 (ii)). We hope that this could demonstrate the importance of channel selection. Performance is measured by increase of error af.ter a certain layer is pruned without fine-tuning, shown in Fig. 4. As expected, error increases as speed-up ratio increases. Our approach is consistently better than other approaches in different convolutional layers under different speed-up ra.tio. Unexpectedly, sometimes max response is even worse than first k. We argue that max response ignores correlations between different filters. Filters with large absolute weight may have strong correlation. Thus selection based on filter weights is less meaningful. Correlation on feature maps is worth exploiting. We can find that channel selection http://www.vlfeat.org/matconvnet/pretrained/ <

> Figure 4. Single layer performance analysis under different speed-up ratios (without fine-tuning), measured by increase of error. To verify the importance of channel selection referred in Sec. 3.1, we considered two naive baselines. first k selects the first k feature maps. max response selects channels based on absolute sum of corresponding weights filter [31]. Our approach is consistently better (smaller is better). <

> Table 1. Accelerating the VGG-16 model [43] using a speedup ratio of 2., 4., or 5. (smaller is better). affects reconstruction error a lot. Therefore, it is important for channel pruning. Also notice that channel pruning gradually becomes hard, from shallower to deeper layers. It indicates that shallower layers have much more redundancy, which is consistent with [52]. We could prune more aggressively on shallower layers in whole model acceleration. 4.1.2 Whole Model Pruning Shown in Table 1, whole model acceleration results under 2., 4., 5. are demonstrated. We adopt whole model pruning proposed in Sec. 3.2. Guided by single layer experiments above, we pruning more aggressive for shallower layers. Remaining channels ratios for shallow lay.ers (conv 1_x to conv 3_x) and deep layers (conv4_x) is 1:1.5. conv 5_x are not pruned, since they only con.tribute 9% computation in total and are not redundant. After fine-tuning, we could reach 2. speed-up without losing accuracy. Under 4., we only suffers 1.0% drops. Consistent with single layer analysis, our approach outperforms previous channel pruning approach (Li et al. [31]) by large margin. This is because we fully exploits channel redundancy within feature maps. Compared with tensor factorization algorithms, our approach is better than Jaderberg et al. [22], without fine-tuning. Though worse than Asym. [52], our combined model outperforms its combined Asym. 3D (Table 2). This may indicate that channel pruning is more challenging than tensor factorization, since removing channels in one layer might dramatically change the input of the following layer. However, channel pruning keeps the original model architecture, do not introduce additional layers, and the absolute speed-up ratio on GPU is much higher (Table 3). Since our approach exploits a new cardinality, we further combine our channel pruning with spatial factorization [22] and channel factorization [52]. Demonstrated in Table 2, <> Table 2. Performance of combined methods on the VGG-16 model [43] using a speed-up ratio of 4. or 5.. Our 3C solution outperforms previous approaches (smaller is better). our 3 cardinalities acceleration (spatial, channel factorization, and channel pruning, denoted by 3C) outperforms previous state-of-the-arts. Asym. 3D [52] (spatial and chan.nel factorization), factorizes a convolutional layer to three parts: <>. We apply spatial factorization, channel factorization, and our channel pruning together sequentially layer-by-layer. We fine-tune the accelerated models for 20 epochs, since they are 3 times deeper than the original ones. After fine-tuning, our 4. model suffers no degradation. Clearly, a combination of different acceleration techniques is better than any single one. This indicates that a model is redundant in each cardinality. 4.1.3 Comparisons of Absolute Performance We further evaluate absolute performance of acceleration on GPU. Results in Table 3 are obtained under Caffe [23], CUDA 8 [37] and cuDNN5 [6], with a mini-batch of 32 on a GPU (GeForce GTX TITAN X). Results are averaged from 50 runs. Tensor factorization approaches decompose weights into too many pieces, which heavily increase over.head. They could not gain much absolute speed-up. Though our approach also encountered performance decadence, it generalizes better on GPU than other approaches. Our re.sults for tensor factorization differ from previous research [52, 22], maybe because current library and hardware prefer single large convolution instead of several small ones. 4.1.4 Comparisons with Training from Scratch Though training a compact model from scratch is time-consuming (usually 120 epochs), it worths comparing our approach and from scratch counterparts. To be fair, we evaluated both from scratch counterpart, and normal setting net.work that has the same computational complexity and same architecture. Shown in Table 4, we observed that it is difficult for from scratch counterparts to reach competitive accuracy. our model outperforms from scratch one. Our approach successfully picks out informative channels and constructs highly compact models. We can safely draw the conclusion that the same model is difficult to be obtained from scratch. This coincides with architecture design researches [20, 1] that the model could be easier to train if there are more channels in shallower layers. However, channel prun.ing favors shallower layers. For from scratch (uniformed), the filters in each layers is reduced by half (eg. reduce conv1_1 from 64 to 32). We can observe that normal setting networks of the same complexity couldn't reach same accuracy either. This consolidates our idea that there is much redundancy in networks while training. However, redundancy can be opt out at inference-time. This maybe an advantage of inference-time acceleration approaches over training-based approaches. Notice that there is a 0.6% gap between the from scratch model and uniformed one, which indicates that there is room for model exploration. Adopting our approach is much faster than training a model from scratch, even for a thin.ner one. Further researches could alleviate our approach to do thin model exploring. 4.1.5 Acceleration for Detection VGG-16 is popular among object detection tasks [42, 41, 33]. We evaluate transfer learning ability of our 2./4. pruned VGG-16, for Faster R-CNN [42] object detections. PASCAL VOC 2007 object detection benchmark [11] contains 5k trainable images and 5k test images. The performance is evaluated by mean Average Precision (mAP). In our experiments, we first perform channel pruning for VGG-16 on the ImageNet. Then we use the pruned model as the pre-trained model for Faster R-CNN. The actual running time of Faster R-CNN is 220ms / im.age. The convolutional layers contributes about 64%. We got actual time of 94ms for 4. acceleration. From Table 5, we observe 0.4% mAP drops of our 2. model, which is not harmful for practice consideration. 4.2. Experiments with Residual Architecture Nets For Multi-path networks [45, 18, 7], we further explore the popular ResNet [18] and latest Xception [7], on Ima.geNet and CIFAR-10. Pruning residual architecture nets is more challenging. These networks are designed for both efficiency and high accuracy. Tensor factorization algorithms [52, 22] have difficult to accelerate these model. Spatially, 1 . 1 convolution is favored, which could hardly be factorized. 4.2.1 ResNet Pruning ResNet complexity uniformly drops on each residual block. Guided by single layer experiments (Sec. 4.1.1), we still prefer reducing shallower layers heavier than deeper ones. Following similar setting as Filter pruning [31], we keep 70% channels for sensitive residual blocks (res5 and blocks close to the position where spatial size <> Table 3. GPU acceleration comparison. We measure forward-pass time per image. Our approach generalizes well on GPU (smaller is better). <> Table 4. Comparisons with training from scratch, under 4. acceleration. Our fine-tuned model outperforms scratch trained counterparts (smaller is better). <> Table 5.Acceleration for Faster R-CNN detection. <> Table 6. 2. acceleration for ResNet-50 on ImageNet, the base.line network is top-5 accuracy is 92.2% (one view). We improve performance with multi-branch enhancement (Sec. 3.3, smaller is better). change, e.g. res3a,res3d). As for other blocks, we keep 30% channels. With multi-branch enhancement, we prune branch 2a more aggressively within each residual block. The remaining channels ratios for branch 2a,branch 2b,branch 2c is 2:4:3 (e.g., Given 30%, we keep 40%, 80%, 60% respectively). We evaluate performance of multi-branch variants of our approach (Sec. 3.3). From Table 6, we improve 4.0% with our multi-branch enhancement. This is because we accounted the accumulated error from shortcut connection which could broadcast to every layer after it. And the large input feature map width at the entry of each residual block is well reduced by our feature map sampling. <> Table 7. Comparisons for Xception-50, under 2. acceleration ra.tio. The baseline network is top-5 accuracy is 92.8%. Our approach outperforms previous approaches. Most structured simplification methods are not effective on Xception architecture (smaller is better). 4.2.2 Xception Pruning Since computational complexity becomes important in model design, separable convolution has been payed much attention [49, 7]. Xception [7] is already spatially optimized and tensor factorization on 1 . 1 convolutional layer is destructive. Thanks to our approach, it could still be accelerated with graceful degradation. For the ease of comparison, we adopt Xception convolution on ResNet-50, denoted by Xception-50. Based on ResNet-50, we swap all convolutional layers with spatial conv blocks. To keep the same computational complexity, we increase the input channels of all branch2b layers by 2.. The baseline Xception.50 has a top-5 accuracy of 92.8% and complexity of 4450 MFLOPs. We apply multi-branch variants of our approach as de.scribed in Sec. 3.3, and adopt the same pruning ratio setting as ResNet in previous section. Maybe because of Xcep.tion block is unstable, Batch Normalization layers must be maintained during pruning. Otherwise it becomes nontrivial to fine-tune the pruned model. Shown in Table 7, after fine-tuning, we only suffer 1.0% increase of error under 2.. Filter pruning [31] could also apply on Xception, though it is designed for small speed.up ratio. Without fine-tuning, top-5 error is 100%. After training 20 epochs which is like training from scratch, in.creased error reach 4.3%. Our results for Xception-50 are not as graceful as results for VGG-16, since modern net.works tend to have less redundancy by design. <> Table 8. 2. speed-up comparisons for ResNet-56 on CIFAR-10, the baseline accuracy is 92.8% (one view). We outperforms previous approaches and scratch trained counterpart (smaller is better). 4.2.3 Experiments on CIFAR-10 Even though our approach is designed for large datasets, it could generalize well on small datasets. We perform experiments on CIFAR-10 dataset [25], which is favored by many acceleration researches. It consists of 50k images for training and 10k for testing in 10 classes. We reproduce ResNet-56, which has accuracy of 92.8% (Serve as a reference, the official ResNet-56 [18] has ac.curacy of 93.0%). For 2. acceleration, we follow similar setting as Sec. 4.2.1 (keep the final stage unchanged, where the spatial size is 8 . 8). Shown in Table 8, our approach is competitive with scratch trained one, without fine-tuning, under 2. speed-up. After fine-tuning, our result is significantly better than Filter pruning [31] and scratch trained one. 5. Conclusion To conclude, current deep CNNs are accurate with high inference costs. In this paper, we have presented an inference-time channel pruning method for very deep net.works. The reduced CNNs are inference efficient networks while maintaining accuracy, and only require off-the-shelf libraries. Compelling speed-ups and accuracy are demonstrated for both VGG Net and ResNet-like networks on Im.ageNet, CIFAR-10 and PASCAL VOC. In the future, we plan to involve our approaches into training time, instead of inference time only, which may also accelerate training procedure. References [1] J. M. Alvarez and M. Salzmann. Learning the number of neurons in deep networks. In Advances in Neural Information Processing Systems, pages 2262fi2270, 2016. 1, 2, 3, 6 [2] S. Anwar, K. Hwang, and W. Sung. Structured prun.ing of deep convolutional neural networks. arXiv preprint arXiv:1512.08571, 2015. 2 [3] S. Anwar and W. Sung. Compact deep convolutional neural networks with coarse pruning. arXiv preprint arXiv:1610.09639, 2016. 1, 2 [4] H. Bagherinezhad, M. Rastegari, and A. Farhadi. Lcnn: Lookup-based convolutional neural network. arXiv preprint arXiv:1611.06473, 2016. 2 [5] L. Breiman. Better subset regression using the nonnegative garrote. Technometrics, 37(4):373fi384, 1995. 3 [6] S. Chetlur, C. Woolley, P. Vandermersch, J. Cohen, J. Tran, B. Catanzaro, and E. Shelhamer. cudnn: Efficient primitives for deep learning. CoRR, abs/1410.0759, 2014. 6 [7] F. Chollet. Xception: Deep learning with depthwise separa.ble convolutions. arXiv preprint arXiv:1610.02357, 2016. 1, 2, 3, 4, 6, 7 [8] M. Courbariaux and Y. Bengio. Binarynet: Training deep neural networks with weights and activations constrained to+ 1 or-1. arXiv preprint arXiv:1602.02830, 2016. 1, 2 [9] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 248fi255. IEEE, 2009. 4 [10] E. L. Denton, W. Zaremba, J. Bruna, Y. LeCun, and R. Fer.gus. Exploiting linear structure within convolutional net.works for efficient evaluation. In Advances in Neural In.formation Processing Systems, pages 1269fi1277, 2014. 2 [11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results. http://www.pascal.network.org/challenges/VOC/voc2007/workshop/index.html. 4, 6 [12] R. Girshick. Fast r-cnn. In Proceedings of the IEEE Inter.national Conference on Computer Vision, pages 1440fi1448, 2015. 2 [13] Y. Gong, L. Liu, M. Yang, and L. Bourdev. Compress.ing deep convolutional networks using vector quantization. arXiv preprint arXiv:1412.6115, 2014. 2 [14] Y. Guo, A. Yao, and Y. Chen. Dynamic network surgery for efficient dnns. In Advances In Neural Information Process.ing Systems, pages 1379fi1387, 2016. 2 [15] S. Han, X. Liu, H. Mao, J. Pu, A. Pedram, M. A. Horowitz, and W. J. Dally. Eie: efficient inference engine on com.pressed deep neural network. In Proceedings of the 43rd International Symposium on Computer Architecture, pages 243fi254. IEEE Press, 2016. 2 [16] S. Han, H. Mao, and W. J. Dally. Deep compression: Com.pressing deep neural network with pruning, trained quantiza.tion and huffman coding. CoRR, abs/1510.00149, 2, 2015. 2 [17] S. Han, J. Pool, J. Tran, and W. Dally. Learning both weights and connections for efficient neural network. In Advances in Neural Information Processing Systems, pages 1135fi1143, 2015. 1, 2, 3 [18] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learn.ing for image recognition. arXiv preprint arXiv:1512.03385, 2015. 1,2,3,4,6,8 [19] H. Hu, R. Peng, Y.-W. Tai, and C.-K. Tang. Network trim.ming: A data-driven neuron pruning approach towards effi.cient deep architectures. arXiv preprint arXiv:1607.03250, 2016. 2 [20] J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara, A. Fathi, I. Fischer, Z. Wojna, Y. Song, S. Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. arXiv preprint arXiv:1611.10012, 2016. 6 [21] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. 4 [22] M. Jaderberg, A. Vedaldi, and A. Zisserman. Speeding up convolutional neural networks with low rank expansions. arXiv preprint arXiv:1405.3866, 2014. 1, 2, 5, 6, 7 [23] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Gir.shick, S. Guadarrama, and T. Darrell. Caffe: Convolu.tional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014. 4, 6 [24] Y.-D. Kim, E. Park, S. Yoo, T. Choi, L. Yang, and D. Shin. Compression of deep convolutional neural networks for fast and low power mobile applications. arXiv preprint arXiv:1511.06530, 2015. 2 [25] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. 2009. 4, 8 [26] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097fi1105, 2012. 2, 3 [27] A. Lavin. Fast algorithms for convolutional neural networks. arXiv preprint arXiv:1509.09308, 2015. 2 [28] V. Lebedev, Y. Ganin, M. Rakhuba, I. Oseledets, and V. Lempitsky. Speeding-up convolutional neural net.works using fine-tuned cp-decomposition. arXiv preprint arXiv:1412.6553, 2014. 2 [29] V. Lebedev and V. Lempitsky. Fast convnets using group-wise brain damage. arXiv preprint arXiv:1506.02515, 2015. 2 [30] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceed.ings of the IEEE, 86(11):2278fi2324, 1998. 2, 3 [31] H. Li, A. Kadav, I. Durdanovic, H. Samet, and H. P. Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710,2016. 1,2,4,5,6,7,8 [32] B. Liu, M. Wang, H. Foroosh, M. Tappen, and M. Pensky. Sparse convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni.tion, pages 806fi814, 2015. 2 [33] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. E. Reed, C. Fu, and A. C. Berg. SSD: single shot multibox detector. CoRR, abs/1512.02325, 2015. 6 [34] Z. Mariet and S. Sra. Diversity networks. arXiv preprint arXiv:1511.05077, 2015. 2 [35] M. Mathieu, M. Henaff, and Y. LeCun. Fast training of convolutional networks through ffts. arXiv preprint arXiv:1312.5851, 2013. 2 [36] V. Nair and G. E. Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807fi814, 2010. 4 [37] J. Nickolls, I. Buck, M. Garland, and K. Skadron. Scalable parallel programming with CUDA. ACM Queue, 6(2):40fi53, 2008. 6 [38] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Ma.chine learning in Python. Journal of Machine Learning Re.search, 12:2825fi2830, 2011. 4 [39] A. Polyak and L. Wolf. Channel-level acceleration of deep face representations. IEEE Access, 3:2163fi2175, 2015. 2 [40] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnor.net: Imagenet classification using binary convolutional neu.ral networks. In European Conference on Computer Vision, pages 525fi542. Springer, 2016. 2 [41] J. Redmon, S. K. Divvala, R. B. Girshick, and A. Farhadi. You only look once: Unified, real-time object detection. CoRR, abs/1506.02640, 2015. 6 [42] S. Ren, K. He, R. B. Girshick, and J. Sun. Faster R-CNN: towards real-time object detection with region proposal net.works. CoRR, abs/1506.01497, 2015. 6 [43] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 4, 5, 6 [44] S. Srinivas and R. V. Babu. Data-free parameter pruning for deep neural networks. arXiv preprint arXiv:1507.06149, 2015. 2 [45] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1fi9, 2015. 1, 3, 6 [46] R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B (Methodological), pages 267fi288, 1996. 3 [47] N. Vasilache, J. Johnson, M. Mathieu, S. Chintala, S. Pi.antino, and Y. LeCun. Fast convolutional nets with fbfft: A gpu performance evaluation. arXiv preprint arXiv:1412.7580, 2014. 1, 2 [48] W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li. Learning structured sparsity in deep neural networks. In Advances In Neural Information Processing Systems, pages 2074fi2082, 2016. 1, 2, 3 [49] S. Xie, R. Girshick, P. Dollfiar, Z. Tu, and K. He. Aggregated residual transformations for deep neural networks. arXiv preprint arXiv:1611.05431, 2016. 7 [50] J. Xue, J. Li, and Y. Gong. Restructuring of deep neural network acoustic models with singular value decomposition. In INTERSPEECH, pages 2365fi2369, 2013. 2 [51] T.-J. Yang, Y.-H. Chen, and V. Sze. Designing energy-efficient convolutional neural networks using energy-aware pruning. arXiv preprint arXiv:1611.05128, 2016. 2 [52] X. Zhang, J. Zou, K. He, and J. Sun. Accelerating very deep convolutional networks for classification and detection. IEEE transactions on pattern analysis and machine intelli.gence, 38(10):1943fi1955, 2016. 1, 2, 3, 5, 6, 7 <> <> <> <> <> <> Convex Neural Networks Yoshua Bengio, Nicolas Le Roux, Pascal Vincent, Olivier Delalleau, Patrice Marcotte Dept. IRO, Universite de Montr´ eal´ P.O. Box 6128, Downtown Branch, Montreal, H3C 3J7, Qc, Canada fbengioy,lerouxni,vincentp,delallea,marcotteg@iro.umontreal.ca Abstract Convexity has recently received a lot of attention in the machine learning community, and the lack of convexity has been seen as a major disad- vantage of many learning algorithms, such as multi-layer artificial neural networks. We show that training multi-layer neural networks in which the number of hidden units is learned can be viewed as a convex optimization problem. This problem involves an infinite number of variables, but can be solved by incrementally inserting a hidden unit at a time, each time finding a linear classifier that minimizes a weighted sum of errors. 1 Introduction The objective of this paper is not to present yet another learning algorithm, but rather to point to a previously unnoticed relation between multi-layer neural networks (NNs),Boosting (Fre- und and Schapire, 1997) and convex optimization. Its main contributions concern the mathe- matical analysis of an algorithm that is similar to previously proposed incremental NNs, with L1 regularization on the output weights. This analysis helps to understand the underlying convex optimization problem that one is trying to solve. This paper was motivated by the unproven conjecture (based on anecdotal experience) that when the number of hidden units is “large”, the resulting average error is rather insensitive to the random initialization of the NN parameters. One way to justify this assertion is that to re- ally stay stuck in a local minimum, one must have second derivatives positive simultaneously in all directions. When the number of hidden units is large, it seems implausible for none of them to offer a descent direction. Although this paper does not prove or disprove the above conjecture, in trying to do so we found an interesting characterization of the optimization problem for NNs as a convex program if the output loss function is convex in the NN out- put and if the output layer weights are regularized by a convex penalty. More specifically, if the regularization is the L1 norm of the output layer weights, then we show that a “rea- sonable” solution exists, involving a finite number of hidden units (no more than the number of examples, and in practice typically much less). We present a theoretical algorithm that is reminiscent of Column Generation (Chvatal, 1983), in which hidden neurons are inserted ´ one at a time. Each insertion requires solving a weighted classification problem, very much like in Boosting (Freund and Schapire, 1997) and in particular Gradient Boosting (Mason et al., 2000; Friedman, 2001). Neural Networks, Gradient Boosting, and Column Generation Denote x~2Rd+1 the extension of vector x2Rd with one element with value 1. What we call “Neural Network” (NN) here is a predictor for supervised learning of the form <> where x is an input vector, <> is obtained from a linear dis- criminant function hi <> with e.g. <>, or <> or <>. A learning algorithm must specify how to select m, the <> i ’s and the vi ’s. The classical solution (Rumelhart, Hinton and Williams, 1986) involves (a) selecting a loss function Q(^y;y)that specifies how to penalize for mismatches between y^(x)and the ob- served y’s (target output or target class), (b) optionally selecting a regularization penalty that favors “small” parameters, and (c) choosing a method to approximately minimize the sum of the losses on the training data D=f(x1 ;y 1 );:::;(xn ;y n )gplus the regularization penalty. Note that in this formulation, an output non-linearity can still be used, by inserting it in the loss function Q. Examples of such loss functions are the quadratic loss jjy^yjj 2 , the hinge loss <> (used in SVMs), the cross-entropy loss <> (used in logistic regression), and the exponential loss <> (used in Boosting). Gradient Boosting has been introduced in (Friedman, 2001) and (Mason et al., 2000) as a non-parametric greedy-stagewise supervised learning algorithm in which one adds a function at a time to the current solution <>, in a steepest-descent fashion, to form an additive model as above but with the functions hi typically taken in other kinds of sets of functions, such as those obtained with decision trees. In a stagewise approach, when the (m+1)-th basis <> is added, only <> is optimized (by a line search), like in matching pursuit algorithms. Such a greedy-stagewise approach is also at the basis of Boosting algorithms (Freund and Schapire, 1997), which is usually applied using decision trees as bases and Qthe exponential loss. It may be difficult to minimize exactly for wm+1 and hm+1 when the previous bases and weights are fixed, so (Friedman, 2001) proposes to “follow the gradient” in function space, i.e., look for a base learner hm+1 that is best correlated with the gradient of the average loss on the <> (that would be the residue <> in the case of the square loss). The algorithm analyzed here also involves maximizing the correlation between Q0 (the derivative of Q with respect to its first argument, evaluated on the training predictions) and the next basis hm+1 . However, we follow a “stepwise”, less greedy, approach, in which all the output weights are optimized at each step, in order to obtain convergence guarantees. Our approach adapts the Column Generation principle (Chvatal, 1983), a decomposition´ technique initially proposed for solving linear programs with many variables and few con- straints. In this framework, active variables, or “columns”, are only generated as they are required to decrease the objective. In several implementations, the column-generation sub- problem is frequently a combinatorial problem for which efficient algorithms are available. In our case, the subproblem corresponds to determining an “optimal” linear classifier. 2 Core Ideas Informally, consider the set Hof all possible hidden unit functions (i.e., of all possible hidden unit weight vectors vi ). Imagine a NN that has all the elements in this set as hidden units. We might want to impose precision limitations on those weights to obtain either a countable or even a finite set. For such a NN, we only need to learn the output weights. If we end up with a finite number of non-zero output weights, we will have at the end an ordinary feedforward NN. This can be achieved by using a regularization penalty on the output weights that yields sparse solutions, such as the L1 penalty. If in addition the loss function is convex in the output layer weights (which is the case of squared error, hinge loss, -tube regression loss, and logistic or softmax cross-entropy), then it is easy to show that the overall training criterion is convex in the parameters (which are now only the output weights). The only problem is that there are as many variables in this convex program as there are elements in the set H, which may be very large (possibly infinite). However, we find that with L1 regularization, a finite solution is obtained, and that such a solution can be obtained by greedily inserting one hidden unit at a time. Furthermore, it is theoretically possible to check that the global optimum has been reached. Definition 2.1.Let Hbe a set of functions from an input space X to R. Elements of H can be understood as “hidden units” in a NN. Let Wbe the Hilbert space of functions from Hto R, with an inner product denoted by <>. An element of W can be understood as the output weights vector in a neural network. Let < R>> the function that maps any element <> of <>. <> can be understood as the vector of activations of hidden units when input x is observed. Let w2 W represent a parameter(the output weights). The NN prediction is denoted <>. Let < RxR>> be a cost function convex in its first argument that takes a scalar prediction y^(x)and a scalar target value y and returns a scalar cost. This is the cost to be minimized on example pair (x;y). Let <> be the training set. Let <> be a convex regularization functional that penalizes for the choice of more “complex” parameters (e.g., <> according to a 1-norm in W, if His countable). We define the convex NN criterion C(H;Q;;D;w)with parameter was follows: <> (1) The following is a trivial lemma, but it is conceptually very important as it is the basis for the rest of the analysis in this paper. Lemma 2.2.The convex NN cost <> is a convex function of w. Proof. <> is convex in w and <<>> is convex in w, by the above construction. C is additive in <> and additive in . Hence C is convex in w. Note that there are no constraints in this convex optimization program, so that at the global minimum all the partial derivatives of C with respect to elements of w cancel. Let jHj be the cardinality of the set H. If it is not finite, it is not obvious that an optimal solution can be achieved in finitely many iterations. Lemma 2.2 says that training NNs from a very large class (with one or more hidden layer) can be seen as convex optimization problems, usually in a very high dimensional space,as long as we allow the number of hidden units to be selected by the learning algorithm. By choosing a regularizer that promotes sparse solutions, we obtain a solution that has a finite number of “active” hidden units (non-zero entries in the output weights vector w). This assertion is proven below, in theorem 3.1, for the case of the hinge loss. However, even if the solution involves a finite number of active hidden units, the convex optimization problem could still be computationally intractable because of the large number of variables involved. One approach to this problem is to apply the principles already suc- cessfully embedded in Gradient Boosting, but more specifically in Column Generation (an optimization technique for very large scale linear programs), i.e., add one hidden unit at a time in an incremental fashion. The important ingredient here is a way to know that we have reached the global optimum, thus not requiring to actually visit all the possible hidden units.We show that this can be achieved as long as we can solve the sub-problem of finding a linear classifier that minimizes the weighted sum of classification errors. This can be done exactly only on low dimensional data sets but can be well approached using weighted linear SVMs, weighted logistic regression, or Perceptron-type algorithms. Another idea (not followed up here) would be to consider first a smaller set H1 , for which the convex problem can be solved in polynomial time, and whose solution can theoretically be selected as initialization for minimizing the criterion <>, with <>, and where H2 may have infinite cardinality (countable or not). In this way we could show that we can find a solution whose cost satisfies <>, i.e., is at least as good as the solution of a more restricted convex optimization problem. The second minimization can be performed with a local descent algorithm, without the necessity to guarantee that the global optimum will be found. 3 Finite Number of Hidden Neurons In this section we consider the special case with <> the hinge loss, and <> regularization, and we show that the global optimum of the convex cost involves at most n+ 1 hidden neurons, using an approach already exploited in (Ratsch, Demiriz and¨ Bennett, 2002) for L1-loss regression Boosting with L1 regularization of output weights. Xn The training criterion is <>. Let us rewrite t=1 this cost function as the constrained optimization problem: <> (C1) <> (C2) Using a standard technique, the above program can be recast as a linear program. Defin- ing <> the vector of Lagrangian multipliers for the constraints C1 , its dual problem (P)takes the form (in the case of a finite number Jof base learners): <> In the case of a finite number Jof base learners, <>. If the number of hidden units is uncountable, then Iis a closed bounded interval of R. Such an optimization problem satisfies all the conditions needed for using Theorem 4.2 from (Hettich and Kortanek, 1993). Indeed: <> it is compact (as a closed bounded interval of <> is a concave function it is even a linear function); <> is convex in <<>> (it is actually linear in <<>>); <> (therefore finite) ( (P)is the largest value of F satisfying the constraints); for every set of n+1 points <>, there exists ~such that <> for <> (one can take <> since K>0). Then, from Theorem 4.2 from (Hettich and Kortanek, 1993), the following theorem holds: Theorem 3.1.The solution of (P) can be attained with constraints C0 and only n+1 constraints C0 (i.e., there exists a subset of n+1 constraints C0 giving rise to the same maximum 1 as when using the whole set of constraints). Therefore, the primal problem associated is the minimization of the cost function of a NN with n+1 hidden neurons. 4 Incremental Convex NN Algorithm In this section we present a stepwise algorithm to optimize a NN, and show that there is a cri- terion that allows to verify whether the global optimum has been reached. This is a specializa- tion of minimizing <>, with <> 1 and <> is the set of soft or hard linear classifiers (depending on choice of s()). Algorithm ConvexNN( D, Q, , s) <> Theorem 4.1.AlgorithmConvexNN Pstops when it reaches the global optimum of <>. Proof.Let wbe the output weights vector when the algorithm stops. Because the set of hidden units Hwe consider is such that when his in H, h is also in H, we can assume all weights to be non-negative. By contradiction, if w0 6=wis the global optimum, with <>, then, since Cis convex in the output weights, for any 2(0;1) , we have <>. For small enough, we can assume all weights in w that are strictly positive to be also strictly positive in w . Let us denote by Ip the set of strictly positive weights in w (and w), by Iz the set of weights set to zero in w but to a non-zero value in w , and by k the difference w;k wk in the weight of hidden unit hk between wand w . We can assume j < 0 for j2Iz , because instead of setting a small positive weight to hj , one can decrease the weight of hj by the same amount, which will give either the same cost, or possibly a lower one when the weight of <> is positive. With o() denoting a quantity such that o()!0 when !0, the difference (w) =XC(w )C(w)can now be written: <> since for i2Ip , thanks to step (7) of the algorithm, we have @C (w) = 0 . Thus the @w inequality <> rewrites into <> which, when !0, yields (note that <> does not depend on ! since j is linear in ): <> (2) i being the optimal classifier chosen in step (5a) or (5c), all hidden units <> verify <> <> <> , contradicting eq. 2. (Mason et al., 2000) prove a related global convergence result for the AnyBoost algorithm, a non-parametric Boosting algorithm that is also similar to Gradient Boosting (Friedman, 2001). Again, this requires solving as a sub-problem an exact minimization to find a function hi 2 H that is maximally correlated with the gradient Q0 on the output. We now show a simple procedure to select a hyperplane with the best weighted classification error. Exact Minimization In step (5a) we are required to find a linear classifier that minimizes the weighted sum of classification errors. Unfortunately, this is an NP-hard problem (w.r.t. d, see theorem 4 in (Marcotte and Savard, 1992)). However, an exact solution can be easily found in O(n3 ) computations for d= 2 inputs. Proposition 4.2.Finding a linear classifier that minimizes the weighted sum of classification error can be achieved in O(n3 )steps when the input dimension is d= 2 . Proof.We want to <> with respect to u and b, the c’s being in <> Consider u fixed and sort the xi ’s according to their dot product with u and denote r the function which maps ito r(i) such that xr(i) is in i-th position in the sort. Depending on P the value of b, we will have n+1 possible sums, respectively <>, <>. It is obvious that those sums only depend on the order of the products <>, <>. When u varies smoothly on the unit circle, as the dot product is a continuous function of its arguments, the changes in the order of the dot products will occur only when there is a pair (i,j) such that <>. Therefore, there are at most as many order changes as there are pairs of different points, i.e., <>. In the case of d=2, we can enumerate all the different angles for which there is a change, namely a1 ;:::;a z with <>. We then need to test at least one <> for each interval a2 i < <>, and also one u for <>, which makes a total of <> possibilities. 2 It is possible to generalize this result in higher dimensions, and as shown in (Marcotte and Savard, 1992), one can achieve <> time. Algorithm 1 Optimal linear classifier search <> Approximate Minimization For data in higher dimensions, the exact minimization scheme to find the optimal linear classifier is not practical. Therefore it is interesting to consider approximate schemes for obtaining a linear classifier with weighted costs. Popular schemes for doing so are the linear SVM (i.e., linear classifier with hinge loss), the logistic regression classifier, and variants of the Perceptron algorithm. In that case, step (5c) of the algorithm is not an exact minimization, and one cannot guarantee that the global optimum will be reached. However, it might be reasonable to believe that finding a linear classifier by minimizing a weighted hinge loss should yield solutions close to the exact minimization. Unfortunately, this is not generally true, as we have found out on a simple toy data set described below. On the other hand, if in step (7) one performs an optimization not only of the output weights wj (ji) but also of the corresponding weight vectors vj , then the algorithm finds a solution close to the global optimum (we could only verify this on 2-D data sets, where the exact solution can be computed easily). It means that at the end of each stage, one first performs a few training iterations of the whole NN (for the hidden units ji) with an ordinary gradient descent mechanism (we used conjugate gradients but stochastic gradient descent would work too), optimizing the wj ’s and the vj ’s, and then one fixes the vj ’s and obtains the optimal wj ’s for these vj ’s (using a convex optimization procedure). In our experiments we used a quadratic Q, for which the optimization of the output weights can be done with a neural network, using the outputs of the hidden layer as inputs. Let us consider now a bit more carefully what it means to tune the v_j’s in step (7). Indeed, changing the weight vector vj of a selected hidden neuron to decrease the cost is equivalent to a change in the output weights w’s. More precisely, consider the step in which the value of vj becomes v0 . This is equivalent to the following operation on the w’s, when wj j is the corresponding output weight value: the output weight associated with the value vj of a hidden neuron is set to 0, and the output weight associated with the value v0 of a hidden j neuron is set to wj . This corresponds to an exchange between two variables in the convex program. We are justified to take any such step as long as it allows us to decrease the cost C(w). The fact that we are simultaneously making such exchanges on all the hidden units when we tune the vj ’s allows us to move faster towards the global optimum. Extension to multiple outputs The multiple outputs case is more involved than the single-output case because it is not P enough to check the condition <>. Consider a new hidden neuron whose output is hi when the input is xi . Let us also denote <> the vector of output weights between the new hidden neuron and the <> output neurons. The gradient with respect to j is <> with <> the value of the j-th output neuron with input <>. This means that if, for a given j, we have <>, moving Pj away from 0 can only increase the cost. Therefore, the right quantity to consider is <>. We must therefore find <>. As before, this sub-problem is not + convex, but it is not as obvious how to approximate it by a convex problem. The stopping P criterion becomes: if there is no j such that <>, then all weights must remain equal to 0 and a global minimum is reached. Experimental Results We performed experiments on the 2-D double moon toy dataset (as used in (Delalleau, Ben- gio and Le Roux, 2005)), to be able to compare with the exact version of the algorithm. In these experiments, <>. The set-up is the following: Select a new linear classifier, either (a) the optimal one or (b) an approximate using logistic regression. Optimize the output weights using a convex optimizer. In case (b), tune both input and output weights by conjugate gradient descent on Cand finally re-optimize the output weights using LASSO regression. Optionally, remove neurons whose output weight has been set to 0. Using the approximate algorithm yielded for 100 training examples an average penalized ( = 1 ) squared error of 17.11 (over 10 runs), an average test classification error of 3.68% and an average number of neurons of 5.5 . The exact algorithm yielded a penalized squared error of 8.09, an average test classification error of 5.3%, and required 3 hidden neurons. A penalty of = 1 was nearly optimal for the exact algorithm whereas a smaller penalty further improved the test classification error of the approximate algorithm. Besides, when running the approximate algorithm for a long time, it converges to a solution whose quadratic error is extremely close to the one of the exact algorithm. 5 Conclusion We have shown that training a NN can be seen as a convex optimization problem, and have analyzed an algorithm that can exactly or approximately solve this problem. We have shown that the solution with the hinge loss involved a number of non-zero weights bounded by the number of examples, and much smaller in practice. We have shown that there exists a stopping criterion to verify if the global optimum has been reached, but it involves solving a sub-learning problem involving a linear classifier with weighted errors, which can be computationally hard if the exact solution is sought, but can be easily implemented for toy data sets (in low dimension), for comparing exact and approximate solutions. The above experimental results are in agreement with our initial conjecture: when there are many hidden units we are much less likely to stall in the optimization procedure, because there are many more ways to descend on the convex cost C(w). They also suggest, based on experiments in which we can compare with the exact sub-problem minimization, that applying Algorithm ConvexNN with an approximate minimization for adding each hidden unit while continuing to tune the previous hidden unit s tends to lead to fast convergence to the global minimum. What can get us stuck in a “local minimum” (in the traditional sense, i.e., of optimizing w’s and v’s together) is simply the inability to find a new hidden unit weight vector that can improve the total cost (fit and regularization term) even if there exists one. Note that as a side-effect of the results presented here, we have a simple way to train P neural networks with hard-threshold hidden units, since increasing <> can be either achieved exactly (at great price) or approximately (e.g. by using a cross-entropy or hinge loss on the corresponding linear classifier). Acknowledgments The authors thank the following for support: NSERC, MITACS, and the Canada Research Chairs. They are also grateful for the feedback and stimulating exchanges with Sam Roweis, Nathan Srebro, and Aaron Courville. References Chvatal, V. (1983).´ Linear Programming. W.H. Freeman. Delalleau, O., Bengio, Y., and Le Roux, N. (2005). Efficient non-parametric function induction in semi-supervised learning. In Cowell, R. and Ghahramani, Z., editors,Proceedings of AIS- TATS’2005, pages 96–103. Freund, Y. and Schapire, R. E. (1997). A decision theoretic generalization of on-line learning and an application to boosting.Journal of Computer and System Science, 55(1):119–139. Friedman, J. (2001). Greedy function approximation: a gradient boosting machine.Annals of Statis- tics, 29:1180. Hettich, R. and Kortanek, K. (1993). Semi-infinite programming: theory, methods, and applications. SIAM Review, 35(3):380–429. Marcotte, P. and Savard, G. (1992). Novel approaches to the discrimination problem.Zeitschrift fr Operations Research (Theory), 36:517–545. Mason, L., Baxter, J., Bartlett, P. L., and Frean, M. (2000). Boosting algorithms as gradient descent. InAdvances in Neural Information Processing Systems 12, pages 512–518. Ratsch, G., Demiriz, A., and Bennett, K. P. (2002). Sparse regression ensembles in infinite and finite¨ hypothesis spaces.Machine Learning. Rumelhart, D., Hinton, G., and Williams, R. (1986). Learning representations by back-propagating errors.Nature, 323:533–536 <> <> <> <> <> <> DEEP COMPRESSION: COMPRESSING DEEP NEURAL NETWORKS WITH PRUNING , T RAINED QUANTIZATION AND HUFFMAN CODING Song Han Stanford University, Stanford, CA 94305, USA songhan@stanford.edu Huizi Mao Tsinghua University, Beijing, 100084, China mhz12@mails.tsinghua.edu.cn William J. Dally Stanford University, Stanford, CA 94305, USA NVIDIA, Santa Clara, CA 95050, USA dally@stanford.edu ABSTRACT Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. To address this limitation, we introduce “deep compression”, a three stage pipeline: pruning, trained quantization and Huffman coding, that work together to reduce the storage requirement of neural networks by 35% to 49% without affecting their accuracy. Our method first prunes the network by learning only the important connections. Next, we quantize the weights to enforce weight sharing, finally, we apply Huffman coding. After the first two steps we retrain the network to fine tune the remaining connections and the quantized centroids. Pruning, reduces the number of connections by 9% to 13%; Quantization then reduces the number of bits that represent each connection from 32 to 5. On the ImageNet dataset, our method reduced the storage required by AlexNet by 35%, from 240MB to 6.9MB, without loss of accuracy. Our method reduced the size of VGG-16 by 49% from 552MB to 11.3MB, again with no loss of accuracy. This allows fitting the model into on-chip SRAM cache rather than off-chip DRAM memory. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. Benchmarked on CPU, GPU and mobile GPU, compressed network has 3% to 4% layerwise speedup and 3% to 7% better energy efficiency. 1 INTRODUCTION Deep neural networks have evolved to the state-of-the-art technique for computer vision tasks (Krizhevsky et al., 2012)(Simonyan & Zisserman, 2014). Though these neural networks are very powerful, the large number of weights consumes considerable storage and memory bandwidth. For example, the AlexNet Caffemodel is over 200MB, and the VGG-16 Caffemodel is over 500MB (BVLC). This makes it difficult to deploy deep neural networks on mobile system. First, for many mobile-first companies such as Baidu and Facebook, various apps are updated via different app stores, and they are very sensitive to the size of the binary files. For example, App Store has the restriction “apps above 100 MB will not download until you connect to Wi-Fi”. As a result, a feature that increases the binary size by 100MB will receive much more scrutiny than one that increases it by 10MB. Although having deep neural networks running on mobile has many great <

> Figure 1: The three stage compression pipeline: pruning, quantization and Huffman coding. Pruning reduces the number of weights by10%, while quantization further improves the compression rate: between27%and31%. Huffman coding gives more compression: between35%and49%. The compression rate already included the meta-data for sparse representation. The compression scheme doesn’t incur any accuracy loss. features such as better privacy, less network bandwidth and real time processing, the large storage overhead prevents deep neural networks from being incorporated into mobile apps. The second issue is energy consumption. Running large neural networks require a lot of memory bandwidth to fetch the weights and a lot of computation to do dot products— which in turn consumes considerable energy. Mobile devices are battery constrained, making power hungry applications such as deep neural networks hard to deploy. Energy consumption is dominated by memory access. Under 45nm CMOS technology, a 32 bit floating point add consumes 0.9PJ, a 32bit SRAM cache access takes 5PJ, while a 32bit DRAM memory access takes 640PJ, which is 3 orders of magnitude of an add operation. Large networks do not fit in on-chip storage and hence require the more costly DRAM accesses. Running a 1 billion connection neural network, for example, at 20fps would require (20Hz)(1G)(640PJ) = 12.8W just for DRAM access - well beyond the power envelope of a typical mobile device. Our goal is to reduce the storage and energy required to run inference on such large networks so they can be deployed on mobile devices. To achieve this goal, we present “deep compression”: a three- stage pipeline (Figure 1) to reduce the storage required by neural network in a manner that preserves the original accuracy. First, we prune the networking by removing the redundant connections, keeping only the most informative connections. Next, the weights are quantized so that multiple connections share the same weight, thus only the codebook (effective weights) and the indices need to be stored. Finally, we apply Huffman coding to take advantage of the biased distribution of effective weights. Our main insight is that, pruning and trained quantization are able to compress the network without interfering each other, thus lead to surprisingly high compression rate. It makes the required storage so small (a few megabytes) that all weights can be cached on chip instead of going to off-chip DRAM which is energy consuming. Based on “deep compression”, the EIE hardware accelerator Han et al. (2016) was later proposed that works on the compressed model, achieving significant speedup and energy efficiency improvement. 2 NETWORK PRUNING Network pruning has been widely studied to compress CNN models. In early work, network pruning proved to be a valid way to reduce the network complexity and over-fitting (LeCun et al., 1989; Hanson & Pratt, 1989; Hassibi et al., 1993; Strom, 1997). Recently Han et al. (2015) pruned state- ¨ of-the-art CNN models with no loss of accuracy. We build on top of that approach. As shown on the left side of Figure 1, we start by learning the connectivity via normal network training. Next, we prune the small-weight connections: all connections with weights below a threshold are removed from the network. Finally, we retrain the network to learn the final weights for the remaining sparse connections. Pruning reduced the number of parameters by9%and13%for AlexNet and VGG-16 model. <

> Figure 2: Representing the matrix sparsity with relative index. Padding filler zero to prevent overflow. <

> Figure 3: Weight sharing by scalar quantization (top) and centroids fine-tuning (bottom). We store the sparse structure that results from pruning using compressed sparse row (CSR) or compressed sparse column (CSC) format, which requires2a+n+1numbers, where a is the number of non-zero elements and n is the number of rows or columns. To compress further, we store the index difference instead of the absolute position, and encode this difference in 8 bits for conv layer and 5 bits for fc layer. When we need an index difference larger than the bound, we the zero padding solution shown in Figure 2: in case when the difference exceeds 8, the largest 3-bit (as an example) unsigned number, we add a filler zero. 3 TRAINED QUANTIZATION AND WEIGHT SHARING Network quantization and weight sharing further compresses the pruned network by reducing the number of bits required to represent each weight. We limit the number of effective weights we need to store by having multiple connections share the same weight, and then fine-tune those shared weights. Weight sharing is illustrated in Figure 3. Suppose we have a layer that has 4 input neurons and 4 output neurons, the weight is a 4x4 matrix. On the top left is the 4x4 weight matrix, and on the bottom left is the 4x4 gradient matrix. The weights are quantized to 4 bins (denoted with 4 colors), all the weights in the same bin share the same value, thus for each weight, we then need to store only a small index into a table of shared weights. During update, all the gradients are grouped by the color and summed together, multiplied by the learning rate and subtracted from the shared centroids from last iteration. For pruned AlexNet, we are able to quantize to 8-bits (256 shared weights) for each CONV layers, and 5-bits (32 shared weights) for each FC layer without any loss of accuracy. To calculate the compression rate, given k clusters, we only need log_2(k) bits to encode the index. In general, for a network with n connections and each connection is represented with b bits, constraining the connections to have only k shared weights will result in a compression rate of: <> (1) For example, Figure 3 shows the weights of a single layer neural network with four input units and four output units. There are4%4 = 16weights originally but there are only4shared weights: similar weights are grouped together to share the same value. Originally we need to store 16 weights each <

> Figure 4: Left: Three different methods for centroids initialization. Right: Distribution of weights (blue) and distribution of codebook before (green cross) and after fine-tuning (red dot). has 32 bits, now we need to store only 4 effective weights (blue, green, red and orange), each has 32 bits, together with 16 2-bit indices giving a compression rate of <> 3.1 WEIGHT SHARING We use k-means clustering to identify the shared weights for each layer of a trained network, so that all the weights that fall into the same cluster will share the same weight. Weights are not shared across layers. We partition n original weights <> into k clusters <>, n%k, so as to minimize the within-cluster sum of squares (WCSS): <> (2) Different from HashNet (Chen et al., 2015) where weight sharing is determined by a hash function before the networks sees any training data, our method determines weight sharing after a network is fully trained, so that the shared weights approximate the original network. 3.2 INITIALIZATION OF SHARED WEIGHTS Centroid initialization impacts the quality of clustering and thus affects the network’s prediction accuracy. We examine three initialization methods: Forgy(random), density-based, and linear initialization. In Figure 4 we plotted the original weights’ distribution of conv3 layer in AlexNet (CDF in blue, PDF in red). The weights forms a bimodal distribution after network pruning. On the bottom it plots the effective weights (centroids) with 3 different initialization methods (shown in blue, red and yellow). In this example, there are 13 clusters. Forgy(random) initialization randomly chooses k observations from the data set and uses these as the initial centroids. The initialized centroids are shown in yellow. Since there are two peaks in the bimodal distribution, Forgy method tend to concentrate around those two peaks. Density-based initialization linearly spaces the CDF of the weights in the y-axis, then finds the horizontal intersection with the CDF, and finally finds the vertical intersection on the x-axis, which becomes a centroid, as shown in blue dots. This method makes the centroids denser around the two peaks, but more scatted than the Forgy method. Linear initialization linearly spaces the centroids between the [min, max] of the original weights. This initialization method is invariant to the distribution of the weights and is the most scattered compared with the former two methods. Larger weights play a more important role than smaller weights (Han et al., 2015), but there are fewer of these large weights. Thus for both Forgy initialization and density-based initialization, very few centroids have large absolute value which results in poor representation of these few large weights. Linear initialization does not suffer from this problem. The experiment section compares the accuracy <

> Figure 5: Distribution for weight (Left) and index (Right). The distribution is biased. of different initialization methods after clustering and fine-tuning, showing that linear initialization works best. 3.3 FEED-FORWARD AND BACK-PROPAGATION The centroids of the one-dimensional k-means clustering are the shared weights. There is one level of indirection during feed forward phase and back-propagation phase looking up the weight table. An index into the shared weight table is stored for each connection. During back-propagation, the gradient for each shared weight is calculated and used to update the shared weight. This procedure is shown in Figure 3. We denote the loss byL, the weight in the ith column and jth row by Wij, the centroid index of element Wij by Iij, the kth centroid of the layer by Ck. By using the indicator function <<1(.)>>, the gradient of the centroids is calculated as: <> (3) 4 HUFFMAN CODING A Huffman code is an optimal prefix code commonly used for lossless data compression(Van Leeuwen, 1976). It uses variable-length codewords to encode source symbols. The table is derived from the occurrence probability for each symbol. More common symbols are represented with fewer bits. Figure 5 shows the probability distribution of quantized weights and the sparse matrix index of the last fully connected layer in AlexNet. Both distributions are biased: most of the quantized weights are distributed around the two peaks; the sparse matrix index difference are rarely above 20. Experiments show that Huffman coding these non-uniformly distributed values saves 20% to 30% of network storage. 5 EXPERIMENTS We pruned, quantized, and Huffman encoded four networks: two on MNIST and two on ImageNet data-sets. The network parameters and accuracy- 1 before and after pruning are shown in Table 1. The compression pipeline saves network storage by 35% to 49% across different networks without loss of accuracy. The total size of AlexNet decreased from 240MB to 6.9MB, which is small enough to be put into on-chip SRAM, eliminating the need to store the model in energy-consuming DRAM memory. Training is performed with the Caffe framework (Jia et al., 2014). Pruning is implemented by adding a mask to the blobs to mask out the update of the pruned connections. Quantization and weight sharing are implemented by maintaining a codebook structure that stores the shared weight, and group-by-index after calculating the gradient of each layer. Each shared weight is updated with all the gradients that fall into that bucket. Huffman coding doesn’t require training and is implemented offline after all the fine-tuning is finished. 5.1 LE NET-300-100 AND LE NET-5 ON MNIST We first experimented on MNIST dataset with LeNet-300-100 and LeNet-5 network (LeCun et al., 1998). LeNet-300-100 is a fully connected network with two hidden layers, with 300 and 100 1 Reference model is from Caffe model zoo, accuracy is measured without data augmentation Table 1: The compression pipeline can save35%to49%parameter storage with no loss of accuracy. <

> Table 2: Compression statistics for LeNet-300-100. P: pruning, Q:quantization, H:Huffman coding. <> Table 3: Compression statistics for LeNet-5. P: pruning, Q:quantization, H:Huffman coding. <> neurons each, which achieves 1.6% error rate on Mnist. LeNet-5 is a convolutional network that has two convolutional layers and two fully connected layers, which achieves 0.8% error rate on Mnist. Table 2 and table 3 show the statistics of the compression pipeline. The compression rate includes the overhead of the codebook and sparse indexes. Most of the saving comes from pruning and quantization (compressed 32%), while Huffman coding gives a marginal gain (compressed 40%) 5.2 ALEX NET ON IMAGE NET We further examine the performance of Deep Compression on the ImageNet ILSVRC-2012 dataset, which has 1.2M training examples and 50k validation examples. We use the AlexNet Caffe model as the reference model, which has 61 million parameters and achieved a top-1 accuracy of 57.2% and a top-5 accuracy of 80.3%. Table 4 shows that AlexNet can be compressed to2:88%of its original size without impacting accuracy. There are 256 shared weights in each CONV layer, which are encoded with 8 bits, and 32 shared weights in each FC layer, which are encoded with only 5 bits. The relative sparse index is encoded with 4 bits. Huffman coding compressed additional 22%, resulting in 35% compression in total. 5.3 VGG-16 ON IMAGE NET With promising results on AlexNet, we also looked at a larger, more recent network, VGG-16 (Si- monyan & Zisserman, 2014), on the same ILSVRC-2012 dataset. VGG-16 has far more convolutional layers but still only three fully-connected layers. Following a similar methodology, we aggressively compressed both convolutional and fully-connected layers to realize a significant reduction in the number of effective weights, shown in Table5. The VGG16 network as a whole has been compressed by49%. Weights in the CONV layers are represented with 8 bits, and FC layers use 5 bits, which does not impact the accuracy. The two largest fully-connected layers can each be pruned to less than 1.6% of their original size. This reduction Table 4: Compression statistics for AlexNet. P: pruning, Q: quantization, H:Huffman coding. <> Table 5: Compression statistics for VGG-16. P: pruning, Q:quantization, H:Huffman coding. <> is critical for real time image processing, where there is little reuse of these layers across images (unlike batch processing). This is also critical for fast object detection algorithms where one CONV pass is used by many FC passes. The reduced layers will fit in an on-chip SRAM and have modest bandwidth requirements. Without the reduction, the bandwidth requirements are prohibitive. 6 DISCUSSIONS 6.1 PRUNING AND QUANTIZATION WORKING TOGETHER Figure 6 shows the accuracy at different compression rates for pruning and quantization together or individually. When working individually, as shown in the purple and yellow lines, accuracy of pruned network begins to drop significantly when compressed below 8% of its original size; accuracy of quantized network also begins to drop significantly when compressed below 8% of its original size. But when combined, as shown in the red line, the network can be compressed to 3% of original size with no loss of accuracy. On the far right side compared the result of SVD, which is inexpensive but has a poor compression rate. The three plots in Figure 7 show how accuracy drops with fewer bits per connection for CONV layers (left), FC layers (middle) and all layers (right). Each plot reports both top-1 and top-5 accuracy. Dashed lines only applied quantization but without pruning; solid lines did both quantization and pruning. There is very little difference between the two. This shows that pruning works well with quantization. Quantization works well on pruned network because unpruned AlexNet has 60 million weights to quantize, while pruned AlexNet has only 6.7 million weights to quantize. Given the same amount of centroids, the latter has less error. <

> Figure 6: Accuracy v.s. compression rate under different compression methods. Pruning and quantization works best when combined. <

> Figure 7: Pruning doesn’t hurt quantization. Dashed: quantization on unpruned network. Solid: quantization on pruned network; Accuracy begins to drop at the same number of quantization bits whether or not the network has been pruned. Although pruning made the number of parameters less, quantization still works well, or even better(3 bits case on the left figure) as in the unpruned network. <

> Figure 8: Accuracy of different initialization methods. Left: top-1 accuracy. Right: top-5 accuracy. Linear initialization gives best result. The first two plots in Figure 7 show that CONV layers require more bits of precision than FC layers. For CONV layers, accuracy drops significantly below 4 bits, while FC layer is more robust: not until 2 bits did the accuracy drop significantly. 6.2 CENTROID INITIALIZATION Figure 8 compares the accuracy of the three different initialization methods with respect to top-1 accuracy (Left) and top-5 accuracy (Right). The network is quantized to2%8bits as shown on x-axis. Linear initialization outperforms the density initialization and random initialization in all cases except at 3 bits. The initial centroids of linear initialization spread equally across the x-axis, from the min value to the max value. That helps to maintain the large weights as the large weights play a more important role than smaller ones, which is also shown in network pruning Han et al. (2015). Neither random nor density-based initialization retains large centroids. With these initialization methods, large weights are clustered to the small centroids because there are few large weights. In contrast, linear initialization allows large weights a better chance to form a large centroid. <

> Figure 9: Compared with the original network, pruned network layer achieved 3% speedup on CPU, 3.5% on GPU and 4.2% on mobile GPU on average. Batch size = 1 targeting real time processing. Performance number normalized to CPU. <

> Figure 10: Compared with the original network, pruned network layer takes 7% less energy on CPU, 3.3% less on GPU and 4.2% less on mobile GPU on average. Batch size = 1 targeting real time processing. Energy number normalized to CPU. 6.3 SPEEDUP AND ENERGY EFFICIENCY Deep Compression is targeting extremely latency-focused applications running on mobile, which requires real-time inference, such as pedestrian detection on an embedded processor inside an autonomous vehicle. Waiting for a batch to assemble significantly adds latency. So when bench- marking the performance and energy efficiency, we consider the case when batch size = 1. The cases of batching are given in Appendix A. Fully connected layer dominates the model size (more than90%) and got compressed the most by Deep Compression (96%weights pruned in VGG-16). In state-of-the-art object detection algorithms such as fast R-CNN (Girshick, 2015), up to 38% computation time is consumed on FC layers on uncompressed model. So it’s interesting to benchmark on FC layers, to see the effect of Deep Compression on performance and energy. Thus we setup our benchmark on FC6, FC7, FC8 layers of AlexNet and VGG-16. In the non-batched case, the activation matrix is a vector with just one column, so the computation boils down to dense / sparse matrix-vector multiplication for original / pruned model, respectively. Since current BLAS library on CPU and GPU doesn’t support indirect look-up and relative indexing, we didn’t benchmark the quantized model. We compare three different off-the-shelf hardware: the NVIDIA GeForce GTX Titan X and the Intel Core i7 5930K as desktop processors (same package as NVIDIA Digits Dev Box) and NVIDIA Tegra K1 as mobile processor. To run the benchmark on GPU, we used cuBLAS GEMV for the original dense layer. For the pruned sparse layer, we stored the sparse matrix in in CSR format, and used cuSPARSE CSRMV kernel, which is optimized for sparse matrix-vector multiplication on GPU. To run the benchmark on CPU, we used MKL CBLAS GEMV for the original dense model and MKL SPBLAS CSRMV for the pruned sparse model. To compare power consumption between different systems, it is important to measure power at a consistent manner (NVIDIA, b). For our analysis, we are comparing pre-regulation power of the entire application processor (AP) / SOC and DRAM combined. On CPU, the benchmark is running on single socket with a single Haswell-E class Core i7-5930K processor. CPU socket and DRAM power are as reported by the pcm-power utility provided by Intel. For GPU, we used nvidia-smi utility to report the power of Titan X. For mobile GPU, we use a Jetson TK1 development board and measured the total power consumption with a power-meter. We assume 15% AC to DC conversion loss,85% regulator efficiency and 15% power consumed by peripheral components (NVIDIA, a) to report the AP+DRAM power for Tegra K1. Table 6: Accuracy of AlexNet with different aggressiveness of weight sharing and quantization. 8/5 bit quantization has no loss of accuracy; 8/4 bit quantization, which is more hardware friendly, has negligible loss of accuracy of 0.01%; To be really aggressive, 4/2 bit quantization resulted in 1.99% and 2.60% loss of accuracy. <

> The ratio of memory access over computation characteristic with and without batching is different. When the input activations are batched to a matrix the computation becomes matrix-matrix multipli- cation, where locality can be improved by blocking. Matrix could be blocked to fit in caches and reused efficiently. In this case, the amount of memory access isO(n2 ), and that of computation is O(n3 ), the ratio between memory access and computation is in the order of1=n. In real time processing when batching is not allowed, the input activation is a single vector and the computation is matrix-vector multiplication. In this case, the amount of memory access isO(n2 ), and the computation isO(n2 ), memory access and computation are of the same magnitude (as opposed to1=n). That indicates MV is more memory-bounded than MM. So reducing the memory footprint is critical for the non-batching case. Figure 9 illustrates the speedup of pruning on different hardware. There are 6 columns for each benchmark, showing the computation time of CPU / GPU / TK1 on dense / pruned network. Time is normalized to CPU. When batch size = 1, pruned network layer obtained 3% to 4% speedup over the dense network on average because it has smaller memory footprint and alleviates the data transferring overhead, especially for large matrices that are unable to fit into the caches. For example VGG16’s FC6 layer, the largest layer in our experiment, contains 400MB data, which is far from the capacity of L3 cache. In those latency-tolerating applications, batching improves memory locality, where weights could be blocked and reused in matrix-matrix multiplication. In this scenario, pruned network no longer shows its advantage. We give detailed timing results in Appendix A. Figure 10 illustrates the energy efficiency of pruning on different hardware. We multiply power consumption with computation time to get energy consumption, then normalized to CPU to get energy efficiency. When batch size = 1, pruned network layer consumes 3% to 7% less energy over the dense network on average. Reported by nvidia-smi, GPU utilization is 99% for both dense and sparse cases. 6.4 RATIO OF WEIGHTS, INDEX AND CODEBOOK Pruning makes the weight matrix sparse, so extra space is needed to store the indexes of non-zero elements. Quantization adds storage for a codebook. The experiment section has already included these two factors. Figure 11 shows the breakdown of three different components when quantizing four networks. Since on average both the weights and the sparse indexes are encoded with 5 bits, their storage is roughly half and half. The overhead of codebook is very small and often negligible. <

> Figure 11: Storage ratio of weight, index and codebook. Table 7: Comparison with other compression methods on AlexNet. (Collins & Kohli, 2014) reduced the parameters by 4% and with inferior accuracy. Deep Fried Conv nets(Yang et al., 2014) worked on fully connected layers and reduced the parameters by less than 4%. SVD save parameters but suffers from large accuracy loss as much as 2%. Network pruning (Han et al., 2015) reduced the parameters by 9%, not including index overhead. On other networks similar to AlexNet, (Denton et al., 2014) exploited linear structure of conv nets and compressed the network by 2.4% to 13.4% layer wise, with 0.9% accuracy loss on compressing a single layer. (Gong et al., 2014) experimented with vector quantization and compressed the network by 16% to 24%, incurring 1% accuracy loss. <

> 7 RELATED WORK Neural networks are typically over-parametrized, and there is significant redundancy for deep learning models(Denil et al., 2013). This results in a waste of both computation and memory usage. There have been various proposals to remove the redundancy: Vanhoucke et al. (2011) explored a fixed- point implementation with 8-bit integer (vs 32-bit floating point) activations. Hwang & Sung (2014) proposed an optimization method for the fixed-point network with ternary weights and 3-bit activations. Anwar et al. (2015) quantized the neural network using L2 error minimization and achieved better accuracy on MNIST and CIFAR-10 datasets.Denton et al. (2014) exploited the linear structure of the neural network by finding an appropriate low-rank approximation of the parameters and keeping the accuracy within 1% of the original model. The empirical success in this paper is consistent with the theoretical study of random-like sparse networks with +1/0/-1 weights (Arora et al., 2014), which have been proved to enjoy nice properties (e.g. reversibility), and to allow a provably polynomial time algorithm for training. Much work has been focused on binning the network parameters into buckets, and only the values in the buckets need to be stored. HashedNets(Chen et al., 2015) reduce model sizes by using a hash function to randomly group connection weights, so that all connections within the same hash bucket share a single parameter value. In their method, the weight binning is pre-determined by the hash function, instead of being learned through training, which doesn’t capture the nature of images. Gong et al. (2014) compressed deep conv nets using vector quantization, which resulted in 1% accuracy loss. Both methods studied only the fully connected layer, ignoring the convolutional layers. There have been other attempts to reduce the number of parameters of neural networks by replacing the fully connected layer with global average pooling. The Network in Network architecture(Lin et al., 2013) and GoogLenet(Szegedy et al., 2014) achieves state-of-the-art results on several benchmarks by adopting this idea. However, transfer learning, i.e. reusing features learned on the ImageNet dataset and applying them to new tasks by only fine-tuning the fully connected layers, is more difficult with this approach. This problem is noted by Szegedy et al. (2014) and motivates them to add a linear layer on the top of their networks to enable transfer learning. Network pruning has been used both to reduce network complexity and to reduce over-fitting. An early approach to pruning was biased weight decay (Hanson & Pratt, 1989). Optimal Brain Damage (LeCun et al., 1989) and Optimal Brain Surgeon (Hassibi et al., 1993) prune networks to reduce the number of connections based on the Hessian of the loss function and suggest that such pruning is more accurate than magnitude-based pruning such as weight decay. A recent work (Han et al., 2015) successfully pruned several state of the art large scale networks and showed that the number of parameters could be reduce by an order of magnitude. There are also attempts to reduce the number of activations for both compression and acceleration Van Nguyen et al. (2015). 8 FUTURE WORK While thE pruned network has been benchmarked on various hardware, the quantized network with weight sharing has not, because off-the-shelf cuSPARSE or MKL SPBLAS library does not support indirect matrix entry lookup, nor is the relative index in CSC or CSR format supported. So the full advantage of Deep Compression that fit the model in cache is not fully unveiled. A software solution is to write customized GPU kernels that support this. A hardware solution is to build custom ASIC architecture specialized to traverse the sparse and quantized network structure, which also supports customized quantization bit width. We expect this architecture to have energy dominated by on-chip SRAM access instead of off-chip DRAM access. 9 CONCLUSION We have presented “Deep Compression” that compressed neural networks without affecting accuracy. Our method operates by pruning the unimportant connections, quantizing the network using weight sharing, and then applying Huffman coding. We highlight our experiments on AlexNet which reduced the weight storage by 35% without loss of accuracy. We show similar results for VGG-16 and LeNet networks compressed by 49% and 39% without loss of accuracy. This leads to smaller storage requirement of putting conv nets into mobile app. After Deep Compression the size of these networks fit into on-chip SRAM cache (5pJ/access) rather than requiring off-chip DRAM memory (640pJ/access). This potentially makes deep neural networks more energy efficient to run on mobile. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. REFERENCES Anwar, Sajid, Hwang, Kyuyeon, and Sung, Wonyong. Fixed point optimization of deep convolutional neural networks for object recognition. InAcoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on, pp. 1131–1135. IEEE, 2015. Arora, Sanjeev, Bhaskara, Aditya, Ge, Rong, and Ma, Tengyu. Provable bounds for learning some deep representations. InProceedings of the 31th International Conference on Machine Learning, ICML 2014, pp. 584–592, 2014. BVLC. Caffe model zoo. URLhttp://caffe.berkeleyvision.org/model_zoo. Chen, Wenlin, Wilson, James T., Tyree, Stephen, Weinberger, Kilian Q., and Chen, Yixin. Compress- ing neural networks with the hashing trick.arXiv preprint arXiv:1504.04788, 2015. Collins, Maxwell D and Kohli, Pushmeet. Memory bounded deep convolutional networks.arXiv preprint arXiv:1412.1442, 2014. Denil, Misha, Shakibi, Babak, Dinh, Laurent, de Freitas, Nando, et al. Predicting parameters in deep learning. InAdvances in Neural Information Processing Systems, pp. 2148–2156, 2013. Denton, Emily L, Zaremba, Wojciech, Bruna, Joan, LeCun, Yann, and Fergus, Rob. Exploiting linear structure within convolutional networks for efficient evaluation. InAdvances in Neural Information Processing Systems, pp. 1269–1277, 2014. Girshick, Ross. Fast r-cnn.arXiv preprint arXiv:1504.08083, 2015. Gong, Yunchao, Liu, Liu, Yang, Ming, and Bourdev, Lubomir. Compressing deep convolutional networks using vector quantization.arXiv preprint arXiv:1412.6115, 2014. Han, Song, Pool, Jeff, Tran, John, and Dally, William J. Learning both weights and connections for efficient neural networks. InAdvances in Neural Information Processing Systems, 2015. Han, Song, Liu, Xingyu, Mao, Huizi, Pu, Jing, Pedram, Ardavan, Horowitz, Mark A, and Dally, William J. EIE: Efficient inference engine on compressed deep neural network.arXiv preprint arXiv:1602.01528, 2016. Hanson, Stephen Jose and Pratt, Lorien Y. Comparing biases for minimal network construction with´ back-propagation. InAdvances in neural information processing systems, pp. 177–185, 1989. Hassibi, Babak, Stork, David G, et al. Second order derivatives for network pruning: Optimal brain surgeon.Advances in neural information processing systems, pp. 164–164, 1993. Hwang, Kyuyeon and Sung, Wonyong. Fixed-point feedforward deep neural network design using weights+ 1, 0, and- 1. InSignal Processing Systems (SiPS), 2014 IEEE Workshop on, pp. 1–6. IEEE, 2014. Jia, Yangqing, Shelhamer, Evan, Donahue, Jeff, Karayev, Sergey, Long, Jonathan, Girshick, Ross, Guadarrama, Sergio, and Darrell, Trevor. Caffe: Convolutional architecture for fast feature embedding.arXiv preprint arXiv:1408.5093, 2014. Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E. Imagenet classification with deep convolutional neural networks. InNIPS, pp. 1097–1105, 2012. LeCun, Yann, Denker, John S, Solla, Sara A, Howard, Richard E, and Jackel, Lawrence D. Optimal brain damage. InNIPs, volume 89, 1989. LeCun, Yann, Bottou, Leon, Bengio, Yoshua, and Haffner, Patrick. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998. Lin, Min, Chen, Qiang, and Yan, Shuicheng. Network in network.arXiv:1312.4400, 2013. NVIDIA. Technical brief: NVIDIA jetson TK1 development kit bringing GPU-accelerated computing to embedded systems, a. URLhttp://www.nvidia.com. NVIDIA. Whitepaper: GPU-based deep learning inference: A performance and power analysis, b. URLhttp://www.nvidia.com/object/white-papers.html. Simonyan, Karen and Zisserman, Andrew. Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014. Strom, Nikko. Phoneme probability estimation with dynamic sparsely connected artificial neural¨ networks.The Free Speech Journal, 1(5):1–41, 1997. Szegedy, Christian, Liu, Wei, Jia, Yangqing, Sermanet, Pierre, Reed, Scott, Anguelov, Dragomir, Erhan, Dumitru, Vanhoucke, Vincent, and Rabinovich, Andrew. Going deeper with convolutions. arXiv preprint arXiv:1409.4842, 2014. Van Leeuwen, Jan. On the construction of huffman trees. InICALP, pp. 382–410, 1976. Van Nguyen, Hien, Zhou, Kevin, and Vemulapalli, Raviteja. Cross-domain synthesis of medical images using efficient location-sensitive deep network. InMedical Image Computing and Computer- Assisted Intervention–MICCAI 2015, pp. 677–684. Springer, 2015. Vanhoucke, Vincent, Senior, Andrew, and Mao, Mark Z. Improving the speed of neural networks on cpus. InProc. Deep Learning and Unsupervised Feature Learning NIPS Workshop, 2011. Yang, Zichao, Moczulski, Marcin, Denil, Misha, de Freitas, Nando, Smola, Alex, Song, Le, and Wang, Ziyu. Deep fried convnets.arXiv preprint arXiv:1412.7149, 2014. A APPENDIX :DETAILED TIMING / POWER REPORTS OF DENSE & SPARSE NETWORK LAYERS Table 8: Average time on different layers. To avoid variance, we measured the time spent on each layer for 4096 input samples, and averaged the time regarding each input sample. For GPU, the time consumed bycudaMallocandcudaMemcpyis not counted. For batch size = 1,gemvis used; For batch size = 64,gemmis used. For sparse case,csrmvandcsrmmis used, respectively. <> Table 9: Power consumption of different layers. We measured the Titan X GPU power with nvidia-smi, Core i7-5930k CPU power withpcm-powerand Tegra K1 mobile GPU power with an external power meter (scaled to AP+DRAM, see paper discussion). During power measurement, we repeated each computation multiple times in order to get stable numbers. On CPU, dense matrix multiplications consume2xenergy than sparse ones because it is accelerated with multi-threading. <> <> <> <> <> <> <> DEEP DOUBLE DESCENT: WHERE BIGGER MODELS AND MORE DATA HURT Preetum Nakkiran Gal Kaplun y Yamini Bansal y Tristan Yang Harvard University Harvard University Harvard University Harvard University Boaz Barak Ilya Sutskever Harvard University OpenAI ABSTRACT We show that a variety of modern deep learning tasks exhibit a “double-descent” phenomenon where, as we increase model size, performance first gets worse and then gets better. Moreover, we show that double descent occurs not just as a function of model size, but also as a function of the number of training epochs. We unify the above phenomena by defining a new complexity measure we call the effective model complexity and conjecture a generalized double descent with respect to this measure. Furthermore, our notion of model complexity allows us to identify certain regimes where increasing (even quadrupling) the number of train samples actually hurts test performance. 1 INTRODUCTION <

> Figure 1:Left:Train and test error as a function of model size, for ResNet18s of varying width on CIFAR-10 with 15% label noise.Right:Test error, shown for varying train epochs. All models trained using Adam for 4K epochs. The largest model (width64) corresponds to standard ResNet18. The bias-variance trade-off is a fundamental concept in classical statistical learning theory (e.g., Hastie et al. (2005)). The idea is that models of higher complexity have lower bias but higher vari- ance. According to this theory, once model complexity passes a certain threshold, models “overfit” with the variance term dominating the test error, and hence from this point onward, increasing model complexity will only decrease performance (i.e., increase test error). Hence conventional wisdom in classical statistics is that, once we pass a certain threshold,“larger models are worse.” However, modern neural networks exhibit no such phenomenon. Such networks have millions of parameters, more than enough to fit even random labels (Zhang et al. (2016)), and yet they perform much better on many tasks than smaller models. Indeed, conventional wisdom among practitioners is that“larger models are better’’ (Krizhevsky et al. (2012), Huang et al. (2018), Szegedy et al. <

> Figure 2:Left:Test error as a function of model size and train epochs. The horizontal line corre- sponds to model-wise double descent–varying model size while training for as long as possible. The vertical line corresponds to epoch-wise double descent, with test error undergoing double-descent as train time increases.RightTrain error of the corresponding models. All models are Resnet18s trained on CIFAR-10 with 15% label noise, data-augmentation, and Adam for up to 4K epochs. (2015), Radford et al. (2019)). The effect of training time on test performance is also up for debate. In some settings, “early stopping” improves test performance, while in other settings training neu- ral networks to zero training error only improves performance. Finally, if there is one thing both classical statisticians and deep learning practitioners agree on is“more data is always better”. In this paper, we present empirical evidence that both reconcile and challenge some of the above “conventional wisdoms.” We show that many deep learning settings have two different regimes. In the under-parameterized regime, where the model complexity is small compared to the number of samples, the test error as a function of model complexity follows the U-like behavior predicted by the classical bias/variance tradeoff. However, once model complexity is sufficiently large to interpolate i.e., achieve (close to) zero training error, then increasing complexity only decreases test error, following the modern intuition of “bigger models are better”. Similar behavior was previously observed in Opper (1995; 2001), Advani & Saxe (2017), Spigler et al. (2018), and Geiger et al. (2019b). This phenomenon was first postulated in generality by Belkin et al. (2018) who named it “double descent”, and demonstrated it for decision trees, random features, and 2-layer neural networks with‘2 loss, on a variety of learning tasks including MNIST and CIFAR-10. Main contributions. We show that double descent is a robust phenomenon that occurs in a variety of tasks, architectures, and optimization methods (see Figure 1 and Section 5; our experiments are summarized in Table A). Moreover, we propose a much more general notion of “double descent” that goes beyond varying the number of parameters. We define the effective model complexity (EMC) of a training procedure as the maximum number of samples on which it can achieve close to zero training error. The EMC depends not just on the data distribution and the architecture of the classifier but also on the training procedure—and in particular increasing training time will increase the EMC. We hypothesize that for many natural models and learning algorithms, double descent occurs as a function of the EMC. Indeed we observe “epoch-wise double descent” when we keep the model fixed and increase the training time, with performance following a classical U-like curve in the underfitting stage (when the EMC is smaller than the number of samples) and then improving with training time once the EMC is sufficiently larger than the number of samples (see Figure 2). As a corollary, early stopping only helps in the relatively narrow parameter regime of critically parameterized models. Sample non-monotonicity. Finally, our results shed light on test performance as a function of the number of train samples. Since the test error peaks around the point where EMC matches the number of samples (the transition from the under- to over-parameterization), increasing the number of samples has the effect of shifting this peak to the right. While in most settings increasing the number of samples decreases error, this shifting effect can sometimes result in a setting wheremore data is worse!For example, Figure 3 demonstrates cases in which increasing the number of samples by a factor of4:5results in worse test performance. Figure 3: Test loss (per-token perplexity) as a function of Transformer model size (embed- ding dimension d model) on language trans- <

> lation (IWSLT‘14 German-to-English). The curve for 18k samples is generally lower than the one for 4k samples, but also shifted to the right, since fitting 18k samples requires a larger model. Thus, for some models, the performance for 18k samples is worse than for 4k samples. 2 OUR RESULTS To state our hypothesis more precisely, we define the notion of effective model complexity. We define a training procedure T to be any procedure that takes as input a set <> of labeled training samples and outputs a classifier <> mapping data to labels. We define the effective model complexity of T (w.r.t. distributionD) to be the maximum number of samples non which T achieves on average <> training error. Definition 1 (Effective Model Complexity)TheEffective Model Complexity(EMC) of a training procedureT, with respect to distribution D and parameter <>, is defined as: <> whereError <> is the mean error of modelMon train samplesS. Our main hypothesis can be informally stated as follows: Hypothesis 1 (Generalized Double Descent hypothesis, informal)For any natural data distribu- tion D, neural-network-based training procedureT, and small <>, if we consider the task of predicting labels based on n samples from D then: Under-parametrized regime.If <> is sufficiently smaller than n, any perturbation of T that increases its effective complexity will decrease the test error. Over-parameterized regime.If <> is sufficiently larger than n, any perturbation of T that increases its effective complexity will decrease the test error. Critically parameterized regime.If <>, then a perturbation of T that increases its effective complexity might decrease or increase the test error. Hypothesis 1 is informal in several ways. We do not have a principled way to choose the parameter <> (and currently heuristically use <>). We also are yet to have a formal specification for “sufficiently smaller” and “sufficiently larger”. Our experiments suggest that there is a critical interval around the interpolation threshold when <>: below and above this interval increasing complexity helps performance, while within this interval it may hurt performance. The width of the critical interval depends on both the distribution and the training procedure in ways we do not yet completely understand. We believe Hypothesis 1 sheds light on the interaction between optimization algorithms, model size, and test performance and helps reconcile some of the competing intuitions about them. The main result of this paper is an experimental validation of Hypothesis 1 under a variety of settings, where we considered several natural choices of datasets, architectures, and optimization algorithms, and we changed the “interpolation threshold” by varying the number of model parameters, the length of training, the amount of label noise in the distribution, and the number of train samples. Model-wise Double Descent.In Section 5, we study the test error of models of increasing size, for a fixed large number of optimization steps. We show that “model-wise double-descent” occurs for various modern datasets (CIFAR-10, CIFAR-100, IWSLT‘14 de-en, with varying amounts of label noise), model architectures (CNNs, ResNets, Transformers), optimizers (SGD, Adam), number of train samples, and training procedures (data-augmentation, and regularization). Moreover, the peak in test error systematically occurs at the interpolation threshold. In particular, we demonstrate realistic settings in which bigger models are worse. Epoch-wise Double Descent.In Section 6, we study the test error of a fixed, large architecture over the course of training. We demonstrate, in similar settings as above, a corresponding peak in test performance when models are trained just long enough to reach <> train error. The test error of a large model first decreases (at the beginning of training), then increases (around the critical regime), then decreases once more (at the end of training)—that is,training longer can correct overfitting. Sample-wise Non-monotonicity.In Section 7, we study the test error of a fixed model and training procedure, for varying number of train samples. Consistent with our generalized double-descent hypothesis, we observe distinct test behavior in the “critical regime”, when the number of samples is near the maximum that the model can fit. This often manifests as a long plateau region, in which taking significantly more data might not help when training to completion (as is the case for CNNs on CIFAR-10). Moreover, we show settings (Transformers on IWSLT‘14 en-de), where this manifests as a peak—and for a fixed architecture and training procedure,more data actually hurts. Remarks on Label Noise.We observe all forms of double descent most strongly in settings with label noise in the train set (as is often the case when collecting train data in the real-world). How- ever, we also show several realistic settings with a test-error peak even without label noise: ResNets (Figure 4a) and CNNs (Figure 20) on CIFAR-100; Transformers on IWSLT‘14 (Figure 8). More- over, all our experiments demonstrate distinctly different test behavior in the critical regime— often manifesting as a “plateau” in the test error in the noiseless case which develops into a peak with added label noise. See Section 8 for further discussion. 3 RELATED WORK Model-wise double descent was first proposed as a general phenomenon by Belkin et al. (2018). Similar behavior had been observed in Opper (1995; 2001), Advani & Saxe (2017), Spigler et al. (2018), and Geiger et al. (2019b). Subsequently, there has been a large body of work studying the double descent phenomenon. A growing list of papers that theoretically analyze it in the tractable setting of linear least squares regression includes Belkin et al. (2019); Hastie et al. (2019); Bartlett et al. (2019); Muthukumar et al. (2019); Bibas et al. (2019); Mitra (2019); Mei & Montanari (2019). Moreover, Geiger et al. (2019a) provide preliminary results for model-wise double descent in con- volutional networks trained on CIFAR-10. Our work differs from the above papers in two crucial aspects: First, we extend the idea of double-descent beyond the number of parameters to incorpo- rate the training procedure under a unified notion of “Effective Model Complexity”, leading to novel insights like epoch-wise double descent and sample non-monotonicity. The notion that increasing train time corresponds to increasing complexity was also presented in Nakkiran et al. (2019). Sec- ond, we provide an extensive and rigorous demonstration of double-descent for modern practices spanning a variety of architectures, datasets optimization procedures. An extended discussion of the related work is provided in Appendix C. 4 EXPERIMENTAL SETUP We briefly describe the experimental setup here; full details are in Appendix B1. We consider three families of architectures: ResNets, standard CNNs, and Transformers.ResNets:We parameterize a family of ResNet18s (He et al. (2016)) by scaling the width (number of filters) of convolutional layers. Specifically, we use layer widths [k;2k;4k;8k] for varying k. The standard ResNet18 corresponds tok= 64. Standard CNNs:We consider a simple family of 5-layer CNNs, with 4 convolutional layers of widths [k;2k;4k;8k] for varying k, and a fully-connected layer. For context, the CNN with width k=64, can reach over 90% test accuracy on CIFAR-10 with data- augmentation.Transformers:We consider the 6 layer encoder-decoder from Vaswani et al. (2017), as implemented by Ott et al. (2019). We scale the size of the network by modifying the embedding dimension d model , and setting the width of the fully-connected layers proportionally (<>). The raw data from our experiments are available at: https://gitlab.com/harvard-machine-learning/double-descent/tree/master For ResNets and CNNs, we train with cross-entropy loss, and the following optimizers: (1) Adam with learning-rate0:0001for 4K epochs; (2) SGD with learning rate/p1 for 500K gradient steps. T We train Transformers for 80K gradient steps, with 10% label smoothing and no drop-out. Label Noise. In our experiments, label noise of probability prefers to training on a samples which have the correct label with probability (<>), and a uniformly random incorrect label otherwise (label noise is sampled only once and not per epoch). Figure 1 plots test error on the noisy distribu- tion, while the remaining figures plot test error with respect to the clean distribution (the two curves are just linear rescaling of one another). 5 MODEL-WISE DOUBLE DESCENT <

> Figure 4:Model-wise double descent for ResNet18s.Trained on CIFAR-100 and CIFAR-10, with varying label noise. Optimized using Adam with LR0:0001for 4K epochs, and data-augmentation. In this section, we study the test error of models of increasing size, when training to completion (for a fixed large number of optimization steps). We demonstrate model-wise double descent across different architectures, datasets, optimizers, and training procedures. The critical region exhibits distinctly different test behavior around the interpolation point and there is often a peak in test error that becomes more prominent in settings with label noise. For the experiments in this section (Figures 4, 5, 6, 7, 8), notice that all modifications which increase the interpolation threshold (such as adding label noise, using data augmentation, and increasing the number of train samples) also correspondingly shift the peak in test error towards larger models. Additional plots showing the early-stopping behavior of these models, and additional experiments showing double descent in settings with no label noise (e.g. Figure 19) are in Appendix E.2. We also observed model-wise double descent for adversarial training, with a prominent robust test error peak even in settings without label noise. See Figure 26 in Appendix E.2. Discussion. Fully understanding the mechanisms behind model-wise double descent in deep neu- ral networks remains an important open question. However, an analog of model-wise double descent occurs even for linear models. A recent stream of theoretical works analyzes this setting (Bartlett et al. (2019); Muthukumar et al. (2019); Belkin et al. (2019); Mei & Montanari (2019); Hastie et al. (2019)). We believe similar mechanisms may be at work in deep neural networks. Informally, our intuition is that for model-sizes at the interpolation threshold, there is effectively only one model that fits the train data and this interpolating model is very sensitive to noise in the <

> Figure 5: Effect of Data Augmentation. 5-layer CNNs on CIFAR10, with and without data- augmentation. Data-augmentation shifts the interpolation threshold to the right, shifting the test error peak accordingly. Optimized using SGD for 500K steps. See Figure 27 for larger models. <

> <

> Figure 6:SGD vs. Adam.5-Layer CNNs Figure 7: Noiseless settings. 5-layer on CIFAR-10 with no label noise, and no CNNs on CIFAR-100 with no label noise; data augmentation. Optimized using SGD note the peak in test error. Trained with for 500K gradient steps, and Adam for 4K SGD and no data augmentation. See Fig- epochs. ure 20 for the early-stopping behavior of these models. train set and/or model mis-specification. That is, since the model is just barely able to fit the train data, forcing it to fit even slightly-noisy or mis-specified labels will destroy its global structure, and result in high test error. (See Figure 28 in the Appendix for an experiment demonstrating this noise sensitivity, by showing that ensembling helps significantly in the critically-parameterized regime). However for over-parameterized models, there are many interpolating models that fit the train set, and SGD is able to find one that “memorizes” (or “absorbs”) the noise while still performing well on the distribution. The above intuition is theoretically justified for linear models. In general, this situation manifests even without label noise for linear models (Mei & Montanari (2019)), and occurs whenever there Figure 8:Transformers on language trans- lation tasks:Multi-head-attention encoder- decoder Transformer model trained for <

> 80k gradient steps with labeled smoothed cross-entropy loss on IWSLT‘14 German- to-English (160K sentences) and WMT‘14 English-to-French (subsampled to 200K sen- tences) dataset. Test loss is measured as per- token perplexity. is model mis-specification between the structure of the true distribution and the model family. We believe this intuition extends to deep learning as well, and it is consistent with our experiments. 6 EPOCH-WISE DOUBLE DESCENT In this section, we demonstrate a novel form of double-descent with respect to training epochs, which is consistent with our unified view of effective model complexity (EMC) and the generalized double descent hypothesis. Increasing the train time increases the EMC—and thus a sufficiently large model transitions from under- to over-parameterized over the course of training. <

> Figure 9:Left:Training dynamics for models in three regimes. Models are ResNet18s on CIFAR10 with 20% label noise, trained using Adam with learning rate0:0001, and data augmentation.Right: Test error over (Model size Epochs). Three slices of this plot are shown on the left. As illustrated in Figure 9, sufficiently large models can undergo a “double descent” behavior where test error first decreases then increases near the interpolation threshold, and then decreases again. In contrast, for “medium sized” models, for which training to completion will only barely reach 0 error, the test error as a function of training time will follow a classical U-like curve where it is better to stop early. Models that are too small to reach the approximation threshold will remain in the “under parameterized” regime where increasing train time monotonically decreases test error. Our experiments (Figure 10) show that many settings of dataset and architecture exhibit epoch-wise double descent, in the presence of label noise. Further, this phenomenon is robust across optimizer variations and learning rate schedules (see additional experiments in Appendix E.1). As in model- wise double descent, the test error peak is accentuated with label noise. Conventional wisdom suggests that training is split into two phases: (1) In the first phase, the net- work learns a function with a small generalization gap (2) In the second phase, the network starts to over-fit the data leading to an increase in test error. Our experiments suggest that this is not the complete picture—in some regimes, the test error decreases again and may achieve a lower value at the end of training as compared to the first minimum (see Fig 10 for 10% label noise). <

> Figure 10:Epoch-wise double descent for ResNet18 and CNN (width=128). ResNets trained using Adam with learning rate0:0001, and CNNs trained with SGD with inverse-square root learning rate. 7 SAMPLE-WISE NON-MONOTONICITY In this section, we investigate the effect of varying the number of train samples, for a fixed model and training procedure. Previously, in model-wise and epoch-wise double descent, we explored behavior in the critical regime, where <>, by varying the EMC. Here, we explore the critical regime by varying the number of train samples n. By increasing n, the same training procedure T can switch from being effectively over-parameterized to effectively under-parameterized. We show that increasing the number of samples has two different effects on the test error vs. model complexity graph. On the one hand, (as expected) increasing the number of samples shrinks the area under the curve. On the other hand, increasing the number of samples also has the effect of “shifting the curve to the right” and increasing the model complexity at which test error peaks. <

> Figure 11: Sample-wise non-monotonicity. These twin effects are shown in Figure 11a. Note that there is a range of model sizes where the effects “cancel out”—and having 4% more train samples does not help test performance when training to completion. Outside the critically-parameterized regime, for sufficiently under- or over- parameterized models, having more samples helps. This phenomenon is corroborated in Figure 12, which shows test error as a function of both model and sample size, in the same setting as Figure 11a. <

> Figure 12:Left:Test Error as a function of model size and number of train samples, for 5-layer CNNs on CIFAR-10 +20% noise. Note the ridge of high test error again lies along the interpolation threshold. Right: Three slices of the left plot, showing the effect of more data for models of different sizes. Note that, when training to completion, more data helps for small and large models, but does not help for near-critically-parameterized models (green). In some settings, these two effects combine to yield a regime of model sizes where more data actually hurts test performance as in Figure 3 (see also Figure 11b). Note that this phenomenon is not unique to DNNs: more data can hurt even for linear models (see Appendix D). 8 CONCLUSION AND DISCUSSION We introduce a generalized double descent hypothesis: models and training procedures exhibit atyp- ical behavior when their Effective Model Complexity is comparable to the number of train samples. We provide extensive evidence for our hypothesis in modern deep learning settings, and show that it is robust to choices of dataset, architecture, and training procedures. In particular, we demon- strate “model-wise double descent” for modern deep networks and characterize the regime where bigger models can perform worse. We also demonstrate “epoch-wise double descent,” which, to the best of our knowledge, has not been previously proposed. Finally, we show that the double descent phenomenon can lead to a regime where training on more data leads to worse test performance. Preliminary results suggest that double descent also holds as we vary the amount of regularization for a fixed model (see Figure 22). We also believe our characterization of the critical regime provides a useful way of thinking for practitioners—if a model and training procedure are just barely able to fit the train set, then small changes to the model or training procedure may yield unexpected behavior (e.g. making the model slightly larger or smaller, changing regularization, etc. may hurt test performance). Early stopping. We note that many of the phenomena that we highlight often do not occur with optimal early-stopping. However, this is consistent with our generalized double descent hypothesis: if early stopping prevents models from reaching0train error then we would not expect to see double- descent, since the EMC does not reach the number of train samples. Further, we show at least one setting where model-wise double descent can still occur even with optimal early stopping (ResNets on CIFAR-100 with no label noise, see Figure 19). We have not observed settings where more data hurts when optimal early-stopping is used. However, we are not aware of reasons which preclude this from occurring. We leave fully understanding the optimal early stopping behavior of double descent as an important open question for future work. Label Noise. In our experiments, we observe double descent most strongly in settings with label noise. However, we believe this effect is not fundamentally about label noise, but rather about model mis-specification. For example, consider a setting where the label noise is not truly random, but rather pseudorandom (with respect to the family of classifiers being trained). In this setting, the performance of the Bayes optimal classifier would not change (since the pseudorandom noise is deterministic, and invertible), but we would observe an identical double descent as with truly random label noise. Thus, we view adding label noise as merely a proxy for making distributions “harder”— i.e. increasing the amount of model mis-specification. Other Notions of Model Complexity. Our notion of Effective Model Complexity is related to classical complexity notions such as Rademacher complexity, but differs in several crucial ways: (1) EMC depends on the true labels of the data distribution, and (2) EMC depends on the training procedure, not just the model architecture. Other notions of model complexity which do not incorporate features (1) and (2) would not suffice to characterize the location of the double-descent peak. Rademacher complexity, for example, is determined by the ability of a model architecture to fit a randomly-labeled train set. But Rademacher complexity and VC dimension are both insufficient to determine the model-wise double descent peak location, since they do not depend on the distribution of labels— and our experiments show that adding label noise shifts the location of the peak. Moreover, both Rademacher complexity and VC dimension depend only on the model family and data distribution, and not on the training procedure used to find models. Thus, they are not capable of capturing train-time double-descent effects, such as “epoch-wise” double descent, and the effect of data-augmentation on the peak location. ACKNOWLEDGMENTS We thank Mikhail Belkin for extremely useful discussions in the early stages of this work. We thank Christopher Olah for suggesting the Model SizeEpoch visualization, which led to the investigation of epoch-wise double descent, as well as for useful discussion and feedback. We also thank Alec Radford, Jacob Steinhardt, and Vaishaal Shankar for helpful discussion and suggestions. P.N. thanks OpenAI, the Simons Institute, and the Harvard Theory Group for a research environment that enabled this kind of work. We thank Dimitris Kalimeris, Benjamin L. Edelman, and Sharon Qian, and Aditya Ramesh for comments on an early draft of this work. This work supported in part by NSF grant CAREER CCF 1452961, BSF grant 2014389, NSF US- ICCS proposal 1540428, a Google Research award, a Facebook research award, a Simons Investiga- tor Award, a Simons Investigator Fellowship, and NSF Awards CCF 1715187, CCF 1565264, CCF 1301976, IIS 1409097, and CNS 1618026. Y.B. would like to thank the MIT-IBM Watson AI Lab for contributing computational resources for experiments. REFERENCES Madhu S Advani and Andrew M Saxe. High-dimensional dynamics of generalization error in neural networks.arXiv preprint arXiv:1710.03667, 2017. Peter L Bartlett, Philip M Long, Gabor Lugosi, and Alexander Tsigler. Benign overfitting in linear´ regression.arXiv preprint arXiv:1906.11300, 2019. Mikhail Belkin, Daniel Hsu, Siyuan Ma, and Soumik Mandal. Reconciling modern machine learning and the bias-variance trade-off.arXiv preprint arXiv:1812.11118, 2018. Mikhail Belkin, Daniel Hsu, and Ji Xu. Two models of double descent for weak features.arXiv preprint arXiv:1903.07571, 2019. Koby Bibas, Yaniv Fogel, and Meir Feder. A new look at an old problem: A universal learning approach to linear regression.arXiv preprint arXiv:1905.04708, 2019. Mauro Cettolo, Christian Girardi, and Marcello Federico. Wit 3 : Web inventory of transcribed and translated talks. InProceedings of the 16 th Conference of the European Association for Machine Translation (EAMT), pp. 261–268, Trento, Italy, May 2012. Mario Geiger, Arthur Jacot, Stefano Spigler, Franck Gabriel, Levent Sagun, Stephane d’Ascoli,´ Giulio Biroli, Clement Hongler, and Matthieu Wyart. Scaling description of generalization with´ number of parameters in deep learning.arXiv preprint arXiv:1901.01608, 2019a. Mario Geiger, Stefano Spigler, Stephane d’Ascoli, Levent Sagun, Marco Baity-Jesi, Giulio Biroli,´ and Matthieu Wyart. Jamming transition as a paradigm to understand the loss landscape of deep neural networks.Physical Review E, 100(1):012115, 2019b. Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples.arXiv preprint arXiv:1412.6572, 2014. Trevor Hastie, Robert Tibshirani, Jerome Friedman, and James Franklin. The elements of statistical learning: data mining, inference and prediction.The Mathematical Intelligencer, 27(2):83–85, 2005. Trevor Hastie, Andrea Montanari, Saharon Rosset, and Ryan J Tibshirani. Surprises in high- dimensional ridgeless least squares interpolation.arXiv preprint arXiv:1903.08560, 2019. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. InEuropean conference on computer vision, pp. 630–645. Springer, 2016. Yanping Huang, Yonglong Cheng, Dehao Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V. Le, and Zhifeng Chen. Gpipe: Efficient training of giant neural networks using pipeline parallelism. CoRR, abs/1811.06965, 2018. URLhttp://arxiv.org/abs/1811.06965. Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo- lutional neural networks. InAdvances in neural information processing systems, pp. 1097–1105, 2012. Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks.arXiv preprint arXiv:1706.06083, 2017. Song Mei and Andrea Montanari. The generalization error of random features regression: Precise asymptotics and double descent curve.arXiv preprint arXiv:1908.05355, 2019. Partha P. Mitra. Understanding overfitting peaks in generalization error: Analytical risk curves for l2 and l1 penalized interpolation.ArXiv, abs/1906.03667, 2019. Vidya Muthukumar, Kailas Vodrahalli, and Anant Sahai. Harmless interpolation of noisy data in regression.arXiv preprint arXiv:1903.09139, 2019. Preetum Nakkiran, Gal Kaplun, Dimitris Kalimeris, Tristan Yang, Benjamin L Edelman, Fred Zhang, and Boaz Barak. Sgd on neural networks learns functions of increasing complexity.arXiv preprint arXiv:1905.11604, 2019. Manfred Opper. Statistical mechanics of learning: Generalization.The Handbook of Brain Theory and Neural Networks, 922-925., 1995. Manfred Opper. Learning to generalize.Frontiers of Life, 3(part 2), pp.763-775., 2001. Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. InProceedings of NAACL-HLT 2019: Demonstrations, 2019. David Page. How to train your resnet. https://myrtle.ai/how-to-train-your-resnet-4-architecture/, 2018. Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. InNeurIPS Autodiff Workshop, 2017. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. InAdvances in neural information processing systems, pp. 1177–1184, 2008. Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units.ArXiv, abs/1508.07909, 2015. Stefano Spigler, Mario Geiger, Stephane d’Ascoli, Levent Sagun, Giulio Biroli, and Matthieu Wyart.´ A jamming transition from under-to over-parametrization affects loss landscape and generaliza- tion.arXiv preprint arXiv:1810.09665, 2018. Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Du- mitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Computer Vision and Pattern Recognition (CVPR), 2015. URLhttp://arxiv.org/abs/ 1409.4842. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need.CoRR, abs/1706.03762, 2017. Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization.ICLR, abs/1611.03530, 2016. A SUMMARY TABLE OF EXPERIMENTAL RESULTS <

> B APPENDIX: EXPERIMENTAL DETAILS B.1 MODELS We use the following families of architectures. The PyTorch Paszke et al. (2017) specification of our ResNets and CNNs are available at https://gitlab.com/harvard-machine-learning/double-descent/tree/master. ResNets. We define a family of ResNet18s of increasing size as follows. We follow the Preac- tivation ResNet18 architecture of He et al. (2016), using 4 ResNet blocks, each consisting of two BatchNorm-ReLU-Convolution layers. The layer widths for the 4 blocks are [k;2k;4k;8k] for varyingk2Nand the strides are [1, 2, 2, 2]. The standard ResNet18 corresponds to k=64 con- volutional channels in the first layer. The scaling of model size withkis shown in Figure 13b. Our implementation is adapted from https://github.com/kuangliu/pytorch-cifar. Standard CNNs. We consider a simple family of 5-layer CNNs, with four Conv-BatchNorm- ReLU-MaxPool layers and a fully-connected output layer. We scale the four convolutional layer widths as [k;2k;4k;8k]. The MaxPool is [1, 2, 2, 8]. For all the convolution layers, the kernel size = 3, stride = 1 and padding=1. This architecture is based on the “backbone” architecture from Page (2018). For k=64, this CNN has 1558026 parameters and can reach >90% test accuracy on CIFAR-10 (Krizhevsky (2009)) with data-augmentation. The scaling of model size with k is shown in Figure 13a. Transformers. We consider the encoder-decoder Transformer model from Vaswani et al. (2017) with 6 layers and 8 attention heads per layer, as implemented by fairseq Ott et al. (2019). We scale the size of the network by modifying the embedding dimension (d model), and scale the width of the fully-connected layers proportionally (dff = 4d model). We train with 10% label smoothing and no drop-out, for 80 gradient steps. <

> Figure 13: Scaling of model size with our parameterization of width & embedding dimension. B.2 IMAGE CLASSIFICATION: EXPERIMENTAL SETUP We describe the details of training for CNNs and ResNets below. Loss function:Unless stated otherwise, we use the cross-entropy loss for all the experiments. Data-augmentation: In experiments where data-augmentation was used, we apply RandomCrop(32, padding=4)andRandomHorizontalFlip. In experiments with added label noise, the label for all augmentations of a given training sample are given the same label. Regularization:No explicit regularization like weight decay or dropout was applied unless explic- itly stated. Initialization:We use the default initialization provided by PyTorch for all the layers. Optimization: Adam: Unless specified otherwise, learning rate was set at constant to 1e^4 and all other parameters were set to their default PyTorch values. SGD: Unless specified otherwise, learning rate schedule inverse-square root (defined be- low) was used with initial learning rate <> and updates every L=512 gradient steps. No momentum was used. We found our results are robust to various other natural choices of optimizers and learning rate schedule. We used the above settings because (1) they optimize well, and (2) they do not require experiment-specific hyperparameter tuning, and allow us to use the same optimization across many experiments. Batch size: All experiments use a batchsize of 128. Learning rate schedule descriptions: Inverse-square root (<>): At gradient stept, the learning rate is set to <>. We set learning-rate with respect to number of gradient steps, and not epochs, <> in order to allow comparison between experiments with varying train-set sizes. Dynamic drop (<>, drop, patience): Starts with an initial learning rate of 0 and drops by a factor of ’drop’ if the training loss has remained constant or become worse for ’patience’ number of gradient steps. B.3 NEURAL MACHINE TRANSLATION: EXPERIMENTAL SETUP Here we describe the experimental setup for the neural machine translation experiments. Training procedure. In this setting, the distributionDconsists of triples <> where V_src and V_tgt are the source and target vocabularies, the stringxis a sentence in the source language,yis its translation in the target language, andiis the index of the token to be predicted by the model. We assume that <> is distributed uniformly on <>. A standard probabilistic model defines an autoregressive factorization of the likelihood: <> Given a set of training samplesS, we define <> In practice,S is not constructed from independent samples from D, but rather by first sampling <<(x,y)>> and then including all <> in S. For training transformers, we replicate the optimization procedure specified in Vaswani et al. (2017) section 5.3, where the learning rate schedule consists of a “warmup” phase with linearly increasing learning rate followed by a phase with inverse square-root decay. We preprocess the data using byte pair encoding (BPE) as described in Sennrich et al. (2015). We use the implementation provided by fairseq (https://github.com/pytorch/fairseq). Datasets. The IWSLT’14 German to English dataset contains TED Talks as described in Cettolo et al. (2012). The WMT’14 English to French dataset is taken from http://www.statmt.org/wmt14/translation-task.html. B.4 PER-SECTION EXPERIMENTAL DETAILS Here we provide full details for experiments in the body, when not otherwise provided. Introduction: Experimental Details Figure 1: All models were trained using Adam with learning- rate 0.0001 for 4K epochs. Plotting means and standard deviations for 5 trials, with random network initialization. Model-wise Double Descent: Experimental Details Figure 7: Plotting means and standard devia- tions for 5 trials, with random network initialization. Sample-wise Nonmonotonicity: Experimental DetailsFigure 11a: All models are trained with SGD for 500K epochs, and data-augmentation. Bottom: Means and standard deviations from 5 trials with random initialization, and random subsampling of the train set. C EXTENDED DISCUSSION OF RELATED WORK Belkin et al. (2018): This paper proposed, in very general terms, that the apparent contradiction between traditional notions of the bias-variance trade-off and empirically successful practices in deep learning can be reconciled under a double-descent curve—as model complexity increases, the test error follows the traditional “U-shaped curve”, but beyond the point of interpolation, the error starts todecrease. This work provides empirical evidence for the double-descent curve with fully connected networks trained on subsets of MNIST, CIFAR10, SVHN and TIMIT datasets. They use thel2 loss for their experiments. They demonstrate that neural networks are not an aberration in this regard—double-descent is a general phenomenon observed also in linear regression with random features and random forests. Theoretical works on linear least squares regression: A variety of papers have attempted to the- oretically analyze this behavior in restricted settings, particularly the case of least squares regression under various assumptions on the training data, feature spaces and regularization method. 1.Advani & Saxe (2017); Hastie et al. (2019) both consider the linear regression problem stated above and analyze the generalization behavior in the asymptotic limit <> using random matrix theory. Hastie et al. (2019) highlight that when the model is mis- specified, the minimum of training error can occur for over-parameterized models 2.Belkin et al. (2019) Linear least squares regression for two data models, where the input data is sampled from a Gaussian and a Fourier series model for functions on a circle. They provide a finite-sample analysis for these two cases 3.Bartlett et al. (2019) provides generalization bounds for the minimuml2 -norm interpolant for Gaussian features 4.Muthukumar et al. (2019) characterize the fundamental limit of of any interpolating solu- tion in the presence of noise and provide some interesting Fourier-theoretic interpretations. 5.Mei & Montanari (2019): This work provides asymptotic analysis for ridge regression over random features Similar double descent behavior was investigated in Opper (1995; 2001) Geiger et al. (2019b) showed that deep fully connected networks trained on the MNIST dataset with hinge loss exhibit a “jamming transition” when the number of parameters exceeds a threshold that allows training to near-zero train loss. Geiger et al. (2019a) provide further experiments on CIFAR- 10 with a convolutional network. They also highlight interesting behavior with ensembling around the critical regime, which is consistent with our informal intuitions in Section 5 and our experiments in Figures 28, 29. Advani & Saxe (2017); Geiger et al. (2019b;a) also point out that double-descent is not observed when optimal early-stopping is used. D RANDOM FEATURES: A CASE STUDY <

> Figure 14:Random Fourier Featureson the Fashion MNIST dataset. The setting is equivalent to two-layer neural network witheix activation, with randomly-initialized first layer that is fixed throughout training. The second layer is trained using gradient flow. In this section, for completeness sake, we show that both the model- and sample-wise double de- scent phenomena are not unique to deep neural networks—they exist even in the setting of Random Fourier Features of Rahimi & Recht (2008). This setting is equivalent to a two-layer neural network with <> activation. The first layer is initialized with aN(0;1 )Gaussian distribution and then fixed throughout training. The width (or embedding dimension) d dof the first layer parameterizes the model size. The second layer is initialized with0s and trained with MSE loss. Figure 14 shows the grid of Test Error as a function of both number of samplesnand model sized. Note that in this settingEMC =d(the embedding dimension). As a result, as demonstrated in the figure, the peak follows the path ofn=d. Both model-wise and sample-wise (see figure 15) double descent phenomena are captured, by horizontally and vertically crossing the grid, respectively. <

> Figure 15: Sample-wise double-descent slice for Random Fourier Features on the Fashion MNIST dataset. In this figure the embedding dimension (number of random features) is 1000. E APPENDIX: ADDITIONAL EXPERIMENTS E.1 EPOCH-WISE DOUBLE DESCENT: ADDITIONAL RESULTS Here, we provide a rigorous evaluation of epoch-wise double descent for a variety of optimizers and learning rate schedules. We train ResNet18 on CIFAR-10 with data-augmentation and 20% label noise with three different optimizers—Adam, SGD, SGD + Momentum (momentum set to 0.9) and three different learning rate schedules—constant, inverse-square root, dynamic drop for differnet values of initial learning rate. We observe that double-descent occurs reliably for all optimizers and learning rate schedules and the peak of the double descent curve shifts with the interpolation point. <

> Figure 16:Epoch-wise double descentfor ResNet18 trained with Adam and multiple learning rate schedules A practical recommendation resulting from epoch-wise double descent is that stopping the training when the test error starts to increase may not always be the best strategy. In some cases, the test error may decrease again after reaching a maximum, and the final value may be lower than the minimum earlier in training. <

> Figure 17:Epoch-wise double descentfor ResNet18 trained with SGD and multiple learning rate schedules <

> Figure 18:Epoch-wise double descentfor ResNet18 trained with SGD+Momentum and multiple learning rate schedules E.2 MODEL-WISE DOUBLE DESCENT: ADDITIONAL RESULTS E.2.1 CLEAN SETTINGS WITH MODEL-WISE DOUBLE DESCENT <

> Figure 19:Top:Train and test performance as a function of both model size and train epochs. Bottom:Test error dynamics of the same model (ResNet18, on CIFAR-100 with no label noise, data-augmentation and Adam optimizer trained for 4k epochs with learning rate 0.0001). Note that even with optimal early stopping this setting exhibits double descent. <

> Figure 20:Top:Train and test performance as a function of both model size and train epochs. Bottom:Test error dynamics of the same models. 5-Layer CNNs, CIFAR-100 with no label noise, no data-augmentation Trained with SGD for 1e6 steps. Same experiment as Figure 7. E.2.2 WEIGHT DECAY <

> Figure 21:Left:Test error dynamics with weight decay of 5e-4 (bottom left) and without weight decay (top left). Right:Test and train error andtest lossfor models with varying amounts of weight decay. All models are 5-Layer CNNs on CIFAR-10 with 10% label noise, trained with data-augmentation and SGD for 500K steps. Here, we now study the effect of varying the level of regularization on test error. We train CIFAR10 with data-augmentation and 20% label noise on ResNet18 for weight decay coefficients <> rang- ing from 0 to 0.1. We train the networks using SGD + inverse-square root learning rate. Figure below shows a picture qualitatively very similar to that observed for model-wise double descent wherein ”model complexity” is now controlled by the regularization parameter. This confirms our generalized double descent hypothesis along yet another axis of Effective Model Complexity. <

> Figure 22: Generalized double descent for weight decay. We found that using the same initial learning rate for all weight decay values led to training instabilities. This resulted in some noise in the Test Error (Weight DecayEpochs) plot shown above. E.2.3 EARLY STOPPING DOES NOT EXHIBIT DOUBLE DESCENT <

> Figure 23: Model-wise test error dynamics for a subsampled IWSLT‘14 dataset. Left: 4k samples, Right: 18k samples. Note that with optimal early-stopping, more samples is always better. <

> Figure 24: Model-wise test error dynamics for a IWSLT‘14 de-en and subsampled WMT‘14 en-fr datasets.Left: IWSLT‘14,Right: subsampled (200k samples) WMT‘14. Note that with optimal early-stopping, the test error is much lower for this task. <

> Figure 25:Top:Train and test performance as a function of both model size and train epochs. Bottom:Test error dynamics of the same model (CNN, on CIFAR-10 with 10% label noise, data-paugmentation and SGD optimizer with learning rate/1= T). E.2.4 TRAINING PROCEDURE <

> Figure 26:Model-wise double descent for adversarial trainingResNet18s on CIFAR-10 (sub- sampled to 25k train samples) with no label noise. We train for L2 robustness of radius <> and <>, using 10-step PGD (Goodfellow et al. (2014); Madry et al. (2017)). Trained using SGD (batch size 128) with learning rate0:1for 400 epochs, then0:01for 400 epochs. <

> Figure 27 E.3 ENSEMBLING <

> Figure 28:Effect of Ensembling (ResNets, 15% label noise). Test error of an ensemble of 5 models, compared to the base models. The ensembled classifier is determined by plurality vote over the 5 base models. Note that emsembling helps most around the critical regime. All models are ResNet18s trained on CIFAR-10 with 15% label noise, using Adam for 4K epochs (same setting as Figure 1). Test error is measured against the original (not noisy) test set, and each model in the ensemble is trained using a train set with independently-sampled 15% label noise. <

> Figure 29:Effect of Ensembling (CNNs, no label noise). Test error of an ensemble of 5 models, compared to the base models. All models are 5-layer CNNs trained on CIFAR-10 with no label noise, using SGD and no data augmentation. (same setting as Figure 7). <> <> <> <> <> <> Deep Residual Learning for Image Recognition Kaiming He Xiangyu Zhang Shaoqing Ren Jian Sun Microsoft Research {kahe, v-xiangz, v-shren, jiansun}@microsoft.com Abstract Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learn.ing residual functions with reference to the layer inputs, in.stead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers 8. deeper than VGG nets [41] but still having lower complex. ity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1 , where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation. 1. Introduction Deep convolutional neural networks [22, 21] have led to a series of breakthroughs for image classification [21, 50, 40]. Deep networks naturally integrate low/mid/high.level features [50] and classifiers in an end-to-end multi.layer fashion, and the levels of features can be enriched by the number of stacked layers (depth). Recent evidence [41, 44] reveals that network depth is of crucial importance, and the leading results [41, 44, 13, 16] on the challenging ImageNet dataset [36] all exploit very deep [41] models, with a depth of sixteen [41] to thirty [16]. Many other non.trivial visual recognition tasks [8, 12, 7, 32, 27] have also <

> Figure 1. Training error (left) and test error (right) on CIFAR-10 with 20-layer and 56-layer plain networks. The deeper network has higher training error, and thus test error. Similar phenomena on ImageNet is presented in Fig. 4. greatly benefited from very deep models. Driven by the significance of depth, a question arises: Is learning better networks as easy as stacking more layers? An obstacle to answering this question was the notorious problem of vanishing/exploding gradients [1, 9], which hamper convergence from the beginning. This problem, however, has been largely addressed by normalized initialization [23, 9, 37, 13] and intermediate normalization layers [16], which enable networks with tens of layers to start con. verging for stochastic gradient descent (SGD) with back-propagation [22]. When deeper networks are able to start converging, a degradation problem has been exposed: with the network depth increasing, accuracy gets saturated (which might be unsurprising) and then degrades rapidly. Unexpectedly, such degradation is not caused by overfitting, and adding more layers to a suitably deep model leads to higher training error, as reported in [11, 42] and thoroughly verified by our experiments. Fig. 1 shows a typical example. The degradation (of training accuracy) indicates that not all systems are similarly easy to optimize. Let us consider a shallower architecture and its deeper counterpart that adds more layers onto it. There exists a solution by construction to the deeper model: the added layers are identity mapping, and the other layers are copied from the learned shallower model. The existence of this constructed solution indicates that a deeper model should produce no higher training error than its shallower counterpart. But experiments show that our current solvers on hand are unable to find solutions that are comparably good or better than the constructed solution (or unable to do so in feasible time). In this paper, we address the degradation problem by introducing a deep residual learning framework. In.stead of hoping each few stacked layers directly fit a desired underlying mapping, we explicitly let these lay.ers fit a residual mapping. Formally, denoting the desired underlying mapping as <>, we let the stacked nonlinear layers fit another mapping of <>. The original mapping is recast into <>. We hypothesize that it is easier to optimize the residual mapping than to optimize the original, unreferenced mapping. To the extreme, if an identity mapping were optimal, it would be easier to push the residual to zero than to fit an identity mapping by a stack of nonlinear layers. The formulation of <> can be realized by feedforward neural networks with shortcut connections (Fig. 2). Shortcut connections [2, 34, 49] are those skipping one or more layers. In our case, the shortcut connections simply perform identity mapping, and their outputs are added to the outputs of the stacked layers (Fig. 2). Identity short.cut connections add neither extra parameter nor computational complexity. The entire network can still be trained end-to-end by SGD with backpropagation, and can be easily implemented using common libraries (e.g., Caffe [19]) without modifying the solvers. We present comprehensive experiments on ImageNet [36] to show the degradation problem and evaluate our method. We show that: 1) Our extremely deep residual nets are easy to optimize, but the counterpart plain nets (that simply stack layers) exhibit higher training error when the depth increases; 2) Our deep residual nets can easily enjoy accuracy gains from greatly increased depth, producing results substantially better than previous networks. Similar phenomena are also shown on the CIFAR-10 set [20], suggesting that the optimization difficulties and the effects of our method are not just akin to a particular dataset. We present successfully trained models on this dataset with over 100 layers, and explore models with over 1000 layers. On the ImageNet classification dataset [36], we obtain excellent results by extremely deep residual nets. Our 152.layer residual net is the deepest network ever presented on ImageNet, while still having lower complexity than VGG nets [41]. Our ensemble has 3.57% top-5 error on the ImageNet test set, and won the 1st place in the ILSVRC 2015 classification competition. The extremely deep representations also have excellent generalization performance on other recognition tasks, and lead us to further win the 1st places on: ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation in ILSVRC & COCO 2015 competitions. This strong evidence shows that the residual learning principle is generic, and we expect that it is applicable in other vision and non-vision problems. 2. Related Work Residual Representations. In image recognition, VLAD [18] is a representation that encodes by the residual vectors with respect to a dictionary, and Fisher Vector [30] can be formulated as a probabilistic version [18] of VLAD. Both of them are powerful shallow representations for image retrieval and classification [4, 48]. For vector quantization, encoding residual vectors [17] is shown to be more effective than encoding original vectors. In low-level vision and computer graphics, for solving Partial Differential Equations (PDEs), the widely used Multigrid method [3] reformulates the system as subproblems at multiple scales, where each subproblem is responsible for the residual solution between a coarser and a finer scale. An alternative to Multigrid is hierarchical basis pre.conditioning [45, 46], which relies on variables that represent residual vectors between two scales. It has been shown [3, 45, 46] that these solvers converge much faster than standard solvers that are unaware of the residual nature of the solutions. These methods suggest that a good reformulation or preconditioning can simplify the optimization. Shortcut Connections. Practices and theories that lead to shortcut connections [2, 34, 49] have been studied for a long time. An early practice of training multi-layer perceptrons (MLPs) is to add a linear layer connected from the network input to the output [34, 49]. In [44, 24], a few intermediate layers are directly connected to auxiliary classifiers for addressing vanishing/exploding gradients. The papers of [39, 38, 31, 47] propose methods for centering layer responses, gradients, and propagated errors, implemented by shortcut connections. In [44], an inception layer is composed of a shortcut branch and a few deeper branches. Concurrent with our work, highway networks [42, 43] present shortcut connections with gating functions [15]. These gates are data-dependent and have parameters, in contrast to our identity shortcuts that are parameter-free. When a gated shortcut is closed (approaching zero), the layers in highway networks represent non-residual functions. On the contrary, our formulation always learns residual functions; our identity shortcuts are never closed, and all information is always passed through, with additional residual functions to be learned. In addition, high.way networks have not demonstrated accuracy gains with extremely increased depth (e.g., over 100 layers). 3. Deep Residual Learning 3.1. Residual Learning Let us consider <> as an underlying mapping to be fit by a few stacked layers (not necessarily the entire net), with x denoting the inputs to the first of these layers. If one hypothesizes that multiple nonlinear layers can asymptotically approximate complicated functions, then it is equivalent to hypothesize that they can asymptotically approximate the residual functions, i.e., <> (assuming that the input and output are of the same dimensions). So rather than expect stacked layers to approximate <>, we explicitly let these layers approximate a residual function <>. The original function thus becomes <>. Although both forms should be able to asymptotically approximate the desired functions (as hypothesized), the ease of learning might be different. This reformulation is motivated by the counterintuitive phenomena about the degradation problem (Fig. 1, left). As we discussed in the introduction, if the added layers can be constructed as identity mappings, a deeper model should have training error no greater than its shallower counterpart. The degradation problem suggests that the solvers might have difficulties in approximating identity mappings by multiple nonlinear layers. With the residual learning reformulation, if identity mappings are optimal, the solvers may simply drive the weights of the multiple nonlinear lay.ers toward zero to approach identity mappings. In real cases, it is unlikely that identity mappings are optimal, but our reformulation may help to precondition the problem. If the optimal function is closer to an identity mapping than to a zero mapping, it should be easier for the solver to find the perturbations with reference to an identity mapping, than to learn the function as a new one. We show by experiments (Fig. 7) that the learned residual functions in general have small responses, suggesting that identity map.pings provide reasonable preconditioning. 3.2. Identity Mapping by Shortcuts We adopt residual learning to every few stacked layers. A building block is shown in Fig. 2. Formally, in this paper we consider a building block defined as: <>. (1) Here x and y are the input and output vectors of the lay.ers considered. The function <> represents the residual mapping to be learned. For the example in Fig. 2 that has two layers, <> in which <> denotes ReLU [29] and the biases are omitted for simplifying notations. The operation <> is performed by a shortcut connection and element-wise addition. We adopt the second nonlinearity after the addition (i.e., <>, see Fig. 2). The shortcut connections in Eqn.(1) introduce neither extra parameter nor computation complexity. This is not only attractive in practice but also important in our comparisons between plain and residual networks. We can fairly compare plain/residual networks that simultaneously have the same number of parameters, depth, width, and computational cost (except for the negligible element-wise addition). The dimensions of x and F must be equal in Eqn.(1). If this is not the case (e.g., when changing the input/output channels), we can perform a linear projection Ws by the shortcut connections to match the dimensions: <>. (2) We can also use a square matrix <> in Eqn.(1). But we will show by experiments that the identity mapping is sufficient for addressing the degradation problem and is economical, and thus Ws is only used when matching dimensions. The form of the residual function F is flexible. Experiments in this paper involve a function F that has two or three layers (Fig. 5), while more layers are possible. But if F has only a single layer, Eqn.(1) is similar to a linear layer: <>, for which we have not observed advantages. We also note that although the above notations are about fully-connected layers for simplicity, they are applicable to convolutional layers. The function <> can represent multiple convolutional layers. The element-wise addition is performed on two feature maps, channel by channel. 3.3. Network Architectures We have tested various plain/residual nets, and have observed consistent phenomena. To provide instances for discussion, we describe two models for ImageNet as follows. Plain Network. Our plain baselines (Fig. 3, middle) are mainly inspired by the philosophy of VGG nets [41] (Fig. 3, left). The convolutional layers mostly have 3.3 filters and follow two simple design rules: (i) for the same output feature map size, the layers have the same number of filters; and (ii) if the feature map size is halved, the number of filters is doubled so as to preserve the time complexity per layer. We perform downsampling directly by convolutional layers that have a stride of 2. The network ends with a global average pooling layer and a 1000-way fully-connected layer with softmax. The total number of weighted layers is 34 in Fig. 3 (middle). It is worth noticing that our model has fewer filters and lower complexity than VGG nets [41] (Fig. 3, left). Our 34 layer baseline has 3.6 billion FLOPs (multiply-adds), which is only 18% of VGG-19 (19.6 billion FLOPs). Residual Network. Based on the above plain network, we insert shortcut connections (Fig. 3, right) which turn the network into its counterpart residual version. The identity shortcuts (Eqn.(1)) can be directly used when the input and output are of the same dimensions (solid line shortcuts in Fig. 3). When the dimensions increase (dotted line shortcuts in Fig. 3), we consider two options: (A) The shortcut still performs identity mapping, with extra zero entries padded for increasing dimensions. This option introduces no extra parameter; (B) The projection shortcut in Eqn.(2) is used to match dimensions (done by 1.1 convolutions). For both options, when the shortcuts go across feature maps of two sizes, they are performed with a stride of 2. 3.4. Implementation Our implementation for ImageNet follows the practice in [21, 41]. The image is resized with its shorter side randomly sampled in [256, 480] for scale augmentation [41]. A 224.224 crop is randomly sampled from an image or its horizontal flip, with the per-pixel mean subtracted [21]. The standard color augmentation in [21] is used. We adopt batch normalization (BN) [16] right after each convolution and before activation, following [16]. We initialize the weights as in [13] and train all plain/residual nets from scratch. We use SGD with a mini-batch size of 256. The learning rate starts from 0.1 and is divided by 10 when the error plateaus, and the models are trained for up to 60 . 104 iterations. We use a weight decay of 0.0001 and a momentum of 0.9. We do not use dropout [14], following the practice in [16]. In testing, for comparison studies we adopt the standard 10-crop testing [21]. For best results, we adopt the fully convolutional form as in [41, 13], and average the scores at multiple scales (images are resized such that the shorter side is in {224, 256, 384, 480, 640}). 4. Experiments 4.1. ImageNet Classification We evaluate our method on the ImageNet 2012 classification dataset [36] that consists of 1000 classes. The models are trained on the 1.28 million training images, and evaluated on the 50k validation images. We also obtain a final result on the 100k test images, reported by the test server. We evaluate both top-1 and top-5 error rates. Plain Networks. We first evaluate 18-layer and 34-layer plain nets. The 34-layer plain net is in Fig. 3 (middle). The 18-layer plain net is of a similar form. See Table 1 for de. tailed architectures. The results in Table 2 show that the deeper 34-layer plain net has higher validation error than the shallower 18-layer plain net. To reveal the reasons, in Fig. 4 (left) we com.pare their training/validation errors during the training procedure. We have observed the degradation problem <

> Table 1. Architectures for ImageNet. Building blocks are shown in brackets (see also Fig. 5), with the numbers of blocks stacked. Down-sampling is performed by conv3 1, conv4 1, and conv5 1 with a stride of 2. Figure 4. Training on ImageNet. Thin curves denote training error, and bold curves denote validation error of the center crops. Left: plain networks of 18 and 34 layers. Right: ResNets of 18 and 34 layers. In this plot, the residual networks have no extra parameter compared to their plain counterparts. <

> Table 2. Top-1 error (%, 10-crop testing) on ImageNet validation. Here the ResNets have no extra parameter compared to their plain counterparts. Fig. 4 shows the training procedures. 34-layer plain net has higher training error throughout the whole training procedure, even though the solution space of the 18-layer plain network is a subspace of that of the 34-layer one. We argue that this optimization difficulty is unlikely to be caused by vanishing gradients. These plain networks are trained with BN [16], which ensures forward propagated signals to have non-zero variances. We also verify that the backward propagated gradients exhibit healthy norms with BN. So neither forward nor backward signals vanish. In fact, the 34-layer plain net is still able to achieve compet.itive accuracy (Table 3), suggesting that the solver works to some extent. We conjecture that the deep plain nets may have exponentially low convergence rates, which impact the reducing of the training error3. The reason for such opti.mization difficulties will be studied in the future. Residual Networks. Next we evaluate 18-layer and 34.layer residual nets (ResNets). The baseline architectures are the same as the above plain nets, expect that a shortcut connection is added to each pair of 3.3 filters as in Fig. 3 (right). In the first comparison (Table 2 and Fig. 4 right), we use identity mapping for all shortcuts and zero-padding for increasing dimensions (option A). So they have no extra parameter compared to the plain counterparts. We have three major observations from Table 2 and Fig. 4. First, the situation is reversed with residual learn.ing fi the 34-layer ResNet is better than the 18-layer ResNet (by 2.8%). More importantly, the 34-layer ResNet exhibits considerably lower training error and is generalizable to the validation data. This indicates that the degradation problem is well addressed in this setting and we manage to obtain accuracy gains from increased depth. Second, compared to its plain counterpart, the 34-layer 3We have experimented with more training iterations (3.) and still ob.served the degradation problem, suggesting that this problem cannot be feasibly addressed by simply using more iterations. <

> Table 3. Error rates (%, 10-crop testing) on ImageNet validation. VGG-16 is based on our test. ResNet-50/101/152 are of option B that only uses projections for increasing dimensions. <> Table 4. Error rates (%) of single-model results on the ImageNet validation set (except fi reported on the test set). <> Table 5. Error rates (%) of ensembles. The top-5 error is on the test set of ImageNet and reported by the test server. <> ResNet reduces the top-1 error by 3.5% (Table 2), resulting from the successfully reduced training error (Fig. 4 right vs. left). This comparison verifies the effectiveness of residual learning on extremely deep systems. Last, we also note that the 18-layer plain/residual nets are comparably accurate (Table 2), but the 18-layer ResNet converges faster (Fig. 4 right vs. left). When the net is not overly deep (18 layers here), the current SGD solver is still able to find good solutions to the plain net. In this case, the ResNet eases the optimization by providing faster convergence at the early stage. Identity vs. Projection Shortcuts. We have shown that Figure 5. A deeper residual function F for ImageNet. Left: a building block (on 56.56 feature maps) as in Fig. 3 for ResNet. <

> parameter-free, identity shortcuts help with training. Next we investigate projection shortcuts (Eqn.(2)). In Table 3 we compare three options: (A) zero-padding shortcuts are used for increasing dimensions, and all shortcuts are parameter-free (the same as Table 2 and Fig. 4 right); (B) projection shortcuts are used for increasing dimensions, and other shortcuts are identity; and (C) all shortcuts are projections. Table 3 shows that all three options are considerably bet. ter than the plain counterpart. B is slightly better than A. We argue that this is because the zero-padded dimensions in A indeed have no residual learning. C is marginally better than B, and we attribute this to the extra parameters introduced by many (thirteen) projection shortcuts. But the small differences among A/B/C indicate that projection shortcuts are not essential for addressing the degradation problem. So we do not use option C in the rest of this paper, to reduce mem.ory/time complexity and model sizes. Identity shortcuts are particularly important for not increasing the complexity of the bottleneck architectures that are introduced below. Deeper Bottleneck Architectures. Next we describe our deeper nets for ImageNet. Because of concerns on the train.ing time that we can afford, we modify the building block as a bottleneck design4. For each residual function F, we use a stack of 3 layers instead of 2 (Fig. 5). The three layers are 1.1, 3.3, and 1.1 convolutions, where the 1.1 layers are responsible for reducing and then increasing (restoring) dimensions, leaving the 3.3 layer a bottleneck with smaller input/output dimensions. Fig. 5 shows an example, where both designs have similar time complexity. The parameter-free identity shortcuts are particularly important for the bottleneck architectures. If the identity short.cut in Fig. 5 (right) is replaced with projection, one can show that the time complexity and model size are doubled, as the shortcut is connected to the two high-dimensional ends. So identity shortcuts lead to more efficient models for the bottleneck designs. 50-layer ResNet: We replace each 2-layer block in the 4Deeper non-bottleneck ResNets (e.g., Fig. 5 left) also gain accuracy from increased depth (as shown on CIFAR-10), but are not as economical as the bottleneck ResNets. So the usage of bottleneck designs is mainly due to practical considerations. We further note that the degradation problem of plain nets is also witnessed for the bottleneck designs. 34-layer net with this 3-layer bottleneck block, resulting in a 50-layer ResNet (Table 1). We use option B for increasing dimensions. This model has 3.8 billion FLOPs. 101-layer and 152-layer ResNets: We construct 101.layer and 152-layer ResNets by using more 3-layer blocks (Table 1). Remarkably, although the depth is significantly increased, the 152-layer ResNet (11.3 billion FLOPs) still has lower complexity than VGG-16/19 nets (15.3/19.6 bil.lion FLOPs). The 50/101/152-layer ResNets are more accurate than the 34-layer ones by considerable margins (Table 3 and 4). We do not observe the degradation problem and thus enjoy significant accuracy gains from considerably increased depth. The benefits of depth are witnessed for all evaluation metrics (Table 3 and 4). Comparisons with State-of-the-art Methods. In Table 4 we compare with the previous best single-model results. Our baseline 34-layer ResNets have achieved very competitive accuracy. Our 152-layer ResNet has a single-model top-5 validation error of 4.49%. This single-model result outperforms all previous ensemble results (Table 5). We combine six models of different depth to form an ensemble (only with two 152-layer ones at the time of submitting). This leads to 3.57% top-5 error on the test set (Table 5). This entry won the 1st place in ILSVRC 2015. 4.2. CIFAR-10 and Analysis We conducted more studies on the CIFAR-10 dataset [20], which consists of 50k training images and 10k test.ing images in 10 classes. We present experiments trained on the training set and evaluated on the test set. Our focus is on the behaviors of extremely deep networks, but not on pushing the state-of-the-art results, so we intentionally use simple architectures as follows. The plain/residual architectures follow the form in Fig. 3 (middle/right). The network inputs are 32.32 images, with the per-pixel mean subtracted. The first layer is 3.3 convolutions. Then we use a stack of 6n layers with 3.3 convolutions on the feature maps of sizes {32, 16, 8} respectively, with 2n layers for each feature map size. The numbers of filters are {16, 32, 64} respectively. The subsampling is per.formed by convolutions with a stride of 2. The network ends with a global average pooling, a 10-way fully-connected layer, and softmax. There are totally 6n+2 stacked weighted layers. The following table summarizes the architecture: <

> When shortcut connections are used, they are connected to the pairs of 3.3 layers (totally 3n shortcuts). On this dataset we use identity shortcuts in all cases (i.e., option A), <> Table 6. Classification error on the CIFAR-10 test set. All meth.ods are with data augmentation. For ResNet-110, we run it 5 times and show best (mean std) as in [43]. so our residual models have exactly the same depth, width, and number of parameters as the plain counterparts. We use a weight decay of 0.0001 and momentum of 0.9, and adopt the weight initialization in [13] and BN [16] but with no dropout. These models are trained with a mini-batch size of 128 on two GPUs. We start with a learning rate of 0.1, divide it by 10 at 32k and 48k iterations, and terminate training at 64k iterations, which is determined on a 45k/5k train/val split. We follow the simple data augmen.tation in [24] for training: 4 pixels are padded on each side, and a 32.32 crop is randomly sampled from the padded image or its horizontal fiip. For testing, we only evaluate the single view of the original 32.32 image. We compare n = {3, 5, 7, 9}, leading to 20, 32, 44, and 56-layer networks. Fig. 6 (left) shows the behaviors of the plain nets. The deep plain nets suffer from increased depth, and exhibit higher training error when going deeper. This phenomenon is similar to that on ImageNet (Fig. 4, left) and on MNIST (see [42]), suggesting that such an optimization difficulty is a fundamental problem. Fig. 6 (middle) shows the behaviors of ResNets. Also similar to the ImageNet cases (Fig. 4, right), our ResNets manage to overcome the optimization difficulty and demon.strate accuracy gains when the depth increases. We further explore n = 18 that leads to a 110-layer ResNet. In this case, we find that the initial learning rate of 0.1 is slightly too large to start converging5. So we use 0.01 to warm up the training until the training error is below 80% (about 400 iterations), and then go back to 0.1 and continue training. The rest of the learning schedule is as done previously. This 110-layer network converges well (Fig. 6, middle). It has fewer parameters than other deep and thin 5With an initial learning rate of 0.1, it starts converging (<90% error) after several epochs, but still reaches similar accuracy. <

> Figure 7. Standard deviations (std) of layer responses on CIFAR. 10. The responses are the outputs of each 3.3 layer, after BN and before nonlinearity. Top: the layers are shown in their original order. Bottom: the responses are ranked in descending order. networks such as FitNet [35] and Highway [42] (Table 6), yet is among the state-of-the-art results (6.43%, Table 6). Analysis of Layer Responses. Fig. 7 shows the standard deviations (std) of the layer responses. The responses are the outputs of each 3.3 layer, after BN and before other nonlinearity (ReLU/addition). For ResNets, this analysis reveals the response strength of the residual functions. Fig. 7 shows that ResNets have generally smaller responses than their plain counterparts. These results support our ba.sic motivation (Sec.3.1) that the residual functions might be generally closer to zero than the non-residual functions. We also notice that the deeper ResNet has smaller magnitudes of responses, as evidenced by the comparisons among ResNet-20, 56, and 110 in Fig. 7. When there are more layers, an individual layer of ResNets tends to modify the signal less. Exploring Over 1000 layers. We explore an aggressively deep model of over 1000 layers. We set n = 200 that leads to a 1202-layer network, which is trained as described above. Our method shows no optimization difficulty, and this 103-layer network is able to achieve training error <0.1% (Fig. 6, right). Its test error is still fairly good (7.93%, Table 6). But there are still open problems on such aggressively deep models. The testing result of this 1202-layer network is worse than that of our 110-layer network, although both <

> Table 7. Object detection mAP (%) on the PASCAL VOC 2007/2012 test sets using baseline Faster R-CNN. See also Ta.ble 10 and 11 for better results. <> Table 8. Object detection mAP (%) on the COCO validation set using baseline Faster R-CNN. See also Table 9 for better results. have similar training error. We argue that this is because of overfitting. The 1202-layer network may be unnecessarily large (19.4M) for this small dataset. Strong regularization such as maxout [10] or dropout [14] is applied to obtain the best results ([10, 25, 24, 35]) on this dataset. In this paper, we use no maxout/dropout and just simply impose regularization via deep and thin architectures by design, without distracting from the focus on the difficulties of optimization. But combining with stronger regularization may im.prove results, which we will study in the future. 4.3. Object Detection on PASCAL and MS COCO Our method has good generalization performance on other recognition tasks. Table 7 and 8 show the object detection baseline results on PASCAL VOC 2007 and 2012 [5] and COCO [26]. We adopt Faster R-CNN [32] as the detection method. Here we are interested in the improvements of replacing VGG-16 [41] with ResNet-101. The detection implementation (see appendix) of using both models is the same, so the gains can only be attributed to better networks. Most remarkably, on the challenging COCO dataset we ob.tain a 6.0% increase in COCOfis standard metric (mAP@[.5, .95]), which is a 28% relative improvement. This gain is solely due to the learned representations. Based on deep residual nets, we won the 1st places in several tracks in ILSVRC & COCO 2015 competitions: Im.ageNet detection, ImageNet localization, COCO detection, and COCO segmentation. The details are in the appendix. References [1] Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is difficult. IEEE Transactions on Neural Networks, 5(2):157fi166, 1994. [2] C. M. Bishop. Neural networks for pattern recognition. Oxford university press, 1995. [3] W. L. Briggs, S. F. McCormick, et al. A Multigrid Tutorial. Siam, 2000. [4] K. Chatfield, V. Lempitsky, A. Vedaldi, and A. Zisserman. The devil is in the details: an evaluation of recent feature encoding methods. In BMVC, 2011. [5] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman. The Pascal Visual Object Classes (VOC) Challenge. IJCV, pages 303fi338, 2010. [6] S. Gidaris and N. Komodakis. Object detection via a multi-region & semantic segmentation-aware cnn model. In ICCV, 2015. [7] R. Girshick. Fast R-CNN. In ICCV, 2015. [8] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hier.archies for accurate object detection and semantic segmentation. In CVPR, 2014. [9] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In AISTATS, 2010. [10] I. J. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and Y. Bengio. Maxout networks. arXiv:1302.4389, 2013. [11] K. He and J. Sun. Convolutional neural networks at constrained time cost. In CVPR, 2015. [12] K. He, X. Zhang, S. Ren, and J. Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. In ECCV, 2014. [13] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. [14] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv:1207.0580, 2012. [15] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735fi1780, 1997. [16] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. [17] H. Jegou, M. Douze, and C. Schmid. Product quantization for nearest neighbor search. TPAMI, 33, 2011. [18] H. Jegou, F. Perronnin, M. Douze, J. Sanchez, P. Perez, and C. Schmid. Aggregating local image descriptors into compact codes. TPAMI, 2012. [19] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv:1408.5093, 2014. [20] A. Krizhevsky. Learning multiple layers of features from tiny im.ages. Tech Report, 2009. [21] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012. [22] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to hand.written zip code recognition. Neural computation, 1989. [23] Y. LeCun,L.Bottou,G.B.Orr,andK.-R.Mfiuller. Efficientbackprop. In Neural Networks: Tricks of the Trade, pages 9fi50. Springer, 1998. [24] C.-Y. Lee, S. Xie, P. Gallagher, Z. Zhang, and Z. Tu. Deeply-supervised nets. arXiv:1409.5185, 2014. [25] M. Lin, Q. Chen, and S. Yan. Network in network. arXiv:1312.4400, 2013. [26] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollfiar, and C. L. Zitnick. Microsoft COCO: Common objects in context. In ECCV. 2014. [27] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. [28] G. Montfiufar, R. Pascanu, K. Cho, and Y. Bengio. On the number of linear regions of deep neural networks. In NIPS, 2014. [29] V. Nair and G. E. Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010. [30] F. Perronnin and C. Dance. Fisher kernels on visual vocabularies for image categorization. In CVPR, 2007. [31] T. Raiko, H. Valpola, and Y. LeCun. Deep learning made easier by linear transformations in perceptrons. In AISTATS, 2012. [32] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015. [33] S. Ren, K. He, R. Girshick, X. Zhang, and J. Sun. Object detection networks on convolutional feature maps. arXiv:1504.06066, 2015. [34] B. D. Ripley. Pattern recognition and neural networks. Cambridge university press, 1996. [35] A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y. Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. [36] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. arXiv:1409.0575, 2014. [37] A. M. Saxe, J. L. McClelland, and S. Ganguli. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. arXiv:1312.6120, 2013. [38] N. N. Schraudolph. Accelerated gradient descent by factor-centering decomposition. Technical report, 1998. [39] N. N. Schraudolph. Centering neural network gradient factors. In Neural Networks: Tricks of the Trade, pages 207fi226. Springer, 1998. [40] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. Le-Cun. Overfeat: Integrated recognition, localization and detection using convolutional networks. In ICLR, 2014. [41] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. [42] R. K. Srivastava, K. Greff, and J. Schmidhuber. Highway networks. arXiv:1505.00387, 2015. [43] R. K. Srivastava, K. Greff, and J. Schmidhuber. Training very deep networks. 1507.06228, 2015. [44] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Er.han, V. Vanhoucke, and A. Rabinovich. Going deeper with convolu.tions. In CVPR, 2015. [45] R. Szeliski. Fast surface interpolation using hierarchical basis func.tions. TPAMI, 1990. [46] R. Szeliski. Locally adapted hierarchical basis preconditioning. In SIGGRAPH, 2006. [47] T. Vatanen, T. Raiko, H. Valpola, and Y. LeCun. Pushing stochas.tic gradient towards second-order methodsfibackpropagation learn.ing with transformations in nonlinearities. In Neural Information Processing, 2013. [48] A. Vedaldi and B. Fulkerson. VLFeat: An open and portable library of computer vision algorithms, 2008. [49] W. Venables and B. Ripley. Modern applied statistics with s-plus. 1999. [50] M. D. Zeiler and R. Fergus. Visualizing and understanding convolu.tional neural networks. In ECCV, 2014. A. Object Detection Baselines In this section we introduce our detection method based on the baseline Faster R-CNN [32] system. The models are initialized by the ImageNet classification models, and then fine-tuned on the object detection data. We have experi.mented with ResNet-50/101 at the time of the ILSVRC & COCO 2015 detection competitions. Unlike VGG-16 used in [32], our ResNet has no hidden fc layers. We adopt the idea of fiNetworks on Conv feature maps (NoC) [33] to address this issue. We compute the full-image shared conv feature maps using those lay.ers whose strides on the image are no greater than 16 pixels (i.e., conv1, conv2 x, conv3 x, and conv4 x, totally 91 conv layers in ResNet-101; Table 1). We consider these layers as analogous to the 13 conv layers in VGG-16, and by doing so, both ResNet and VGG-16 have conv feature maps of the same total stride (16 pixels). These layers are shared by a region proposal network (RPN, generating 300 proposals) [32] and a Fast R-CNN detection network [7]. RoI pool.ing [7] is performed before conv5 1. On this RoI-pooled feature, all layers of conv5 x and up are adopted for each region, playing the roles of VGG-16fis fc layers. The final classification layer is replaced by two sibling layers (classi.fication and box regression [7]). For the usage of BN layers, after pre-training, we compute the BN statistics (means and variances) for each layer on the ImageNet training set. Then the BN layers are fixed during fine-tuning for object detection. As such, the BN layers become linear activations with constant offsets and scales, and BN statistics are not updated by fine-tuning. We fix the BN layers mainly for reducing memory consumption in Faster R-CNN training. PASCAL VOC Following [7, 32], for the PASCAL VOC 2007 test set, we use the 5k trainval images in VOC 2007 and 16k train-val images in VOC 2012 for training (fi07+12fi). For the PASCAL VOC 2012 test set, we use the 10k trainval+test images in VOC 2007 and 16k trainval images in VOC 2012 for training (fi07++12fi). The hyper-parameters for train.ing Faster R-CNN are the same as in [32]. Table 7 shows the results. ResNet-101 improves the mAP by >3% over VGG-16. This gain is solely because of the improved features learned by ResNet. MS COCO The MS COCO dataset [26] involves 80 object cate.gories. We evaluate the PASCAL VOC metric (mAP @ IoU = 0.5) and the standard COCO metric (mAP @ IoU = .5:.05:.95). We use the 80k images on the train set for train.ing and the 40k images on the val set for evaluation. Our detection system for COCO is similar to that for PASCAL VOC. We train the COCO models with an 8-GPU imple.mentation, and thus the RPN step has a mini-batch size of 8 images (i.e., 1 per GPU) and the Fast R-CNN step has a mini-batch size of 16 images. The RPN step and Fast R.CNN step are both trained for 240k iterations with a learn.ing rate of 0.001 and then for 80k iterations with 0.0001. Table 8 shows the results on the MS COCO validation set. ResNet-101 has a 6% increase of mAP@[.5, .95] over VGG-16, which is a 28% relative improvement, solely con.tributed by the features learned by the better network. Re.markably, the mAP@[.5, .95]fis absolute increase (6.0%) is nearly as big as mAP@.5fis (6.9%). This suggests that a deeper network can improve both recognition and localiza.tion. B. Object Detection Improvements For completeness, we report the improvements made for the competitions. These improvements are based on deep features and thus should benefit from residual learning. MS COCO Box refinement. Our box refinement partially follows the it.erative localization in [6]. In Faster R-CNN, the final output is a regressed box that is different from its proposal box. So for inference, we pool a new feature from the regressed box and obtain a new classification score and a new regressed box. We combine these 300 new predictions with the orig.inal 300 predictions. Non-maximum suppression (NMS) is applied on the union set of predicted boxes using an IoU threshold of 0.3 [8], followed by box voting [6]. Box re.finement improves mAP by about 2 points (Table 9). Global context. We combine global context in the Fast R-CNN step. Given the full-image conv feature map, we pool a feature by global Spatial Pyramid Pooling [12] (with a fisingle-levelfi pyramid) which can be implemented as fiRoIfi pooling using the entire imagefis bounding box as the RoI. This pooled feature is fed into the post-RoI layers to obtain a global context feature. This global feature is con.catenated with the original per-region feature, followed by the sibling classification and box regression layers. This new structure is trained end-to-end. Global context im.proves mAP@.5 by about 1 point (Table 9). Multi-scale testing. In the above, all results are obtained by single-scale training/testing as in [32], where the imagefis shorter side is s = 600 pixels. Multi-scale training/testing has been developed in [12, 7] by selecting a scale from a feature pyramid, and in [33] by using maxout layers. In our current implementation, we have performed multi-scale testing following [33]; we have not performed multi-scale training because of limited time. In addition, we have per.formed multi-scale testing only for the Fast R-CNN step (but not yet for the RPN step). With a trained model, we compute conv feature maps on an image pyramid, where the imagefis shorter sides are s .{200, 400, 600, 800, 1000}. <> Table 9. Object detection improvements on MS COCO using Faster R-CNN and ResNet-101. <> Table 10. Detection results on the PASCAL VOC 2007 test set. The baseline is the Faster R-CNN system. The system fibaseline+++fi include box refinement, context, and multi-scale testing in Table 9. system net data mAP areo bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv baseline baseline baseline+++ VGG-16 ResNet-101 ResNet-101 07++12 07++12 COCO+07++12 70.4 73.8 83.8 84.9 79.8 74.3 53.9 49.8 77.5 75.9 88.5 45.6 77.1 55.3 86.9 81.7 80.9 79.6 40.1 72.6 60.9 81.2 61.5 86.5 81.6 77.2 58.0 51.0 78.6 76.6 93.2 48.6 80.4 59.0 92.1 85.3 84.8 80.7 48.1 77.3 66.5 84.7 65.6 92.1 88.4 84.8 75.9 71.4 86.3 87.8 94.2 66.8 89.4 69.2 93.9 91.9 90.9 89.6 67.9 88.2 76.8 90.3 80.0 Table 11. Detection results on the PASCAL VOC 2012 test set (http://host.robots.ox.ac.uk:8080/leaderboard/displaylb.php?challengeid=11&compid=4). The baseline is the Faster R-CNN system. The system baseline+++ include box refinement, context, and multi-scale testing in Table 9. We select two adjacent scales from the pyramid following [33]. RoI pooling and subsequent layers are performed on the feature maps of these two scales [33], which are merged by maxout as in [33]. Multi-scale testing improves the mAP by over 2 points (Table 9). Using validation data. Next we use the 80k+40k trainval set for training and the 20k test-dev set for evaluation. The test.dev set has no publicly available ground truth and the result is reported by the evaluation server. Under this setting, the results are an mAP@.5 of 55.7% and an mAP@[.5, .95] of 34.9% (Table 9). This is our single-model result. Ensemble. In Faster R-CNN, the system is designed to learn region proposals and also object classifiers, so an ensemble can be used to boost both tasks. We use an ensemble for proposing regions, and the union set of proposals are pro.cessed by an ensemble of per-region classifiers. Table 9 shows our result based on an ensemble of 3 networks. The mAP is 59.0% and 37.4% on the test-dev set. This result won the 1st place in the detection task in COCO 2015. We revisit the PASCAL VOC dataset based on the above model. With the single model on the COCO dataset (55.7% mAP@.5 in Table 9), we fine-tune this model on the PAS.CAL VOC sets. The improvements of box refinement, con.text, and multi-scale testing are also adopted. By doing so we achieve 85.6% mAP on PASCAL VOC 2007 (Table 10) and 83.8% on PASCAL VOC 2012 (Table 11)6. The result on PASCAL VOC 2012 is 10 points higher than the previ.ous state-of-the-art result [6]. <> Table 12. Our results (mAP, %) on the ImageNet detection dataset. Our detection system is Faster R-CNN [32] with the improvements in Table 9, using ResNet-101. ImageNet Detection The ImageNet Detection (DET) task involves 200 object categories. The accuracy is evaluated by mAP@.5. Our object detection algorithm for ImageNet DET is the same as that for MS COCO in Table 9. The networks are pre.trained on the 1000-class ImageNet classification set, and are fine-tuned on the DET data. We split the validation set into two parts (val1/val2) following [8]. We fine-tune the detection models using the DET training set and the val1 set. The val2 set is used for validation. We do not use other ILSVRC 2015 data. Our single model with ResNet-101 has <> Table 13. Localization error (%) on the ImageNet validation. In the column of fiLOC error on GT classfi ([41]), the ground truth class is used. In the fitestingfi column, fi1-cropfi denotes testing on a center crop of 224.224 pixels, fidensefi denotes dense (fully convolutional) and multi-scale testing. <> Table 14. Comparisons of localization error (%) on the ImageNet dataset with state-of-the-art methods. 58.8% mAP and our ensemble of 3 models has 62.1% mAP on the DET test set (Table 12). This result won the 1st place in the ImageNet detection task in ILSVRC 2015, surpassing the second place by 8.5 points (absolute). C. ImageNet Localization The ImageNet Localization (LOC) task [36] requires to classify and localize the objects. Following [40, 41], we assume that the image-level classifiers are first adopted for predicting the class labels of an image, and the localiza.tion algorithm only accounts for predicting bounding boxes based on the predicted classes. We adopt the fiper-class re.gressionfi (PCR) strategy [40, 41], learning a bounding box regressor for each class. We pre-train the networks for Im.ageNet classification and then fine-tune them for localiza.tion. We train networks on the provided 1000-class Ima.geNet training set. Our localization algorithm is based on the RPN frame.work of [32] with a few modifications. Unlike the way in [32] that is category-agnostic, our RPN for localization is designed in a per-class form. This RPN ends with two sib.ling 1.1 convolutional layers for binary classification (cls) and box regression (reg), as in [32]. The cls and reg layers are both in a per-class from, in contrast to [32]. Specifi.cally, the cls layer has a 1000-d output, and each dimension is binary logistic regression for predicting being or not be.ing an object class; the reg layer has a 1000.4-d output consisting of box regressors for 1000 classes. As in [32], our bounding box regression is with reference to multiple translation-invariant fianchorfi boxes at each position. As in our ImageNet classification training (Sec. 3.4), we randomly sample 224.224 crops for data augmentation. We use a mini-batch size of 256 images for fine-tuning. To avoid negative samples being dominate, 8 anchors are ran.domly sampled for each image, where the sampled positive and negative anchors have a ratio of 1:1 [32]. For testing, the network is applied on the image fully-convolutionally. Table 13 compares the localization results. Following [41], we first perform fioraclefi testing using the ground truth class as the classification prediction. VGGfis paper [41] re-ports a center-crop error of 33.1% (Table 13) using ground truth classes. Under the same setting, our RPN method us.ing ResNet-101 net significantly reduces the center-crop er.ror to 13.3%. This comparison demonstrates the excellent performance of our framework. With dense (fully convolu.tional) and multi-scale testing, our ResNet-101 has an error of 11.7% using ground truth classes. Using ResNet-101 for predicting classes (4.6% top-5 classification error, Table 4), the top-5 localization error is 14.4%. The above results are only based on the proposal network (RPN) in Faster R-CNN [32]. One may use the detection network (Fast R-CNN [7]) in Faster R-CNN to improve the results. But we notice that on this dataset, one image usually contains a single dominate object, and the proposal regions highly overlap with each other and thus have very similar RoI-pooled features. As a result, the image-centric training of Fast R-CNN [7] generates samples of small variations, which may not be desired for stochastic training. Motivated by this, in our current experiment we use the original R-CNN [8] that is RoI-centric, in place of Fast R-CNN. Our R-CNN implementation is as follows. We apply the per-class RPN trained as above on the training images to predict bounding boxes for the ground truth class. These predicted boxes play a role of class-dependent proposals. For each training image, the highest scored 200 proposals are extracted as training samples to train an R-CNN classi.fier. The image region is cropped from a proposal, warped to 224.224 pixels, and fed into the classification network as in R-CNN [8]. The outputs of this network consist of two sibling fc layers for cls and reg, also in a per-class form. This R-CNN network is fine-tuned on the training set us.ing a mini-batch size of 256 in the RoI-centric fashion. For testing, the RPN generates the highest scored 200 proposals for each predicted class, and the R-CNN network is used to update these proposalsfi scores and box positions. This method reduces the top-5 localization error to 10.6% (Table 13). This is our single-model result on the validation set. Using an ensemble of networks for both clas.sification and localization, we achieve a top-5 localization error of 9.0% on the test set. This number significantly out.performs the ILSVRC 14 results (Table 14), showing a 64% relative reduction of error. This result won the 1st place in the ImageNet localization task in ILSVRC 2015. <> <> <> <> <> <> Direct Feedback Alignment Scales to Modern Deep Learning Tasks and Architectures Julien Launay 1;2 Iacopo Poli 1 François Boniface 1 Florent Krzakala 1;2 1 LightOn 2 École Normale Supérieure Abstract Despite being the workhorse of deep learning, the backpropagation algorithm is no panacea. It enforces sequential layer updates, thus preventing efficient paral- lelization of the training process. Furthermore, its biological plausibility is being challenged. Alternative schemes have been devised; yet, under the constraint of synaptic asymmetry, none have scaled to modern deep learning tasks and architec- tures. Here, we challenge this perspective, and study the applicability of Direct Feedback Alignment to neural view synthesis, recommender systems, geometric learning, and natural language processing. In contrast with previous studies lim- ited to computer vision tasks, our findings show that it successfully trains a large range of state-of-the-art deep learning architectures, with performance close to fine-tuned backpropagation. At variance with common beliefs, our work supports that challenging tasks can be tackled in the absence of weight transport. 1 Introduction While the backpropagation algorithm (BP) [1,2] is at the heart of modern deep learning achievements, it is not without pitfalls. For one, its weight updates are non-local and rely on upstream layers. Thus, they cannot be easily parallelized [3], incurring important memory and compute costs. Moreover, its biological implementation is problematic [4,5]. For instance, BP relies on the transpose of the weights to evaluate updates. Hence, synaptic symmetry is required between the forward and backward path: this is implausible in biological brains, and known as the weight transport problem [6]. Consequently, alternative training algorithms have been developed. Some of these algorithms are explicitly biologically inspired [7–13], while others focus on making better use of available compute resources [3,14–19]. Despite these enticing characteristics, none has been widely adopted, as they are often demonstrated on a limited set of tasks. Moreover, as assessed in [20], their performance on challenging datasets under the constraint of synaptic asymmetry is disappointing. We seek to broaden this perspective, and demonstrate the applicability of Direct Feedback Alignment (DFA) [19] in state-of-the-art settings: from applications of fully connected networks such as neural view synthesis and recommender systems, to geometric learning with graph convolutions, and natural language processing with Transformers. Our results define new standards for learning without weight transport and show that challenging tasks can indeed be tackled under synaptic asymmetry. All code needed to reproduce our experiments is available at https://github.com/lightonai/dfa-scales-to-modern-deep-learning. 1.1 Related work Training a neural network is a credit assignment problem: an update is derived for each parameter from its contribution to a cost function. To solve this problem, a spectrum of algorithms exists [21]. Biologically motivated methods Finding a training method applicable under the constraints of biological brains remains an open problem. End-to-end propagation of gradients is unlikely to occur [22], implying local learning is required. Furthermore, the weight transport problem enforces synaptic asymmetry [6]. Inspired by auto-encoders, target propagation methods (TP) [10–12] train distinct feedback connections to invert the feedforward ones. Feedback alignment (FA) [13] replaces the transpose of the forward weights used in the backward pass by a random matrix. Throughout training, the forward weights learn to align with the arbitrary backward weights, eventually approximating BP. Beyond biological considerations As deep learning models grow bigger, large-scale distributed training is increasingly desirable. Greedy layer-wise training [14] allows networks to be built layer by layer, limiting the depth of backpropagation. To enable parallelization of the backward pass, updates must only depend on local quantities. Unsupervised learning is naturally suited for this, as it relies on local losses such as Deep InfoMax [17] and Greedy InfoMax [18]. More broadly, synthetic gradient methods, like decoupled neural interfaces [3,15] and local error signals (LES) [16], approximate gradients using layer-wise trainable feedback networks. DFA [19] expands on FA and directly projects a global error to each layer. A shared feedback path is still needed, but it only depends on a simple random projection operation. Performance of alternative methods Local training methods are successful in unsupervised learn- ing [18]. Even in a supervised setting, they scale to challenging datasets like CIFAR-100 or ImageNet [14,16]. Thus, locality is not too penalizing. However, TP, FA, and DFA are unable to scale to these tasks [20]. In fact, DFA is unable to train convolutional layers [23]. To enable feedback alignment techniques to perform well on challenging datasets, some form of weight transport is necessary: either by explicitly sharing sign information [24–26], or by introducing dedicated phases of alignment for the forward and backward weights where some information is shared [27]. To the best of our knowledge, no method compatible with the weight transport problem has ever been demonstrated on challenging tasks. 1.2 Motivations and contributions We focus on DFA, a compromise between biological and computational considerations. Notably, DFA is compatible with synaptic asymmetry: this asymmetry raises important challenges, seemingly preventing learning in demanding settings. Moreover, it allows for asynchronous weight updates, and puts a single operation at the center of the training stage. This enables new classes of training co-processors [28, 29], leveraging dedicated hardware to perform the random projection. Extensive survey We apply DFA in a large variety of settings matching current trends in machine learning. Previous works have found that DFA is unsuitable for computer vision tasks [20,23]; but computer vision alone cannot be the litmus test of a training method. Instead, we consider four vastly different domains, across eight tasks, and with eleven different architectures. This constitutes a survey of unprecedented scale for an alternative training method, and makes a strong case for the possibility of learning without weight transport in demanding scenarios. Challenging settings We demonstrate the ability of DFA to tackle challenging tasks. We success- fully learn and render real-world 3D scenes (section 3.1.1); we perform recommendation at scale (section 3.1.2); we explore graph-based citation networks (section 3.2); and we consider language modelling with a Transformer (section 3.3). We study tasks at the state-of-the-art level, that have only been recently successfully tackled with deep learning. Modern architectures We prove that the previously established failure of DFA to train convolutions does not generalize. By evaluating performance metrics, comparing against a shallow baseline, measuring alignment, and visualizing t-SNE embeddings, we show that learning indeed occurs in layers involving graph convolutions and attention. This significantly broadens the applicability of DFA–previously thought to be limited to simple problems like MNIST and CIFAR-10. 2 Methods Forward pass In a fully connected network, at layer i out of N, neglecting its biases, with W_i its weight matrix, f_i its non-linearity, and hi its activations, the forward pass is: <> (1) <> is the input data, and <> are the predictions. A task-specific cost function <> is computed to quantify the quality of the predictions with respect to the targets y. Backward pass with BP The weight updates are computed by backpropagation of the error vector. Using the chain-rule of derivatives, each neuron is updated based on its contribution to the cost function. Leaving aside the specifics of the optimizer used, the equation for the weight updates is: <> (2) Backward pass with DFA The gradient signal <> of the (i+1)-th layer violates synaptic asymmetry. DFA replaces it with a random projection of the topmost derivative of the loss, <>. For common classification and regression losses such as the mean squared error or the negative log likelihood, this corresponds to a random projection of the global error <>. With B_i, a fixed random matrix of appropriate shape drawn at initialization for each layers: <> (3) 3 Experiments We study the applicability of DFA to a diverse set of applications requiring state-of-the-art architec- tures. We start with fully connected networks, where DFA has already been demonstrated, and address new challenging settings. We then investigate geometric learning: we apply DFA to graph neural net- works in classification tasks on citation networks, as well as graph autoencoders. These architectures feature graph convolutions and attention layers. Finally, we use DFA to train a transformer-based Natural Language Processing (NLP) model on a dataset of more than 100 million tokens. 3.1 Fully connected architectures DFA has been successful at training fully connected architectures, with performance on-par with backpropagation [19,20]. However, only computer vision tasks have been considered, where fully connected networks considerably underperform their convolutional counterpart. Here, we focus on tasks where fully connected architectures are state-of-the-art. Moreover, the architectures considered are deeper and more complex than those necessary to solve a simple task like MNIST. 3.1.1 Neural view synthesis with Neural Radiance Fields The most recent state-of-the-art neural view synthesis methods are based on large fully connected networks: this is an ideal setting for a first evaluation of DFA on a challenging task. Background There has been growing interest in methods capable of synthesizing novel renders of a 3D scene using a dataset of past renders. The network is trained to learn an inner representation of the scene, and a classical rendering system can then query the model to generate novel views. With robust enough methods, real-world scenes can also be learned from a set of pictures. Until recently, most successful neural view synthesis methods were based on sampled volumetric representations [30–32]. In this context, Convolutional Neural Networks (CNNs) can be used to smooth out the discrete sampling of 3D space [33,34]. However, these methods scale poorly to higher resolutions, as they still require finer and finer sampling. Conversely, alternative schemes based on a continuous volume representation have succeeded in generating high-quality renders [35], even featuring complex phenomenons such as view-dependant scattering [36]. These schemes make point-wise predictions, and use fully connected neural networks to encode the scene. <

> Figure 1: Comparisons of NeRF-DFA with state-of-the-art methods trained with BP on the most challenging synthetic and real-world scenes. While NeRF-DFA generates render of lower quality, they maintain multi-view consistency and exhibit no geometric artifacts. BP results from [36]. Setting We employ Neural Radiance Fields (NeRF) [36], the state-of-the-art for neural view synthesis. NeRF represents scenes as a continuous 5D function of space–three spatial coordinates, two viewing angles–and outputs a point-wise RGB radiance and opacity. A ray-casting renderer can then query the network to generate arbitrary views of the scene. The network modeling the continuous function is 10 layers deep. Two identical networks are trained: the coarse network predictions inform the renderer about the spatial coordinates that the fine network should preferentially evaluate to avoid empty space and occluded regions. Results We report quantitative results of training NeRF with DFA in Table 1. Neural view synthesis methods are often better evaluated qualitatively: we showcase some renders in Figure 1. On a dataset of renders featuring complex scenes with non-Lambertian materials (NeRF-Synthetic [36]), NeRF-DFA outperforms two previous fine-tuned state-of-the-art methods–Scene Representation Networks (SRN) [35] and Local Light Field Fusion (LLFF) [32]–and nearly matches the performance of Neural Volumes (NV) [34]. While DFA underperforms alternative methods trained with BP on the real world view dataset (LLFF-Real [32]), its performance remains significant: real world view synthesis is a challenging tasks, and this level of PSNR indicates that learning is indeed happening. In particular, we find that NeRF-DFA retains the key characteristics of NeRF-BP: it can render view- dependant effects, and is multi-view consistent. The last point is an especially important achievement, and most visible in videos, as it is a challenge for most algorithms [30–32,35]. The main drawback of NeRF-DFA appears to be a seemingly lower render definition. The NeRF architecture has not Table 1: Peak Signal to Noise Ratio (PSNR, higher is better) of neural view synthesis methods trained with backpropagation against NeRF trained with DFA. Even when trained with DFA, NeRF outperforms two state-of-the-art methods on a synthetic dataset (NeRF-Synthetic), and achieves fair performance on a challenging real world views datasets (LLFF-Real). BP results from [36]. <

> been fine-tuned to achieve these results: DFA works out-of-the-box on this advanced method. Future research focusing on architectural changes to NeRF could improve performance with DFA; some preliminary results are included in the supplementary material. 3.1.2 Click-through rate prediction with recommender systems We have demonstrated that DFA can train large fully connected networks on the difficult task of neural view synthesis. We now seek to use DFA in more complex heterogeneous architectures, combining the use of fully connected networks with other machine learning methods.Recommender systems are an ideal application for such considerations. Background Recommender systems are used to model the behavior of users and predict future interactions. In particular, in the context of click-through rate (CTR) prediction, these systems model the probability of a user clicking on a given item. Building recommender systems is hard [37]: their input is high-dimensional and sparse, and the model must learn to extract high-order combinatorial features from the data. Moreover, they need to do so efficiently, as they are used to make millions of predictions and the training data may contain billions of examples. Factorization Machines (FM) [38] use inner-products of latent vectors between features to extract pairwise feature interactions. They constitute an excellent baseline for shallow recommender systems, but fail to efficiently transcribe higher-level features. To avoid extensive feature engineering, it has been suggested that deep learning can be used in conjunction with wide shallow models to extract these higher-level features [39]. In production, these systems are regularly retrained on massive datasets: the speedup allowed by backward unlocking in DFA is thus of particular interest. Setting Deep Factorization Machines (DeepFM) [40] combine FM and a deep fully connected neural network, which we train with DFA. The input embedding is still trained directly via gradient descent, as weight transport is not necessary to backpropagate through the FM. Deep & Cross Networks (DCN) [41] replace the FM with a Cross Network, a deep architecture without non- linearities capable of extracting high-degree interactions across features. We train the fully connected network, the deep cross network, and the embeddings with DFA. Finally, Adaptative Factorization Network (AFN) [42] uses Logarithmic Neural Networks [43] to enhance the representational power of its deep component. We evaluate these methods on the Criteo dataset [44], which features nearly 46 million samples of one million sparse features. This is a difficult task, where performance improvements of the AUC on the 0.001-level can enhance CTR significantly [39]. Results Performance metrics are reported in Table 2. To obtain these results, a simple hyperpa- rameter grid search over optimization and regularization parameters was performed for BP and DFA independently. DFA successfully trains all methods above the FM baseline, and in fact matches BP performance in both DeepFM and AFN. Because of their complexity, recommender systems require intensive tuning and feature engineering to perform at the state-of-the-art level–and reproducing existing results can be challenging [45]. Hence, it is not surprising that a performance gap exists with Deep&Cross–further fine-tuning may be necessary for DFA to reach BP performance. Alignment measurements corroborate that learning is indeed occurring in the special layers of Deep&Cross and AFN–see supplementary for details. Our results on recommender systems support that DFA can learn in a large variety of settings, and that weight transport is not necessary to solve a difficult recommendation task. Table 2: AUC (higher is better) and log loss (lower is better) of recommender systems trained on the Criteo dataset [44]. Even in complex heterogeneous architectures, DFA performance is in line with BP. Values in bold indicate DFA AUC within 0.001 from the BP AUC or better. <> 3.2 Geometric Learning with Graph Convolutional Networks The use of sophisticated architectures beyond fully connected layers is necessary for certain tasks, such as geometric learning[46], where information lies in a complex structured domain. To address geometric learning tasks, methods capable of handling graph-based data are commonly needed. Graph convolutional neural networks (GCNNs) [47–50] have demonstrated the ability to process large-scale graph data efficiently. We study the applicability of DFA to these methods, including recent architectures based on an attention mechanism. Overall, this is an especially interesting setting, as DFA fails to train more classic 2D image convolutional layers [23]. Background Complex data like social networks or brain connections lie on irregular or non- Euclidean domains. They can be represented as graphs, and efficient processing in the spectral domain is possible. Non-spectral techniques to apply neural networks to graphs have also been developed [51–53], but they exhibit unfavorable scaling properties. The success of CNNs in deep learning can be attributed to their ability to efficiently process structured high-dimensional data by sharing local filters. Thus, a generalization of the convolution operator to the graph domain is desirable: [47] first proposed a spectral convolution operation for graphs, and [48] introduced a form of regularization to enforce spatial locality of the filters. We use DFA to train different such GCNNs implementations. We study both spectral and non-spectral convolutions, as well as methods inspired by the attention mechanism. We consider the task of semi-supervised node classification: nodes from a graph are classified using their relationship to other nodes as well as node-wise features. Setting Fast Localized Convolutions (ChebConv) [49] approximate the graph convolution kernel with Chebyshev polynomials, and are one of the first scalable convolution methods on graph. Graph Convolutions (GraphConv) [50] remove the need for an explicit parametrization of the kernel by enforcing linearity of the convolution operation on the graph Laplacian spectrum. It is often considered as the canonical graph convolution. More recent methods do not operate in the spectral domain. Spline Convolutions (SplineConv) [54] use a spline-based kernel, enabling the inclusion of information about the relative positioning of nodes, enhancing their representational power–for instance in the context of 3D meshes. Graph Attention Networks (GATConv) [55] use self-attention [56] layers to enable predictions at a given node to attend more specifically to certain parts of its neighborhood. Finally, building upon Jumping Knowledge Network [57], Just Jump (DNAConv) [58] uses multi- head attention [59] to enhance the aggregation process in graph convolutions and enable deeper architectures. We use PyTorch Geometric [60] for reference implementation of all of these methods. We evaluate performance on three citation network datasets: Cora, CiteSeer, and PubMed [61]. Results We report classification accuracy in Table 3. BP and DFA regularization and optimiza- tion hyperparameters are fine-tuned separately on the Cora dataset. In general, we find that less regularization and lower learning rates are needed with DFA. DFA successfully trains all graph methods, independent of whether they use the spectral domain or not, and even if they use attention. Furthermore, for GraphConv, SplineConv, and GATConv DFA performance nearly matches BP. As GCNNs struggle with learning meaningful representations when stacking many layers [62], all architectures but DNAConv are quite shallow (two layers). However, DFA performance is still significantly higher than that of a shallow training method–see supplementary for details. The lower performance on DNAConv is not a failure to learn: alignment measurements show that learning is indeed occurring. It may be explained instead by a need for more in-depth fine-tuning, as this is a deep architecture with 5 successive attention layers. Table 3: Classification accuracy (%, higher is better) of graph convolution methods trained with BP and DFA, on citation networks [61]. But for ChebConv and DNAConv, DFA performance nearly matches BP performance. Values in bold when DFA is within 2.5% of BP. <> Table 4: AUC and Average Precision Figure 2: t-SNE visualization of the hidden layer (AP, higher is better) for a Graph- activations of a two-layer GraphConv trained on Conv GAE trained with BP or DFA Cora with DFA. Classes forms clear clusters, indicating that a useful intermediary representation is learned. Colors represent different classes. on citation networks. DFA reproduces BP performance. We further demonstrate that DFA helps graph convolutions learn meaningful representations by applying t-SNE [63,64] to the hidden layer activations in GraphConv (Figure 2). Cluster of classes are well-separated, indicating that a useful intermediary representation is derived by the first layer. Graph autoencoders We consider one last application of graph convolutions, in the context of graph autoencoders (GAE). We train a non-probabilistic GAE [65] based on GraphConv with DFA, and report results in Table 4. DFA performance is always in line with BP. 3.3 Natural Language Processing with Transformers We complete our study by training a Transformer [59] on a language modelling task. Transformers have proved successful in text, image, music generation, machine translation, and many supervised NLP tasks [59,66–69]. Here, we demonstrate that DFA can train them, and we show the influence of tuning the optimizer hyperparameters in narrowing the gap with BP. Background NLP has largely benefited from advances in deep learning. Recurrent Neural Net- works were responsible for early breakthroughs, but their sequential nature prevented efficient parallelization of data processing. Transformers are attention-based models that do not rely on recurrence or convolution. Their ability to scale massively has allowed the training of models with several billion parameters [70,71], obtaining state-of-the-art results on all NLP tasks: Transformers now top the prominent SQuAD 2.0 [72,73] and SuperGLUE [74] benchmarks. In parallel, transfer learning in NLP has leaped forward thanks to language modelling, the unsupervised task of predicting the next word. It can leverage virtually unlimited data from web scraping [75]. This enabled the training of universal language models[76] on extremely large and diversified text corpora. These models are useful across a wide range of domains, and can solve most NLP tasks after fine-tuning. Setting The prominence of both language modelling and Transformers gives us the ideal candidate for our NLP experiments: we train a Transformer to predict the next word on the WikiText-103 dataset [77], a large collection of good and featured Wikipedia articles. We use byte-pair-encoding [78] with 32,000 tokens. Our setup is similar to GPT [66]: we adapt the Transformer, originally an encoder-decoder model designed for machine translation, to language modelling. We keep only the encoder and mask the tokens to predict. Our architecture consists in 6 layers, 8 attention heads, a model dimension of 512, and a hidden size of 2048 in the feed-forward blocks. The text is sliced in chunks of 128 tokens and batches of 64 such chunks, resulting in 8192 tokens per batch. Our baseline is trained with BP using the optimization setup of [59]. We found perplexity after 20 epochs to be an excellent indicator of perplexity at convergence; to maximize the number of experiments we could perform, we report the best validation perplexity after 20 epochs. We study two ways of implementing DFA: applying the feedback after every encoder block (macro) or after every layer in those blocks (micro). The input embedding layer receives gradients from the next feedback point through BP. This leaves some amount of weight transport even in the micro-case. Table 5: Best validation perplexity after 20 epochs of a Transformer trained on WikiText-103 (lower is better). The BP and DFA baselines share all hyper-parameters. In Macro the feedback is applied after every transformer layer, while in Micro the feedback is applied after every sub-layer. The learning rate of Adam without the learning rate scheduler is <>. With the scheduler, the initial learning rate is <> and it is multiplied by 0.2 when performance plateaus, with a patience of 1. * score after 22 epochs to let the learning rate scheduler take effect <> Results Our results are summarized in Table 5. Hyper-parameters fine-tuned for BP did not fare well with DFA, but changes in the optimizer narrowed the gap between BP and DFA considerably. The learning rate schedule used on top of Adam [79] in [59] proved detrimental. Using Adam alone required reducing the learning rate between BP and DFA. Increasing 2 from 0.98 [59] to 0.999 improved performance significantly. Finally, a simple scheduler that reduces the learning rate when the validation perplexity plateaus helped reducing it further. Considering that the perplexity of the shallow baseline is over 400, DFA is clearly able to train Transformers. However, our results are not on par with BP, especially in the micro setting. A substantial amount of work remains to make DFA competitive with BP, even more so in a minimal weight transport scenario. The large performance improvements brought by small changes in the optimizer indicate that intensive fine-tuning, common in publications introducing state-of-the-art results, could close the gap between BP and DFA. 4 Conclusion and outlooks We conducted an extensive study demonstrating the ability of DFA to train modern architectures. We considered a broad selection of domains and tasks, with complex models featuring graph convolutions and attention. Our results on large networks like NeRF and Transformers are encouraging, suggesting that with further tuning, such leading architectures can be effectively trained with DFA. Future work on principled training with DFA–in particular regarding the influence of common practices and whether new procedures are required–will help close the gap with BP. More broadly, we verified for the first time that learning under synaptic asymmetry is possible beyond fully-connected layers, and in tasks significantly more difficult than previously considered. This addresses a notable concern in biologically-plausible architectures. DFA still requires an implausible global feedback pathway; however, local training has already been demonstrated at scale. The next step towards biologically-compatible learning is a local method without weight transport. While the tasks and architectures we have considered are not biologically inspired, they constitute a good benchmark for behavioral realism[20]. Any learning algorithm claiming to approximate the brain should reproduce its ability to solve complex and unseen task. Furthermore, even though the current implementation of mechanisms like attention is devoid of biological considerations, they represent broader concepts applicable to human brains [80]. Understanding how our brain learns is a gradual process, and future research could incorporate further realistic elements, like spiking neurons. Finally, unlocking the backward pass in large architectures like Transformers is promising. More opti- mized implementation of DFA–built at a lower-level of existing ML libraries–could unlock significant speed-up. Leveraging the use of a single random projection as the cornerstone of training, dedicated accelerators may employ more exotic hardware architectures. This will open new possibilities in the asynchronous training of massive models. Broader Impact Of our survey This study is the first experimental validation of DFA as an effective training method in a wide range of challenging tasks and neural networks architectures. This significantly broadens the applications of DFA, and more generally brings new insight on training techniques alternative to back- propagation. From neural rendering and recommender systems, to natural language processing or geometric learning, each of these applications has its own potential impact. Our task selection process was motivated by current trends in deep learning, as well as by technically appealing mechanisms (graph convolutions, attention). A limit of our survey is that our–arguably biased–selection of tasks cannot be exhaustive. Our experiments required substantial cloud compute resources, with state-of- the-art GPU hardware. Nevertheless, as this study provides new perspectives for hardware accelerator technologies, it may favor the application of neural networks in fields previously inaccessible because of computational limits. Future research on DFA should continue to demonstrate its use in novel contexts of interest as they are discovered. Of the considered applications Each of the applications considered in our study has a wide potential impact, consider for example the impact of textual bias in pretrained word embeddings [81]. We refer to [82] and references therein for a discussion of ethical concerns of AI applications. Of DFA as a training method DFA enables parallelization of the backward pass and places a single operation at the center of the training process, opening the prospect of reducing the power consumption of training chips by an order of magnitude [28]. Not only is more efficient training a path to more environmentally responsible machine learning [83], but it may lower the barrier of entry, supporting equality and sustainable development goals. A significant downside of moving from BP to DFA is a far more limited understanding of how to train models and how the trained models behave. There is a clear empirical understanding of the impact of techniques such as batch normalization or skip connections on the performance of BP; new insights need to be obtained for DFA. BP also enjoys decades of works on topics like adversarial attacks, interpretability, and fairness. Much of this work has to be cross-checked for alternative training methods, something we encourage further research to consider as the next step towards safely and responsively scaling up DFA. Of biologically motivated method Finally, a key motivation for this study was to demonstrate that learning challenging tasks was possible without weight transport. Biologically motivated methods are a more foundational research direction, and as such the possible long-term impact of our findings is harder to estimate under this light. However, fundamental research of this kind is important to open new pathways for ML and neuroscience. Acknowledgments and Disclosure of Funding We thank Igor Carron and Laurent Daudet for the general guidance on the subject of this investigation and the insightful comments, as well as the larger LightOn team for their support. References [1]P. J. Werbos.Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences. PhD thesis, Harvard University, 1974. [2]D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning internal representations by error propagation. InParallel Distributed Processing, volume 1, pages 318–362. MIT Press, 1986. [3]Max Jaderberg, Wojciech Marian Czarnecki, Simon Osindero, Oriol Vinyals, Alex Graves, David Silver, and Koray Kavukcuoglu. Decoupled neural interfaces using synthetic gradients. InProceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1627–1635, 2017. [4]Francis Crick. The recent excitement about neural networks.Nature, 337(6203):129–132, 1989. [5]Adam H Marblestone, Greg Wayne, and Konrad P Kording. Toward an integration of deep learning and neuroscience.Frontiers in computational neuroscience, 10:94, 2016. [6]Stephen Grossberg. Competitive learning: From interactive activation to adaptive resonance. Cognitive science, 11(1):23–63, 1987. [7]Javier R Movellan. Contrastive hebbian learning in the continuous hopfield model. InConnec- tionist models, pages 10–17. Elsevier, 1991. [8]Randall C O’Reilly. Biologically plausible error-driven learning using local activation differ- ences: The generalized recirculation algorithm.Neural computation, 8(5):895–938, 1996. [9]Ruslan Salakhutdinov and Geoffrey Hinton. Deep boltzmann machines. InArtificial intelligence and statistics, pages 448–455, 2009. [10]Yann Le Cun. Learning process in an asymmetric threshold network. InDisordered systems and biological organization, pages 233–240. Springer, 1986. [11]Yoshua Bengio. How auto-encoders could provide credit assignment in deep networks via target propagation.arXiv preprint arXiv:1407.7906, 2014. [12]Dong-Hyun Lee, Saizheng Zhang, Asja Fischer, and Yoshua Bengio. Difference target propaga- tion. InJoint european conference on machine learning and knowledge discovery in databases, pages 498–515. Springer, 2015. [13]Timothy P Lillicrap, Daniel Cownden, Douglas B Tweed, and Colin J Akerman. Random synap- tic feedback weights support error backpropagation for deep learning.Nature communications, 7(1):1–10, 2016. [14]Eugene Belilovsky, Michael Eickenberg, and Edouard Oyallon. Greedy layerwise learning can scale to imagenet. InInternational Conference on Machine Learning, pages 583–593, 2019. [15]Wojciech M Czarnecki, Simon Osindero, Max Jaderberg, Grzegorz Swirszcz, and Razvan Pascanu. Sobolev training for neural networks. InAdvances in Neural Information Processing Systems, pages 4278–4287, 2017. [16]Arild Nøkland and Lars Hiller Eidnes. Training neural networks with local error signals. In International Conference on Machine Learning, pages 4839–4850, 2019. [17]R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. InInternational Conference on Learning Representations, 2019. URLhttps://openreview.net/forum?id=Bklr3j0cKX. [18]Sindy Löwe, Peter O’Connor, and Bastiaan Veeling. Putting an end to end-to-end: Gradient- isolated learning of representations. InAdvances in Neural Information Processing Systems, pages 3033–3045, 2019. [19] Arild Nøkland. Direct feedback alignment provides learning in deep neural networks. In Advances in neural information processing systems, pages 1037–1045, 2016. [20]Sergey Bartunov, Adam Santoro, Blake Richards, Luke Marris, Geoffrey E Hinton, and Timothy Lillicrap. Assessing the scalability of biologically-motivated deep learning algorithms and architectures. InAdvances in Neural Information Processing Systems, pages 9368–9378, 2018. [21]Timothy P Lillicrap, Adam Santoro, Luke Marris, Colin J Akerman, and Geoffrey Hinton. Backpropagation and the brain.Nature Reviews Neuroscience, pages 1–12, 2020. [22]Natalia Caporale and Yang Dan. Spike timing–dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci., 31:25–46, 2008. [23]Julien Launay, Iacopo Poli, and Florent Krzakala. Principled training of neural networks with direct feedback alignment.arXiv preprint arXiv:1906.04554, 2019. [24]Qianli Liao, Joel Z Leibo, and Tomaso Poggio. How important is weight symmetry in back- propagation? InThirtieth AAAI Conference on Artificial Intelligence, 2016. [25]Theodore H Moskovitz, Ashok Litwin-Kumar, and LF Abbott. Feedback alignment in deep convolutional networks.arXiv preprint arXiv:1812.06488, 2018. [26]Will Xiao, Honglin Chen, Qianli Liao, and Tomaso Poggio. Biologically-plausible learning algorithms can scale to large datasets. InInternational Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=SygvZ209F7. [27]Mohamed Akrout, Collin Wilson, Peter C Humphreys, Timothy Lillicrap, and Douglas Tweed. Using weight mirrors to improve feedback alignment.arXiv preprint arXiv:1904.05391, 2019. [28]Julien Launay, Iacopo Poli, Kilian Müller, Igor Carron, Laurent Daudet, Florent Krzakala, and Sylvain Gigan. Light-in-the-loop: using a photonics co-processor for scalable training of neural networks, 2020. [29]Charlotte Frenkel.Bottom-Up and Top-Down Neuromorphic Processor Design: Unveiling Roads to Embedded Cognition. PhD thesis, UCL-Université Catholique de Louvain, 2020. [30]Eric Penner and Li Zhang. Soft 3d reconstruction for view synthesis.ACM Transactions on Graphics (TOG), 36(6):1–11, 2017. [31]John Flynn, Michael Broxton, Paul Debevec, Matthew DuVall, Graham Fyffe, Ryan Overbeck, Noah Snavely, and Richard Tucker. Deepview: View synthesis with learned gradient descent. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2367–2376, 2019. [32]Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines.ACM Transactions on Graphics (TOG), 38(4):1–14, 2019. [33]Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2437–2446, 2019. [34]Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images.ACM Transactions on Graphics (TOG), 38(4):65, 2019. [35]Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. InAdvances in Neural Information Processing Systems, pages 1119–1130, 2019. [36]Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis.arXiv preprint arXiv:2003.08934, 2020. [37]H Brendan McMahan, Gary Holt, David Sculley, Michael Young, Dietmar Ebner, Julian Grady, Lan Nie, Todd Phillips, Eugene Davydov, Daniel Golovin, et al. Ad click prediction: a view from the trenches. InProceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1222–1230, 2013. [38]Steffen Rendle. Factorization machines. In2010 IEEE International Conference on Data Mining, pages 995–1000. IEEE, 2010. [39]Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al. Wide & deep learning for recommender systems. InProceedings of the 1st workshop on deep learning for recommender systems, pages 7–10, 2016. [40]Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. Deepfm: a factorization-machine based neural network for ctr prediction.arXiv preprint arXiv:1703.04247, 2017. [41]Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. Deep & cross network for ad click predictions. InProceedings of the ADKDD’17, ADKDD’17, New York, NY, USA, 2017. Association for Computing Machinery. ISBN 9781450351942. doi: 10.1145/3124749.3124754. URLhttps://doi.org/10.1145/3124749.3124754. [42]Weiyu Cheng, Yanyan Shen, and Linpeng Huang. Adaptive factorization network: Learning adaptive-order feature interactions. InThirty-Fourth AAAI Conference on Artificial Intelligence, 2020. [43]J Wesley Hines. A logarithmic neural network architecture for unbounded non-linear function approximation. InProceedings of International Conference on Neural Networks (ICNN’96), volume 2, pages 1245–1250. IEEE, 1996. [44]Criteo. Kaggle contest dataset is now available for academic use!http://labs.criteo.com/ 2014/09/kaggle-contest-dataset-now-available-academic-use/, 2014. accessed on the 2020-05-20. [45]Maurizio Ferrari Dacrema, Paolo Cremonesi, and Dietmar Jannach. Are we really making much progress? a worrying analysis of recent neural recommendation approaches. InProceedings of the 13th ACM Conference on Recommender Systems, pages 101–109, 2019. [46]Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. Geometric deep learning: going beyond euclidean data.IEEE Signal Processing Magazine, 34 (4):18–42, 2017. [47]Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann Lecun. Spectral networks and locally connected networks on graphs. InInternational Conference on Learning Representations, pages http–openreview, 2014. [48]Mikael Henaff, Joan Bruna, and Yann LeCun. Deep convolutional networks on graph-structured data.arXiv preprint arXiv:1506.05163, 2015. [49]Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. InAdvances in neural information processing systems, pages 3844–3852, 2016. [50]Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations (ICLR), 2017. [51]Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. InProceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2, pages 729–734. IEEE, 2005. [52]Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model.IEEE Transactions on Neural Networks, 20(1):61–80, 2008. [53]Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. InInternational Conference on Learning Representations, 2016. [54]Matthias Fey, Jan Eric Lenssen, Frank Weichert, and Heinrich Müller. Splinecnn: Fast geometric deep learning with continuous b-spline kernels. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 869–877, 2018. [55]Petar Velickoviˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua´ Bengio. Graph attention networks. InInternational Conference on Learning Representations, 2018. URLhttps://openreview.net/forum?id=rJXMpikCZ. [56] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In3rd International Conference on Learning Representations, ICLR 2015, 2015. [57]Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InInternational Conference on Machine Learning, 2018. [58]Matthias Fey. Just jump: Dynamic neighborhood aggregation in graph neural networks. In ICLR Workshop on Representation Learning on Graphs and Manifolds, 2019. [59]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in neural information processing systems, pages 5998–6008, 2017. [60]Matthias Fey and Jan E. Lenssen. Fast graph representation learning with PyTorch Geometric. InICLR Workshop on Representation Learning on Graphs and Manifolds, 2019. [61]Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi- Rad. Collective classification in network data.AI magazine, 29(3):93–93, 2008. [62]Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InInternational Conference on Learning Representations, 2019. URLhttps: //openreview.net/forum?id=ryGs6iA5Km. [63]Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of machine learning research, 9(Nov):2579–2605, 2008. [64]David M Chan, Roshan Rao, Forrest Huang, and John F Canny. Gpu accelerated t-distributed stochastic neighbor embedding.Journal of Parallel and Distributed Computing, 131:1–13, 2019. [65]Thomas N Kipf and Max Welling. Variational graph auto-encoders.NIPS Workshop on Bayesian Deep Learning, 2016. [66]Alec Radford, Karthik Narasimhan, Time Salimans, and Ilya Sutskever. Improving language understanding with unsupervised learning.Technical report, OpenAI, 2018. [67]Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer.ArXiv, abs/1802.05751, 2018. [68]Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music.arXiv preprint arXiv:2005.00341, 2020. [69]Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 Confer- ence of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1423. URLhttps://www.aclweb.org/anthology/N19-1423. [70]Mohammad Shoeybi, Mostofa Ali Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using model parallelism.ArXiv, abs/1909.08053, 2019. [71]Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.arXiv preprint arXiv:2005.14165, 2020. [72]Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. SQuAD: 100,000+ questions for machine comprehension of text. InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin, Texas, Novem- ber 2016. Association for Computational Linguistics. doi: 10.18653/v1/D16-1264. URL https://www.aclweb.org/anthology/D16-1264. [73]Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don’t know: Unanswerable questions for SQuAD. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784–789, Melbourne, Australia, July 2018. Association for Computational Linguistics. doi: 10.18653/v1/P18-2124. URL https://www.aclweb.org/anthology/P18-2124. [74]Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. Superglue: A stickier benchmark for general-purpose language understanding systems. InAdvances in Neural Information Processing Systems, pages 3261–3275, 2019. [75]The Common Crawl Team. Common Crawl.https://commoncrawl.org, 2020. [76]Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classifica- tion. InACL. Association for Computational Linguistics, 2018. URLhttp://arxiv.org/ abs/1801.06146. [77]Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models.ArXiv, abs/1609.07843, 2017. [78]Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715–1725, Berlin, Germany, August 2016. Association for Computational Linguistics. doi: 10.18653/v1/P16-1162. URL https://www.aclweb.org/anthology/P16-1162. [79]Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization.International Conference on Learning Representations, 12 2014. [80]Grace W Lindsay. Attention in psychology, neuroscience, and machine learning.Frontiers in Computational Neuroscience, 14:29, 2020. [81]Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in neural information processing systems, pages 4349–4357, 2016. [82]Alexandra Luccioni and Yoshua Bengio. On the morality of artificial intelligence.arXiv preprint arXiv:1912.11945, 2019. [83]Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep learning in nlp.arXiv preprint arXiv:1906.02243, 2019. [84]Yi Tay, Dara Bahri, Donald Metzler, Da-Cheng Juan, Zhe Zhao, and Che Zheng. Synthesizer: Rethinking self-attention in transformer models.arXiv preprint arXiv:2005.00743, 2020. [85]Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond.arXiv preprint arXiv:1908.03265, 2019. [86]Alessandro Raganato, Yves Scherrer, and Jörg Tiedemann. Fixed encoder self-attention patterns in transformer-based machine translation.arXiv preprint arXiv:2002.10260, 2020. [87]Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high- performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché- Buc, E. Fox, and R. Garnett, editors,Advances in Neural Information Processing Systems 32, pages 8024–8035. Curran Associates, Inc., 2019. URLhttp://papers.neurips.cc/paper/ 9015-pytorch-an-imperative-style-high-performance-deep-learning-library. pdf. Appendix We first provide additional elements to corroborate our findings: alignment measurement (Section A), and shallow baselines (Section B). We then discuss the process of adapting the considered architectures for DFA (Section C), and the issue of weight transport in attention layers (Section D). We provide some supplementary results for NeRF (Section E), including details of performance on each scene of each datatset, and a discussion on possible mitigation of DFA shortcomings. Finally, we outline steps necessary for reproduction of this work (Section F). A Alignment Alignment measurement In feedback alignment methods, the forward weights learn toalignwith the random backward weights, making the delivered updates useful. This alignment can be quantified by measuring the cosine similarity between the gradient signal delivered by DFABi ay and the gradient signal BP would have deliveredWT <>. For learning to occur and DFA to work as a training method, there must be alignment. This can be measured numerically [23]. Measuring alignments allows to check whether or not the layers are effectively being trained by DFA, regardless of performance metrics. We note that any alignment value superior to 0 signifies that learning is occuring. Values closer to 1 indicate a better match with BP, but small alignment values are sufficient to enable learning. We report values measured at the deepest DFA layer. Recommender systems We measure alignment on the Criteo dataset, in the two architectures featuring non-conventional fully-connected layers: Deep & Cross and AFN. Alignment is measured after 15 epochs of training, and averaged over a random batch of 512 samples. Results are reported in table A.1. These alignment measurements indicate that learning is indeed occurring in the cross and logarithmic layers. High-variance of alignment in the cross layers is unique: it may be explained by the absence of non-linearity, and account for the difference in performance between BP and DFA on this architecture–which is higher than on the others. Table A.1: Alignment cosine similarity (higher is better, standard deviation in parenthesis) of recommender systems as measured on the Criteo dataset. Learning occurs in both architectures, and high variance may explain the larger performance gap on Deep & Cross compared to other methods. <> Graph convolutions We measure alignment on the Cora dataset, after 250 epochs of training, averaging values over every sample available–train, validation, and test split included. Results are reported in Table A.2. We observe high alignment values in all architectures, indicative that learning is indeed occuring. Slightly lower values in SplineConv and GATConv may be explained by the use of the Exponential Linear Unit (ELU) instead of the Rectified Linear Unit (ReLU) used as activation in other architectures. Table A.2: Alignment cosine similarity (standard deviation in parenthesis) of various graph convolu- tions architectures as measured on the Cora dataset. These values corroborate that DFA successfully trains all architectures considered. <> B Shallow baselines Shallow learning We compare DFA to BP, but also to shallow learning–where only the topmost layer is trained. While DFA may not reach the performance level of BP, it should still vastly Figure A.1: Comparisons of Tiny-NeRF trained with BP, DFA, and a shallow approach. Shallow training is insufficient to learn scene geometry. Lego scene from the NeRF synthetic dataset. <

> outperform shallow learning: failure to do so would mean that the weight updates delivered by DFA are useless. On a simple task like MNIST, a shallow baseline may be as high as 90%. However, given the difficulty of the tasks we consider, the shallow baseline is here usually much lower. NeRF Because NeRF models are expensive to train–up to 15 hours on a V100–we consider a simplified setup for the shallow baseline, NeRF-Tiny. This setup operates at half the full resolution of the training images available, runs for 5000 iterations only, and does away with view-dependant characteristics. Furthermore, the network is cut down to 3 layers of half the width of NeRF, and no coarse network is used to inform the sampling. We train this network on the Lego scene of the NeRF-Synthetic dataset, and compare results. Figure A.1 presents renders generated by NeRF-Tiny trained with BP, DFA, and a shallow approach. While BP and DFA delivers similar renders, shallow training fails to reproduce even basic scene geometry, instead outputting a diffuse cloud of colors. This highlights that while DFA may not reach a level of performance on-par with BP on NeRF, it nonetheless delivers meaningful updates enabling the learning of complex features. Recommender systems Because recommender systems require fine-tuning, we perform the same hyperparameter search for shallow learning than for DFA and BP. Results are detailed in Table A.3. Performance of shallow training is always well under BP and DFA–remember that0.001-levelmatter in recommender systems. In particular, in Deep & Cross, where there was the biggest gap between BP and DFA, the performance of the shallow method is extremely poor, well below the FM baseline. Finally, it is expected to see that DeepFM recovers more or less the performance of FM even with a shallow baseline. Table A.3: Shallow baseline for recommender system models on the Criteo dataset. Performance is always well below BP and DFA, as expected. <

> Graph convolutions We use the same hyperparameters as for DFA to produce the shallow baseline on graph datasets. Results are reported in Table A.4. Performance is always much worse than BP and DFA. GATConv recovers the best performance: random attention layers may still deliver useful features [84], as do random convolutions. Transformers In the baseline setting (optimizer and hyper-parameters of [59]), a Transformer trained in the shallow regime yields a perplexity of 428 on WikiText-103. We do not consider Table A.4: Shallow baseline for GCNNs on Cora, CiteSeer, and PubMed [61]. Performance is always well below BP and DFA. <> other settings, as the cost of training a Transformer is high and we do not expect any meaningful improvements–as with NeRF above. C Adapting architectures to DFA NeRF We use an architecture identical to the one used in [36], but based on the effective code implementation rather than the description in the paper 1 . During our tests, we have found that lowering the learning rate to <> rather than <> works best with DFA. Recommender systems For all training methods (BP, DFA, and shallow), we have conducted independent hyperparameter searches. We performed a grid search over the learning rate, from <> to <> in <> steps, as well as over the dropout probability, from <> to in <> steps (where applicable). On DeepFM, this search leads to reduce the learning rate from <> with BP to <> with DFA, but to keep the 0.5 dropout rate. On Deep & Cross, we reduce learning rate from <> to <>, with no dropout in both cases. In AFN, we reduce dropout from <> to <> and dropout from 0.3 to 0. Graph convolutions We manually test for a few hyperparameters configuration on the Cora dataset, focusing on learning rate, weight decay, and dropout. We do not consider architectural changes, such as changing the number of filters or of attention heads. For ChebConv and GraphConv, we reduce weight decay to <> instead of <>, and set the dropout rate to 0 and 0.1 respectively, instead of 0.5 with BP. For SplineConv, we find that no change in the hyperparameters are necessary. For GATConv, we reduce weight decay to <> instead of <> and reduce dedicated dropout layer to 0.1 instead of 0.6 but keep the 0.6 dropout rate within the GAT layer. Finally, on DNAConv we disable weight decay entirely, instead of an original value of <>, double the learning rate from <> to <>, and disable dropout entirely. In all cases, we share the backward random matrix across all nodes in a graph. Transformers The model hyper-parameters were fixed across all of our experiments, except for the number of attention heads in one case, that we will precise below, and dropout. We tested several values of dropout probability between 0 and 0.5, but found the original value of 0.1 to perform best. We manually tested a number of optimizers, optimizer parameters and attention mechanisms. We tested four combinations of optimizers and schedulers : Adam with the scheduler used in [59], Adam alone, RAdam [85] alone, and Adam with a scheduler that reduces the learning rate when the validation perplexity plateaus. We found it necessary to reduce the initial learning rate of Adam from <> to <>, although it could be set back to <> with a scheduler. We tried two values of 0.98 and 0.999. We also tried to change <> and observed some small differences that were not significant enough for the main text. Finally, we tried three attention mechanisms in addition to the standard multihead scaled dot-product attention: the dense and random (learnable) Synthesizers of [84], as well as the fixed attention patterns of [86]. The latter needed to be adapted to language modelling to prevent attending to future tokens, which led us to reduced the number of attention heads to 4. The backward random matrix is always shared across all tokens and batches. D Weight transport and attention We consider an attention layer operating on inputx. The queries, keys, and values are respectively <>, and <> is the dimension of the queries and keys. The layer performs: <> (4) When using DFA on attention, we deliver the random feedback to the top of the layer. Accordingly, to obtain updates toWQ ;WK ;andWV we still to have to backpropagate through the attention mechanism itself. This involves weight transport onWV , sacrificing some biological realism for simplicity. Overall weight transport between layers still does not occur, and updating the layers in parallel remains possible. Beside using FA or DFA within the attention layer, alternative mechanisms like the synthesizer [84]–which uses random attention in place of the query and key system–or fixed attention [86] can remove the need for weight transport. Implementing these mechanisms in DFA-trained Transformers, or other attention-powered architectures, will require further research. E Supplementary NeRF results Quantitative results We report per-scene scores for each dataset in Table A.5. BP values are taken from [36]. On three scenes of the synthetic datasets, NeRF-DFA even outperforms past state-of-the-art methods trained with BP. Note that Neural Volumes (NV) is not applicable to forward-facing view synthesis–as is required in LLFF-Real–and thus no results are reported. Qualitative results We report sample renders from the NeRF-Synthetic dataset (Figure A.2) and the LLFF-Real dataset (Figure A.2), for every scene available. However, we recommend readers to consult the supplementary video to make better sense of characteristics like multi-view consistency and view-dependent effects (most visible on the LLFF-Real Room scene). Table A.5: Per-scene PSNR for NeRF DFA and BP against other state-of-the-art methods on the Nerf-Synthetic and LLFF-Real. DFA performance is fairly homogeneous across each dataset and in line with the differences in other methods. <> Possible future directions Despite retranscribing scene geometry in a multi-view consistent way, NeRF produces renders of a lower quality when trained with DFA instead of BP. In particular, it struggles to transcribe small-scale details, resulting in "blurry" renders. Moreover, it displays high- frequency artefacts: not in the scene geometry, but in individual pixels taking values very distant from their neighborhood. Interestingly, this noise phenomenon is unique to NeRF-DFA: it is not observed on NeRF-BP with similar PSNR values (achieved during training) or on other methods with similar or lower PSNR. This leads us to hypothesize this is an aspect unique to DFA, possibly due to the alignment process. Indeed, DFA creates a bias on the weights, by encouraging them to be "aligned" with an arbitrary values dependant on the random matrix used. It is possible this could introduce random noise in the final renders–though we leave a more principled experiment to future research. To attempt to alleviate this issue, we first consider NeRF-Dual. In NeRF-Dual, we average the pixel-wise prediction between the fine and coarse network, to attempt to remove some of the noise. To do so, we first still use the coarse network to create a probability distribution for the hierarchical sampling. Then, we evaluate again both the coarse and fine networks at the locations informed by this probability distribution. Compared to vanilla NeRF, this requires an extra batch of evaluation of the coarse network for all rays–rougly speaking, this increases inference time by 30-50% depending on the coarse network architecture considered. We note that this is not applied during training, so that training times remain identical. Figure A.2 and Figure A.3 showcase comparisons between NeRF and NeRF-Dual trained with DFA on all scenes. When viewed at high resolution–such as in our supplementary video–the NeRF-Dual renders are more pleasing, especially for the full scenes. They remove most of the high-frequency noise, leading to smoother renders. However, this averaging process further blurs small-scale details in the render. This is especially visible in the NeRF-Synthetic dataset, on scenes like Ficus. Furthermore, NeRF-Dual introduces novel artefacts in the Mic and Ship scenes, with areas improperly colored with a violet tint. The cause for these artefacts is unknown, but they show that NeRF-Dual is far from a silver bullet. The PSNR is also minimally increased, by less than 0.5 per scene. Nevertheless, this shows some promise in possibilities to allievate the shortcomings of NeRF-DFA. It is possible that changes to the overall rendering process, or the use of classic image processing techniques, may help enhance the NeRF-DFA images. Finally, we also experimented with increasing the capacity of the fine network, by widening its layers to 512 neurons. We call this architecture NeRF-XL. However, we have not succeeded in getting PSNR values higher than with vanilla NeRF on DFA. In particular, the training process becomes much more cumbersome, as multi-GPU parallelism is needed to fit the model. It is possible that higher network capacity may help learning both the task at hand and to align simultaneously, but further work is required. F Reproducibility Hardware used All main experiments require at most a single NVIDIA V100 GPU with 16GB of memory to reproduce. Alignment measurement on large architectures (NeRF and Transformers) require a second identical GPU to keep a copy of the network to evaluate BP gradients. We estimate that a total of around 10,000 GPU-hours on V100s were necessary for this paper. Accordingly, we estimate the cloud-computing carbon impact of this paper to be of 1700 kgCO 2 eq 2 . However, without hyperparameter searches, our results can be reproduced with less than 500 GPU- hours on V100s, with most of that budget going to NeRF and Transformers. Implementation We use the shared random matrix trick from [23] to reduce memory use in DFA and enable its scaling to large networks. We use PyTorch [87] for all experiments. For reference implementation of the methods considered, we relied on various sources. Our NeRF implementation is based on the PyTorch implementation by Krishna Murthy 3 , with modifications to allow for proper test and validation, as well as DFA and multi-GPU support. For recommender systems, we use PyTorch Geometric [60] for all graph operations. Our Transformer implementation is our own. Our code is available as supplementary material. NeRF We provide training, testing, and rendering code along with the configurations used to obtain our results. An example to reproduce our results is given in the supplementary code repository. Given the computing cost associated with training a NeRF, we also provide our trained models. Recommender systems We provide bash scripts to reproduce the results in Table 2 and A.3, with the results of our hyperparameter search. We provide code to reproduce the results in Table A.1. Graph convolutions We provide the code to reproduce all of our results. Note that the t-SNE results are not exactly reproducible, as the CUDA implementation used is non-deterministic. Transformers We provide bash scripts to reproduce Table 5 and the shallow results. <

> Figure A.2: Sample renders for every scene of the NeRF-Synthetic dataset, for NeRF and NeRF-Dual trained with DFA. <

> Figure A.3: Sample renders for every scene of the LLFF-Real dataset, for NeRF and NeRF-Dual trained with DFA. <> <> <> <> <> <> Efficient Behavior of Small-World Networks We introduce the concept of efficiency of a network, measuring how efficiently it exchanges information. By using this simple measure small-world networks are seen as systems that are both globally and locally efficient. This allows to give a clear physical meaning to the concept of small-world, and also to perform a precise quantitative analysis of both weighted and unweighted networks. We study neural networks and man-made communication and transportation systems and we show that the underlying general principle of their construction is in fact a small-world principle of high efficiency. PACS numbers 89.70.+c, 05.90.+m, 87.18.Sn, 89.40.+k We live in a world of networks. In fact any complex system in nature can be modeled as a network, where vertices are the elements of the system and edges represent the interactions between them. Coupled biological and chemical systems, neural networks, social interacting species, computer networks or the Internet are only few of such examples [1]. Characterizing the structural properties of the networks is then of fundamental importance to understand the complex dynamics of these systems. A recent paper [2] has shown that the connection topology of some biological and social networks is neither completely regular nor completely random. These networks, there named small-worlds, in analogy with the concept of small-world phenomenon developed 30 years ago in social psychology [3], are in fact highly clustered like regular lattices, yet having small characteristics path lengths like random graphs. The original paper has triggered a large interest in the study of the properties of small-worlds (see ref. [4] for a recent review). Researchers have focused their attention on different aspects: study of the inset mechanism [5,7], dynamics [8] and spreading of diseases on small-worlds [9], applications to social net.works [10,11] and to the Internet [12,13]. In this letter we introduce the concept of efficiency of a network, measuring how efficiently information is exchanged over the net.work. By using efficiency small-world networks results as systems that are both globally and locally efficient. This formalization gives a clear physical meaning to the concept of small-world, and also allows a precise quantitative analysis of unweighted and weighted networks. We study several systems, like brains, communication and transportation networks, and show that the underlying general principle of their construction is in fact a small-world principle, provided attention is taken not to ignore an important observational property (closure). We start by reexamining the original formulation pro.posed in ref. [2]. There, a generic graph G with N vertices and K edges is considered. G is assumed to be unweighted, i.e. edges are all equal, sparse <<(K . N (N . 1)/2)>>, and connected. i.e. there exists at least one path connecting any couple of vertices with a infinite number of steps. G is therefore represented by simply giving the adjacency (or connection) matrix, i.e. the NxN matrix whose entry a_ij is 1 if there is an edge joining vertex i to vertex j, and 0 otherwise. An important quantity of G is the degree of vertex i, i.e. the number ki of edges incident with vertex i (the number of neighbors of i). The average value of ki is <>. Once {<>} is given it can be used to calculate the matrix of the short.est path lengths d_ij between two generic vertices i and j. The fact that G is assumed to be connected implies that dij is positive and infinite .i = j. In order to quantify the structural properties of G, [2] proposes to evaluate two different quantities: the characteristic path length L and the clustering coefficient C. L is the average distance be- tween two generic vertices <>, and C is a local property defined as <>. Here C_i is the number of edges existing in Gi, the subgraph of the neighbors of i, divided by the maximum possible number ki(ki . 1)/2. In [2] a simple method is considered to pro.duce a class of graphs with increasing randomness. The initial graph G is taken to be a one-dimensional lattice with each vertex connected to its k neighbors and with periodic boundary conditions. Rewiring each edge at ran.dom with probability p, G can be tuned in a continuous way from a regular lattice (p = 0) into a random graph (p = 1). For the regular lattice we expect <> and a high clustering coefficient <>, while for a random graph <> and <> [14,5]. Although in the two limit cases a large C is associated to a large L and vice versa a small C to a small L, the numerical experiment reveals an intermediate regime at small p where the system is highly clustered like regular lattices, yet having small characteristics path lengths like random graphs. This behavior is there called small-world and it is found to be a property of some social and biological networks analyzed [2]. Now we propose a more general set-up to investigate real networks. We will show that: the definition of small-world behavior can be given in terms of a single variable with a physical meaning, the efficiency E of the network. -1/L and C can be seen as first approximations of E evaluated resp. on a global and on a local scale. -we can drop all the restrictions on the system, like unweightedness, connectedness and sparseness. We represent a real network as a generic weighted (and possibly even non sparse and non connected) graph G. Such a graph needs two matrices to be described: the adjacency matrix {a_ij} defined as for the unweighted graph, and the matrix {<>} of physical distances. The number <> can be the space distance between the two vertices or the strength of their possible interaction: we suppose <> can be the geographical distance between stations in transportation systems (in such a case <> respects the triangle equality, though this is not a necessary assumption), the time taken to ex.change a packet of information between routers in the Internet, or the inverse v<>ity of chemical reactions along a direct connection in a biological system. Of course, in the particular case of an unweighted graph <>. The shortest path length dij between two generic points i and j is the smallest sum of the physical distances throughout all the possible paths in the graph from i to j. The matrix {<>} is therefore calculated by using the information contained both in matrix {a_ij} and in matrix {<>}. We have <>, the equality being valid when there is an edge between i and j. Let us now suppose that the system is parallel, i.e. every vertex sends information concurrently along the network, through its edges. The efficiency <> in the communication between vertex i and j can be then defined to be inversely proportional to the shortest distance: <>. When there is no path in the graph between i and j, <> and consistently <>. The average efficiency of G can be defined as: <> To normalize E we consider the ideal case G_id in which the graph G has all the <> possible edges. In such a case the information is propagated in the most efficient way since dij = .ij .i, j, and E assumes its maxi- <>. The efficiency <> <> considered in the following of the paper is always divided by <> and therefore <>. Though the equality E = 1 is valid when there is an edge between each couple of vertices, real networks can reach a high value of E. In our formalism, we can define the small-world be.haviour by using the single measure E to analyze both the local and global behavior, rather than two different variables L and C. The quantity in eq. (1) is the global efficiency of G and we therefore name it E_glob. Since E is defined also for a disconnected graph we can characterize the local properties of G by evaluating for each vertex i the efficiency of G_i, the subgraph of the neighbors of i. We define the local efficiency as the average efficiency of the local subgraphs, E loc =1/N E(Gi). This quantity plays a role similar to the clustering co.efficient C. Since <>, the local efficiency <> tells how much the system is fault tolerant, thus how efficient is the communication between the first neighbors of i when i is removed [15]. The definition of small-world can now be rephrased and generalized in terms of the information <>: small-world networks have high <> and <>, i.e. are very efficient in global and local communication. This definition is valid both for unweighted and weighted graphs, and can also be applied to disconnected and/or non sparse graphs. It is interesting to see the correspondence between our measure and the quantities L and C of [2] (or, correspondingly, <<1/L>> and C). The fundamental difference is that E_glob is the efficiency of a parallel systems, where all the nodes in the network concurrently exchange pack.ets of information (such are all the systems in [2], for example), while 1/L measures the efficiency of a sequential system (i.e. only one packet of information goes along the network). <> is a reasonable approximation of <>when there are not huge differences among the distances in the graph, and this can explain why L works reasonably well in the unweighted examples of [2]. But, in general 1/L can significantly depart from E_glob. For instance, in the Internet, having few computers with an extremely slow connection does not mean that the whole Internet diminishes by far its efficiency: in practice, the presence of such very slow computers goes unnoticed, be.cause the other thousands of computers are exchanging packets among them in a very efficient way. Here 1/L would give a number very close to zero (strictly 0 in the particular case when a computer is disconnected from the others and <>, while E_glob gives the correct efficiency measure of the Internet. We turn now our attention to the local properties of a network. C is only one among the many possible intuitive measures [10] of how well connected a cluster is. It can be shown that when in a graph most of its local subgraphs Gi are not sparse, then C is a good approximation of E_loc. Summing up there are not two different kinds of analysis to be done for the global and local scales, but just one with a very precise physical meaning: the efficiency in transporting information. We now illustrate the onset of the small-world in an un.weighted graph by means of the same example used in [2]. A regular lattice with <> and <> is rewired 2 with probability p and <> and <> are reported in <> as functions of p [16]. For <

> we expect the system to be inefficient on a global scale (E_glob . k/N log(N/K)) but locally efficient. The situation is inverted for the ran.dom graph. In fact at p =1 E_glob assumes a maximum value of 0.4, meaning 40% the efficiency of the ideal graph with an edge between each couple of vertices. This at the expenses of the fault tolerance (<>). <

> FIG. 1. FIG.1 Global and local efficiency for the graph example considered in [2]. A regular lattice with <> and <> is rewired with probability p. The small-world behavior results from the increase of E_glob caused by the introduction of only a few rewired edges (short cuts), which on the other side do not affect <>. At p � 0.1, E_glob has almost reached the value of the random graph, though <> has only diminished by very little from the value of 0.82 of the regular lattice. Small worlds have high E_glob and <>. The small-world behavior appears for intermediate values of p. It results from the fast increase of E_glob (for small p we find a linear increase of E_glob in the logarithmic horizontal scale) caused by the introduction of only a few rewired edges (short cuts), which on the other side do not affect <>. At p . 0.1, E_glob has almost reached the maximum value of 0.4, though <> has only diminished by very little from the maximum value of 0.82. For an unweighted case the description in terms of network efficiency resembles the approximation given in [2]. In particular we have checked that a good agreement with curves L(p) and C(p) [2] can be obtained by reporting <> and <>. Of course in such an example the short cuts connect at almost no cost vertices that would otherwise be much farther apart (because <>). On the other hand this is not true when we consider a weighted network. As real networks we consider first different examples of natural systems (neural networks), and then we turn our attention to man-made communication and transportation systems. 1) Neural Networks. Thanks to recent experiments neural structures can be studied at several levels of scale. Here we focus first on the analysis of the neuro-anatomical structure of cerebral cortex, and then on a simple nervous system at the level of wiring between neurons. The anatomical connections between cortical areas are of particular importance for their intricate relationship with the functional connectivity of the cerebral cortex [18]. We analyze two databases of cortico-cortical connections in the macaque and in the cat [19]. Tab.1 indicates the two networks are small-worlds [16]: they have high E_glob, respectively 52% and 69% the efficiency of the ideal graph with an edge between each couple of vertices (just slightly smaller than the best possible values of 57% and 70% obtained in random graphs) and high <>, respectively 70% and 83%, i.e. high fault tolerance [22]. These results indicate that in neural cortex each region is intermingled with the others and has grown following a perfect balance between local necessities (fault tolerance) and wide-scope interactions. Next we consider the neural network of C. elegans, the only case of a nervous system completely mapped at the level of neurons and chemical synapses [23]. Tab.1 shows that this is also a small-world network: C. elegans achieves both a 50% of global and local efficiency. Moreover the value of E_glob is similar to <>. This is a difference from cortex databases where fault tolerance is slightly privileged with respect to global communication. 2) Communication Networks. We have considered two of the most important large-scale communication net.works present nowadays: the World Wide Web and the Internet. Tab.2 shows that they have relatively high val.ues of E_glob (slightly smaller than the best possible val.ues obtained for random graphs) and <>. Despite the WWW is a virtual network and the Internet is a physical network, at a global scale they transport information essentially in the same way (as their E_glob�s are almost equal). At a local scale, the bigger <> in the WWW case can be explained both by the tendency in the WWW to create Web communities (where pages talking about the same subject tend to link to each other), and by the fact that many pages within the same site are often quickly connected to each other by some root or menu page. 3) Transport Networks. differently from previous databases the Boston subway transportation system (MBTA) can be better described by a weighted graph, the matrix {.ij } being given by the geographical distances between stations. If we consider the MBTA as an unweighted graph we obtain that it is apparently neither locally nor globally efficient (see Tab.3). On the other hand, when we take into account the geographical distances, we have E_glob =0.63: this shows the MBTA is a very efficient transportation system on a global scale, only 37% less efficient than the ideal subway with a di.rect tunnel from each station to the others. Even in the weighted case <> stays low (0.03), indicating a poor local behavior: differently from a neural network the MBTA is not fault tolerant and a damage in a station will dramatically affect the connection between the previous and the next station. The difference with respect to neural networks comes from different needs and priorities in the construction and evolution mechanism: when we build a subway system, the priority is given to the achievement of global efficiency, and not to fault tolerance. In fact a temporary problem in a station can be solved by other means: for example, walking, or taking a bus from the previous to the next station. That is to say, the MBTA is not a closed system: it can be considered, after all, as a subgraph of a wider transportation network. This property is of fundamental importance when we analyze a system: while global efficiency is without doubt the major characteristic, it is closure that somehow leads a system to have high local efficiency (without alternatives, there should be high fault-tolerance). The MBTA is not a closed system, and thus this explains why, unlike in the case of the brain, fault tolerance is not a critical issue. Indeed, if we increase the precision of the analysis and change the MBTA subway network by taking into account, for example, the Boston Bus System, this ex.tended transportation system comes back to be a small-world network (<>). Qualitatively similar results, confirming the similarity of construction principles, have been obtained for other undergrounds and for a wider transportation system consisting of all the main airplane and highway connections throughout the world [25]. Considering all the transportation alter.natives available at that scale makes again the system closed (there are no other reasonable routing alternatives), and so fault-tolerance comes back as a leading construction principle. Summing up, the introduction of the efficiency mea.sure allows to give a definition of small-world with a clear physical meaning, and provides important hints on why the original formulas of [2] work reasonably well in some cases, and where they fail. The efficiency measure al.lows a precise quantitative analysis of the information flow, and works both in the unweighted abstraction, and in the more realistic assumption of weighted networks. Finally, analysis of real data indicates that various existing (neural, communication and transport) networks exhibit the small-world behavior (even, in some cases, when their unweighted abstractions do not), substantiating the idea that the diffusion of small-world networks can be interpreted as the need to create networks that are both globally and locally efficient. [1] Y. Bar-Yam, Dynamics of Complex Systems (Addison-Wesley, Reading Mass, 1997). [2] D.J. Watts and S.H. Strogatz, Nature 393, 440 (1998). [3] S. Milgram, Physicol. Today, 2, 60 (1967). [4] M.E.J. Newman, cond-mat/0001118. [5] A. Barrat, M. Weigt, Europ. Phys. J. B 13, 547 (2000) [6] M. Marchiori and V. Latora, Physica A285, 539 (2000). [7] M. Barthelemy, L. Amaral, Phys. Rev. Lett. 82, 3180 (1999). [8] L. F. Lago-Fernandez et al, Phys. Rev. Lett. 84, 2758 (2000). [9] C. Moore and M.E.J. Newman, Phys. Rev. E61, 5678 (2000). [10] M.E.J. Newman, cond-mat/0011144. [11] L. A. N. Amaral, A. Scala, M. Barth�el�emy, and H. E. Stanley, Proc. Natl. Acad. Sci. 97, 11149 (2000). [12] R. Albert, H. Jeong, and A.-L. Barab�asi, Nature 401, 130 (1999). [13] A.-L. Barab�asi and R. Albert, Science 286, 509 (1999). [14] B. Bollob�as, Random Graphs (Academic, London, 1985). [15] Our concept of fault tolerance is different from the one adopted in R. Albert, H. Jeong, and A.-L. Barab�asi, Na.ture 406, 378 (2000); R. Cohen et al. Phys. Rev. Lett. 85, 2758 (2000), where the authors consider the response of the entire network to the removal of i. [16] Here and in the following the matrix {dij }i,j2G has been computed by using two different methods: the Floyd-Warshall (O(N 3 )) [17] and the Dijkstra algorithm (O(N 2 logN )) [10]. [17] G. Gallo and S. Pallottino, Ann. Oper. Res. 13, 3 (1988). [18] O. Sporns, G. Tononi, G.M. Edelman, Celebral Cortex 10, 127 (2000). [19] J.W.Scannell, Nature 386, 452 (1997). [20] M.P. Young, Phil.Trans.R.Soc B252, 13 (1993). [21] J.W. Scannell, M.P. Young and C. Blakemore, J. Neu.rosci. 15, 1463 (1995). [22] E. Sivan, H. Parnas and D. Dolev, Biol. Cybern. 81, 11.23 (1999). [23] J.G. White et. al., Phil. Trans. R. Soc. London B314,1 (1986). [24] T.B. Achacoso and W.S. Yamamoto, AY�s Neuroanatomy of C. elegans for Computation (CRC Press, FL, 1992). [25] M. Marchiori and V. Latora, in preparation. TABLE I. Macaque and cat cortico-cortical connections [19]. The macaque database contains N = 69 cortical areas and K = 413 connections [20]. The cat database has N = 55 cortical areas (including hippocampus, amygdala, entorhinal cortex and subiculum) and K = 564 (revised database and cortical parcellation from [21]). The nervous system of C. elegans consists of N = 282 neurons and K = 2462 links which can be either synaptic connections or gap junctions [24]. <

> TABLE II. Communication networks. Data on the World Wide Web from http://www.nd.edu/�networks contains N = 325729 documents and K = 1090108 links [12], while the Internet database is taken from http://moat.nlanr.net and has N = 6474 nodes and K = 12572 links. <> TABLE III. The Boston underground transportation system (MBTA) consists of N = 124 stations and K = 124 tunnels. The matrix {.ij } of the spatial distances between stations, used for the weighted case, has been calculated us.ing databases from http://www.mbta.com/ and the U.S. Na.tional Mapping Division. <> <> <> <> <> <> <> Efficient Processing of Deep Neural Networks: A Tutorial and Survey Vivienne Sze,Senior Member, IEEE,Yu-Hsin Chen,Student Member, IEEE,Tien-Ju Yang,Student Member, IEEE,Joel Emer,Fellow, IEEE Abstract Deep neural networks (DNNs) are currently widely representation of an input space. This is different from earlier used for many artificial intelligence (AI) applications including approaches that use hand-crafted features or rules designed by While DNNs experts. deliver state-of-the-art accuracy on many AI tasks, it comes at the The superior accuracy of DNNs, however, comes at the cost of high computational complexity. Accordingly, techniques that enable efficient processing of DNNs to improve energy cost of high computational complexity. While general-purpose efficiency and throughput without sacrificing application accuracy compute engines, especially graphics processing units (GPUs), or increasing hardware cost are critical to the wide deployment have been the mainstay for much DNN processing, increasingly of DNNs in AI systems. there is interest in providing more specialized acceleration of This article aims to provide a comprehensive tutorial and the DNN computation. This article aims to provide an overview survey about the recent advances towards the goal of enabling efficient processing of DNNs. Specifically, it will provide an of DNNs, the various tools for understanding their behavior, overview of DNNs, discuss various hardware platforms and and the techniques being explored to efficiently accelerate their architectures that support DNNs, and highlight key trends in computation. reducing the computation cost of DNNs either solely via hardware This paper is organized as follows: design changes or via joint hardware design and DNN algorithm changes. It will also summarize various development resources Section II provides background on the context of why that enable researchers and practitioners to quickly get started DNNs are important, their history and applications. in this field, and highlight important benchmarking metrics and Section III gives an overview of the basic components of design considerations that should be used for evaluating the DNNs and popular DNN models currently in use. rapidly growing number of DNN hardware designs, optionally including algorithmic co-designs, being proposed in academia Section IV describes the various resources used for DNN and industry. research and development. The reader will take away the following concepts from this Section V describes the various hardware platforms used article: understand the key design considerations for DNNs; be to process DNNs and the various optimizations used able to evaluate different DNN hardware implementations with to improve throughput and energy efficiency without benchmarks and comparison metrics; understand the trade-offs impacting application accuracy (i.e., produce bit-wise between various hardware architectures and platforms; be able to evaluate the utility of various DNN design techniques for efficient identical results). processing; and understand recent implementation trends and Section VI discusses how mixed-signal circuits and new opportunities. memory technologies can be used for near-data processing to address the expensive data movement that dominates throughput and energy consumption of DNNs. I. INTRODUCTION Section VII describes various joint algorithm and hardware Deep neural networks (DNNs) are currently the foundation optimizations that can be performed on DNNs to improve for many modern artificial intelligence (AI) applications [1]. both throughput and energy efficiency while trying to Since the breakthrough application of DNNs to speech recogni- minimize impact on accuracy. tion [2] and image recognition [3], the number of applications Section VIII describes the key metrics that should be that use DNNs has exploded. These DNNs are employed in a considered when comparing various DNN designs. myriad of applications from self-driving cars [4], to detecting cancer [5] to playing complex games [6]. In many of these II. B ACKGROUND ON DEEP NEURAL NETWORKS (DNN) domains, DNNs are now able to exceed human accuracy. The In this section, we describe the position of DNNs in thesuperior performance of DNNs comes from its ability to extract context of AI in general and some of the concepts that motivatedhigh-level features from raw sensory data after using statistical its development. We will also present a brief chronology oflearning over a large amount of data to obtain an effective the major steps in its history, and some current domains to which it is being applied. V. Sze, Y.-H. Chen and T.-J. Yang are with the Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technol- ogy, Cambridge, MA 02139 USA. (e-mail: sze@mit.edu; yhchen@mit.edu, A. Artificial Intelligence and DNNs tjy@mit.edu) J. S. Emer is with the Department of Electrical Engineering and Computer DNNs, also referred to as deep learning, are a part of Science, Massachusetts Institute of Technology, Cambridge, MA 02139 USA, the broad field of AI, which is the science and engineering and also with Nvidia Corporation, Westford, MA 01886 USA. (e-mail: jsemer@mit.edu) of creating intelligent machines that have the ability to 2 <

> Fig. 2. Connections to a neuron in the brain. <>,<>,<>, and b are the activations, weights, non-linear function and bias, respectively. (Figure adopted from [7].)Fig. 1. Deep Learning in the context of Artificial Intelligence. to be10 14 to10 15 synapses in the average human brain. achieve goals like humans do, according to John McCarthy, A key characteristic of the synapse is that it can scale the the computer scientist who coined the term in the 1950s. signal (x_i) crossing it as shown in Fig. 2. That scaling factor The relationship of deep learning to the whole of artificial can be referred to as a weight (<>), and the way the brain is intelligence is illustrated in Fig. 1. believed to learn is through changes to the weights associated Within artificial intelligence is a large sub-field called with the synapses. Thus, different weights result in different machine learning, which was defined in 1959 by Arthur Samuel responses to an input. Note that learning is the adjustment as the field of study that gives computers the ability to learn of the weights in response to a learning stimulus, while the without being explicitly programmed. That means a single organization (what might be thought of as the program) of the program, once created, will be able to learn how to do some brain does not change. This characteristic makes the brain an intelligent activities outside the notion of programming. This is excellent inspiration for a machine-learning-style algorithm. in contrast to purpose-built programs whose behavior is defined Within the brain-inspired computing paradigm there is a by hand-crafted heuristics that explicitly and statically define subarea called spiking computing. In this subarea, inspiration their behavior. is taken from the fact that the communication on the dendrites The advantage of an effective machine learning algorithm and axons are spike-like pulses and that the information being is clear. Instead of the laborious and hit-or-miss approach of conveyed is not just based on a spike’s amplitude. Instead, creating a distinct, custom program to solve each individual it also depends on the time the pulse arrives and that the problem in a domain, the single machine learning algorithm computation that happens in the neuron is a function of not just simply needs to learn, via a processes called training, to handle a single value but the width of pulse and the timing relationship each new problem. between different pulses. An example of a project that was Within the machine learning field, there is an area that is inspired by the spiking of the brain is the IBM TrueNorth [8]. often referred to as brain-inspired computation. Since the brain In contrast to spiking computing, another subarea of brain- is currently the best ‘machine’ we know for learning and inspired computing is called neural networks, which is the solving problems, it is a natural place to look for a machine focus of this article. 1 learning approach. Therefore, a brain-inspired computation is a program or algorithm that takes some aspects of its basic B. Neural Networks and Deep Neural Networks (DNNs) form or functionality from the way the brain works. This is in Neural networks take their inspiration from the notion that contrast to attempts to create a brain, but rather the program a neuron’s computation involves a weighted sum of the input aims to emulate some aspects of how we understand the brain values. These weighted sums correspond to the value scaling to operate. performed by the synapses and the combining of those values Although scientists are still exploring the details of how the in the neuron. Furthermore, the neuron doesn’t just output that brain works, it is generally believed that the main computational weighted sum, since the computation associated with a cascade element of the brain is the neuron. There are approximately of neurons would then be a simple linear algebra operation. 86 billion neurons in the average human brain. The neurons Instead there is a functional operation within the neuron that themselves are connected together with a number of elements is performed on the combined inputs. This operation appears entering them called dendrites and an element leaving them to be a non-linear function that causes a neuron to generate called an axon as shown in Fig. 2. The neuron accepts the an output only if the inputs cross some threshold. Thus by signals entering it via the dendrites, performs a computation on analogy, neural networks apply a non-linear function to the those signals, and generates a signal on the axon. These input weighted sum of the input values. We look at what some of and output signals are referred to as activations. The axon of those non-linear functions are in Section III-A1. one neuron branches out and is connected to the dendrites of many other neurons. The connections between a branch of the 1 Note: Recent work using TrueNorth in a stylized fashion allows it to be used to compute reduced precision neural networks [9]. These types of neural axon and a dendrite is called asynapse. There are estimated networks are discussed in Section VII-A. 3 <

> Fig. 3. Simple neural network example and terminology (Figure adopted (a) Compute the gradient of the loss (b) Compute the gradient of the lossfrom [7]). relative to the filter inputs relative to the weights <

> Fig. 4. An example of backpropagation through a neural network. <

> Fig. 3(a) shows a diagrammatic picture of a computational neural network. The neurons in the input layer receive some values and propagate them to the neurons in the middle layer and is referred to as training the network. Once trained, the of the network, which is also frequently called a ‘hidden program can perform its task by computing the output of layer’. The weighted sums from one or more hidden layers are the network using the weights determined during the training ultimately propagated to the output layer, which presents the process. Running the program with these weights is referred final outputs of the network to the user. To align brain-inspired to as inference. terminology with neural networks, the outputs of the neurons In this section, we will use image classification, as shown are often referred to as activations, and the synapses are often in Fig. 6, as a driving example for training and using a DNN. referred to as weights as shown in Fig. 3(a). We will use the When we perform inference using a DNN, we give an input activation/weight nomenclature in this article. image and the output of the DNN is a vector of scores, one for Fig. 3(b) shows an example of the computation at each each object class; the class with the highest score indicates the most likely class of object in the image. The overarching goal layer: <>, where W_ij ,x_i and y_j are the for training a DNN is to determine the weights that maximize weights, input activations and output activations, respectively, i=1 the score of the correct class and minimize the scores of the and <> is a non-linear function described in SectionIII-A1. incorrect classes. When training the network the correct class The bias term b is omitted from Fig. 3(b) for simplicity. is often known because it is given for the images used for Within the domain of neural networks, there is an area called training (i.e., the training set of the network). The gap between deep learning, in which the neural networks have more than the ideal correct scores and the scores computed by the DNN three layers, i.e., more than one hidden layer. Today, the typical based on its current weights is referred to as theloss(L). numbers of network layers used in deep learning range from Thus the goal of training DNNs is to find a set of weights to five to more than a thousand. In this article, we will generally minimize the average loss over a large training set. use the terminologydeep neural networks (DNNs)to refer to When training a network, the weights (wij ) are usually the neural networks used in deep learning. updated using a hill-climbing optimization process called DNNs are capable of learning high-level features with more gradient descent. A multiple of the gradient of the loss relative complexity and abstraction than shallower neural networks. An to each weight, which is the partial derivative of the loss with example that demonstrates this point is using DNNs to process respect to the weight, is used to update the weight (i.e., updated visual data. In these applications, pixels of an image are fed into <>, where <> is called the learning rate). Note <> the first layer of a DNN, and the outputs of that layer can be that this gradient indicates how the weights should change in ij interpreted as representing the presence of different low-level order to reduce the loss. The process is repeated iteratively to features in the image, such as lines and edges. At subsequent reduce the overall loss. layers, these features are then combined into a measure of the An efficient way to compute the partial derivatives of likely presence of higher level features, e.g., lines are combined the gradient is through a process called backpropagation. into shapes, which are further combined into sets of shapes. Backpropagation, which is a computation derived from the And finally, given all this information, the network provides a chain rule of calculus, operates by passing values backwards probability that these high-level features comprise a particular through the network to compute how the loss is affected by object or scene. This deep feature hierarchy enables DNNs to each weight. achieve superior performance in many tasks. This backpropagation computation is, in fact, very similar in form to the computation used for inference as shown in Fig. 4 [10]. 2 Thus, techniques for efficiently performing C. Inference versus Training Since DNNs are an instance of a machine learning algorithm, 2 To backpropagate through each filter: (1) compute the gradient of the loss the basic program does not change as it learns to perform its relative to the weights from the filter inputs (i.e., the forward activations) and given tasks. In the specific case of DNNs, this learning involves the gradients of the loss relative to the filter outputs; (2) compute the gradient of the loss relative to the filter inputs from the filter weights and the gradients determining the value of the weights (and bias) in the network, of the loss relative to the filter outputs. 4 inference can sometimes be useful for performing training. DNN Timeline It is, however, important to note a couple of points. First, backpropagation requires intermediate outputs of the network 1940s - Neural networks were proposed to be preserved for the backwards computation, thus training 1960s - Deep neural networks were proposed has increased storage requirements. Second, due to the gradients 1989 - Neural networks for recognizing digits (LeNet) use for hill-climbing, the precision requirement for training 1990s - Hardware for shallow neural nets (Intel ETANN) is generally higher than inference. Thus many of the reduced 2011 - Breakthrough DNN-based speech recognition (Microsoft)precision techniques discussed in Section VII are limited to inference only. 2012 - DNNs for vision start supplanting hand-crafted approaches (AlexNet)A variety of techniques are used to improve the efficiency and robustness of training. For example, often the loss from 2014+ - Rise of DNN accelerator research (Neuflow, DianNao...)multiple sets of input data, i.e., abatch, are collected before a single pass of weight update is performed; this helps to speed Fig. 5. A concise history of neural networks. ’Deep’ refers to the number of up and stabilize the training process. layers in the network. There are multiple ways to train the weights. The most common approach, as described above, is called supervised learning, where all the training samples are labeled (e.g., with amount of available information to train the networks. To learn the correct class).Unsupervised learning is another approach a powerful representation (rather than using a hand-crafted where all the training samples are not labeled and essentially approach) requires a large amount of training data. For example, the goal is to find the structure or clusters in the data.Semi- Facebook receives over 350 millions images per day, Walmart supervised learning falls in between the two approaches where creates 2.5 Petabytes of customer data hourly and YouTube only a small subset of the training data is labeled (e.g., use has 300 hours of video uploaded every minute. As a result, unlabeled data to define the cluster boundaries, and use the the cloud providers and many businesses have a huge amount small amount of labeled data to label the clusters). Finally, of data to train their algorithms. reinforcement learning can be used to the train weights such The second factor is the amount of compute capacity that given the state of the current environment, the DNN can available. Semiconductor device and computer architecture output what action the agent should take next to maximize advances have continued to provide increased computing expected rewards; however, the rewards might not be available capability, and we appear to have crossed a threshold where the immediately after an action, but instead only after a series of large amount of weighted sum computation in DNNs, which actions. is required for both inference and training, can be performed Another commonly used approach to determine weights is in a reasonable amount of time. fine-tuning, where previously-trained weights are available and The successes of these early DNN applications opened the are used as a starting point and then those weights are adjusted floodgates of algorithmic development. It has also inspired the for a new dataset (e.g., transfer learning) or for a new constraint development of several (largely open source) frameworks that (e.g., reduced precision). This results in faster training than make it even easier for researchers and practitioners to explore starting from a random starting point, and can sometimes result and use DNNs. Combining these efforts contributes to the third in better accuracy. factor, which is the evolution of the algorithmic techniques that This article will focus on the efficient processing of DNN have improved application accuracy significantly and broadened inference rather than training, since DNN inference is often the domains to which DNNs are being applied. performed on embedded devices (rather than the cloud) where An excellent example of the successes in deep learning can resources are limited as discussed in more details later. be illustrated with the ImageNet Challenge [14]. This challenge is a contest involving several different components. One of the components is an image classification task where algorithmsD. Development History are given an image and they must identify what is in the image,Although neural nets were proposed in the 1940s, the first as shown in Fig. 6. The training set consists of 1.2 millionpractical application employing multiple digital neurons didn’t images, each of which is labeled with one of 1000 objectappear until the late 1980s with the LeNet network for hand- categories that the image contains. For the evaluation phase,written digit recognition [11]3 . Such systems are widely used the algorithm must accurately identify objects in a test set ofby ATMs for digit recognition on checks. However, the early images, which it hasn’t previously seen.2010s have seen a blossoming of DNN-based applications with Fig. 7 shows the performance of the best entrants in thehighlights such as Microsoft’s speech recognition system in ImageNet contest over a number of years. One sees that 2011 [2] and the AlexNet system for image recognition in the accuracy of the algorithms initially had an error rate2012 [3]. A brief chronology of deep learning is shown in of 25% or more. In 2012, a group from the University ofFig. 5. Toronto used graphics processing units (GPUs) for their highThe deep learning successes of the early 2010s are believed compute capability and a deep neural network approach, namedto be a confluence of three factors. The first factor is the AlexNet, and dropped the error rate by approximately 10% [3]. Their accomplishment inspired an outpouring of deep learning In the early 1960s, single analog neuron systems were used for adaptive style algorithms that have resulted in a steady stream of filtering [12, 13]. 5 Speech and LanguageDNNs have significantly improved the accuracy of speech recognition [21] as well as many related tasks such as machine translation [2], natural language processing [22], and audio generation [23]. Machines Learning MedicalDNNs have played an important role in genomic to gain insight into the genetics of diseases such as autism, cancers, and spinal muscular atrophy [24–27]. <

> They have also been used in medical imaging to detect skin cancer [5], brain cancer [28] and breast cancer [29]. Fig. 6. Example of an image classification task. The machine learning platform takes in an image and outputs the confidence scores for a predefined set of classes. Game PlayRecently, many of the grand AI challenges involving game play have been overcome using DNNs. These successes also required innovations in training techniques and many rely on reinforcement learning [30]. DNNs have surpassed human level accuracy in playing Atari [31] as well as Go [6], where an exhaustive search of all possibilities is not feasible due to the unimaginably huge number of possible moves. RoboticsDNNs have been successful in the domain of <

> robotic tasks such as grasping with a robotic arm [32], motion planning for ground robots [33], visual navigation [4,34], control to stabilize a quadcopter [35] and Fig. 7. Results from the ImageNet Challenge [14]. driving strategies for autonomous vehicles [36]. DNNs are already widely used in multimedia applications today (e.g., computer vision, speech recognition). Looking improvements. forward, we expect that DNNs will likely play an increasingly In conjunction with the trend to deep learning approaches important role in the medical and robotics fields, as discussed for the ImageNet Challenge, there has been a corresponding above, as well as finance (e.g., for trading, energy forecasting, increase in the number of entrants using GPUs. From 2012 and risk assessment), infrastructure (e.g., structural safety, and when only 4 entrants used GPUs to 2014 when almost all traffic control), weather forecasting and event detection [37]. the entrants (110) were using them. This reflects the almost The myriad application domains pose new challenges to the complete switch from traditional computer vision approaches efficient processing of DNNs; the solutions then have to be to deep learning-based approaches for the competition. adaptive and scalable in order to handle the new and varied In 2015, the ImageNet winning entry, ResNet [15], exceeded forms of DNNs that these applications may employ. human-level accuracy with a top-5 error rate 4 below 5%. Since then, the error rate has dropped below 3% and more focus F. Embedded versus Cloud is now being placed on more challenging components of the The various applications and aspects of DNN processing competition, such as object detection and localization. These (i.e., training versus inference) have different computational successes are clearly a contributing factor to the wide range needs. Specifically, training often requires a large dataset 5 and of applications to which DNNs are being applied. significant computational resources for multiple weight-update iterations. In many cases, training a DNN model still takes several hours to multiple days and thus is typically performed E. Applications of DNN Many applications can benefit from DNNs ranging from in the cloud. Inference, on the other hand, can happen either multimedia to medical space. In this section, we will provide in the cloud or at the edge (e.g., IoT or mobile). examples of areas where DNNs are currently making an impact In many applications, it is desirable to have the DNN and highlight emerging areas where DNNs hope to make an inference processing near the sensor. For instance, in computer impact in the future. vision applications, such as measuring wait times in stores Image and VideoVideo is arguably the biggest of the or predicting traffic patterns, it would be desirable to extract big data. It accounts for over 70% of today’s Internet meaningful information from the video right at the image traffic [16]. For instance, over 800 million hours of video sensor rather than in the cloud to reduce the communication is collected daily worldwide for video surveillance [17]. cost. For other applications such as autonomous vehicles, Computer vision is necessary to extract meaningful infor- drone navigation and robotics, local processing is desired since mation from video. DNNs have significantly improved the the latency and security risks of relying on the cloud are accuracy of many computer vision tasks such as image too high. However, video involves a large amount of data, classification [14], object localization and detection [18], which is computationally complex to process; thus, low cost image segmentation [19], and action recognition [20]. hardware to analyze video is challenging yet critical to enabling 4 The top-5 error rate is measured based on whether the correct answer 5 One of the major drawbacks of DNNs is their need for large datasets to appears in one of the top 5 categories selected by the algorithm. prevent over-fitting during training. 6 attention has been given to hardware acceleration specifically Feed Forward Recurrent Fully-Connected Sparsely-Connected for RNNs. DNNs can be composed solely offully-connected(FC) layers (also referred to as multi-layer perceptrons, or MLP) as shown in the leftmost layer of Fig. 8(b). In a FC layer, all output activations are composed of a weighted sum of all input activations (i.e., all outputs are connected to all inputs). This requires a significant amount of storage and Thankfully, in many applications, we can remove current) networks some connections between the activations by setting the weights to zero without affecting accuracy. This results in a sparsely connected layer. A sparsely connected layer is illustrated in the rightmost layer of Fig. 8(b).these applications. Speech recognition enables us to seamlessly We can also make the computation more efficient by limitinginteract with electronic devices, such as smartphones. While the number of weights that contribute to an output. This sort ofcurrently most of the processing for applications such as Apple structured sparsity can arise if each output is only a functionSiri and Amazon Alexa voice services is in the cloud, it is of a fixed-size window of inputs. Even further efficiency canstill desirable to perform the recognition on the device itself to be gained if the same set of weights are used in the calculationreduce latency and dependency on connectivity, and to improve of every output. This repeated use of the same weight values is privacy and security. calledweight sharingand can significantly reduce the storageMany of the embedded platforms that perform DNN infer- requirements for weights.ence have stringent energy consumption, compute and memory An extremely popular windowed and weight-shared DNNcost limitations; efficient processing of DNNs have thus become layer arises by structuring the computation as a convolution,of prime importance under these constraints. Therefore, in this as shown in Fig. 9(a), where the weighted sum for each outputarticle, we will focus on the compute requirements for inference activation is computed using only a small neighborhood of inputrather than training. activations (i.e., all weights beyond beyond the neighborhood are set to zero), and where the same set of weights are shared for every output (i.e., the filter is space invariant). Such convolution- III. OVERVIEW OF DNN'S DNNs come in a wide variety of shapes and sizes depending based layers are referred to as convolutional (CONV) layers. on the application. The popular shapes and sizes are also evolving rapidly to improve accuracy and efficiency. In all A. Convolutional Neural Networks (CNNs)cases, the input to a DNN is a set of values representing the A common form of DNNs isConvolutional Neural Netsinformation to be analyzed by the network. For instance, these (CNNs), which are composed of multiple CONV layers asvalues can be pixels of an image, sampled amplitudes of an shown in Fig. 10. In such networks, each layer generates aaudio wave or the numerical representation of the state of some successively higher-level abstraction of the input data, calledsystem or game. afeature map(fmap), which preserves essential yet uniqueThe networks that process the input come in two major information. Modern CNNs are able to achieve superior per-forms: feed forward and recurrent as shown in Fig. 8(a). In formance by employing a very deep hierarchy of layers. CNNfeed-forward networks all of the computation is performed as a are widely used in a variety of applications including imagesequence of operations on the outputs of a previous layer. The understanding [3], speech recognition [39], game play [6],final set of operations generates the output of the network, for robotics [32], etc. This paper will focus on its use in imageexample a probability that an image contains a particular object, processing, specifically for the task of image classification [3].the probability that an audio sequence contains a particular Each of the CONV layers in CNN is primarily composed ofword, a bounding box in an image around an object or the high-dimensional convolutions as shown in Fig. 9(b). In thisproposed action that should be taken. In such DNNs, the computation, the input activations of a layer are structured asnetwork has no memory and the output for an input is always a set of 2-Dinput feature maps(ifmaps), each of which isthe same irrespective of the sequence of inputs previously given called achannel. Each channel is convolved with a distinctto the network. 2-D filter from the stack of filters, one for each channel; thisIn contrast, recurrent neural networks (RNNs), of which stack of 2-D filters is often referred to as a single 3-D filter.Long Short-Term Memory networks (LSTMs) [38] are a The results of the convolution at each point are summed acrosspopular variant, have internal memory to allow long-term all the channels. In addition, a 1-D bias can be added to thedependencies to affect the output. In these networks, some filtering results, but some recent networks [15] remove itsintermediate operations generate values that are stored internally usage from parts of the layers. The result of this computationin the network and used as inputs to other operations in is the output activations that comprise one channel ofoutputconjunction with the processing of a later input. In this article, feature map(ofmap). Additional 3-D filters can be used onwe will focus on feed-forward networks since (1) the major computation in RNNs is still the weighted sum, which is 6 Note: the structured sparsity in CONV layers is orthogonal to the sparsity covered by the feed-forward networks, and (2) to-date little that occurs from network pruning as described in Section VII-B2. 7 after the CONV layers for classification purposes. A FC layer Fully Connected also applies filters on the ifmaps as in the CONV layers, but × × the filters are of the same size as the ifmaps. Therefore, it does not have the weight sharing property of CONV layers. Optional Eq. (1) still holds for the computation of FC layers with a Fig. 10. Convolutional Neural Networks. few additional constraints on the shape parameters: <>, <>,<>, and <>. In addition to CONV and FC layers, various optional layers the same input to create additional output channels. Finally, can be found in a DNN such as the non-linearity, pooling, multiple input feature maps may be processed together as a and normalization. The function and computations for each of batchto potentially improve reuse of the filter weights. these layers are discussed next. Given the shape parameters in Table I, the computation of 1) Non-Linearity:A non-linear activation function is typically applied after each CONV or FC layer. Various non-linear functions are used to introduce non-linearity into the DNN as shown in Fig. 11. These include historically conventional non- <> <> linear functions such as sigmoid or hyperbolic tangent as well <> as rectified linear unit (ReLU) [40], which has become popular <>; in recent years due to its simplicity and its ability to enable <>; fast training. Variations of ReLU, such as leaky ReLU [41], (1) parametric ReLU [42], and exponential LU [43] have also been O,I,W and B are the matrices of the of_maps, if_maps, filters explored for improved accuracy. Finally, a non-linearity called and biases, respectively.Uis a given stride size. Fig. 9(b) maxout, which takes the max value of two intersecting linear shows a visualization of this computation (ignoring biases). functions, has shown to be effective in speech recognition To align the terminology of CNNs with the generic DNN, tasks [44, 45]. filters are composed of weights (i.e., synapses) 2) Pooling: A variety of computations that reduce the input and output feature maps (if_maps, of_maps) are dimensionality of a feature map are referred to as pooling. composed of activations (i.e., input and output neurons) Pooling, which is applied to each channel separately, enables DNN is run only once), which is more consistent with what would likely be deployed in real-time and/or energy-constrained LeNet[11] was one of the first CNN approaches introduced in 1989. It was designed for the task of digit classification in <

> grayscale images of size 28x28. The most well known version, LeNet-5, contains two CONV layers and two FC layers [48]. Fig. 12. Various forms of pooling (Figure adopted from Caffe Tutorial [46]). Each CONV layer uses filters of size 5x5 (1 channel per filter) with 6 filters in the first layer and 16 filters in the second layer. the network to be robust and invariant to small shifts and Average pooling of 2x2 is used after each convolution and a distortions. Pooling combines, or pools, a set of values in sigmoid is used for the non-linearity. In total, LeNet requires its receptive field into a smaller number of values. It can be 60k weights and 341k multiply-and-accumulates (MACs) per configured based on the size of its receptive field (e.g., 2x2) image. LeNet led to CNNs’ first commercial success, as it was and pooling operation (e.g., max or average), as shown in deployed in ATMs to recognize digits for check deposits. Fig. 12. Typically pooling occurs on non-overlapping blocks AlexNet[3] was the first CNN to win the ImageNet Challenge (i.e., the stride is equal to the size of the pooling). Usually a in 2012. It consists of five CONV layers and three FC layers. stride of greater than one is used such that there is a reduction Within each CONV layer, there are 96 to 384 filters and the in the dimension of the representation (i.e., feature map). filter size ranges from 3x3 to 11x11, with 3 to 256 channels 3) Normalization:Controlling the input distribution across each. In the first layer, the 3 channels of the filter correspond layers can help to significantly speed up training and improve to the red, green and blue components of the input image. accuracy. Accordingly, the distribution of the layer input A ReLU non-linearity is used in each layer. Max pooling of activations <> are normalized such that it has a zero mean 3x3 is applied to the outputs of layers 1, 2 and 5. To reduce and a unit standard deviation. In batch normalization (BN), computation, a stride of 4 is used at the first layer of the the normalized value is further scaled and shifted, as shown network. AlexNet introduced the use of LRN in layers 1 and in Eq. (2), where the parameters <> are learned from 2 before the max pooling, though LRN is no longer popular training [47].X is a small constant to avoid numerical problems. in later CNN models. One important factor that differentiates Prior to this, local response normalization (LRN) [3] was AlexNet from LeNet is that the number of weights is much used, which was inspired by lateral inhibition in neurobiology larger and the shapes vary from layer to layer. To reduce the where excited neurons (i.e., high value activations) should amount of weights and computation in the second CONV layer, subdue its neighbors (i.e., cause low value activations); however, the 96 output channels of the first layer are split into two groups BN is now considered standard practice in the design of of 48 input channels for the second layer, such that the filters in CNNs while LRN is mostly deprecated. Note that while LRN the second layer only have 48 channels. Similarly, the weights usually is performed after the non-linear function, BN is mostly in fourth and fifth layer are also split into two groups. In total, performed between the CONV or FC layer and the non-linear AlexNet requires 61M weights and 724M MACs to process one 227x227 input image. Overfeat[49] has a very similar architecture to AlexNet with <> (2) five CONV layers and three FC layers. The main differences <> are that the number of filters is increased for layers 3 (384 to 512), 4 (384 to 1024), and 5 (256 to 1024), layer 2 is notB. Popular DNN Models split into two groups, the first fully connected layer only has Many DNN models have been developed over the past 3072 channels rather than 4096, and the input size is 231x231 two decades. Each of these models has a different ‘network rather than 227x227. As a result, the number of weights grows architecture’ in terms of number of layers, layer types, layer to 146M and the number of MACs grows to 2.8G per image. shapes (i.e., filter size, number of channels and filters), and Overfeat has two different models: fast (described here) and connections between layers. Understanding these variations accurate. The accurate model used in the ImageNet Challenge and trends is important for incorporating the right flexibility gives a 0.65% lower top-5 error rate than the fast model at the in any efficient DNN engine. cost of 1.9% more MACs In this section, we will give an overview of various popular VGG-16[50] goes deeper to 16 layers consisting of 13 DNNs such as LeNet [48] as well as those that competed in CONV layers and 3 FC layers. In order to balance out the and/or won the ImageNet Challenge [14] as shown in Fig. 7, cost of going deeper, larger filters (e.g., 5x5) are built from most of whose models with pre-trained weights are publicly multiple smaller filters (e.g., 3x3), which have fewer weights, available for download; the DNN models are summarized in to achieve the same receptive fields as shown in Fig. 13(a). Table II. Two results for top-5 error results are reported. In the As a result, all CONV layers have the same filter size of 3x3. first row, the accuracy is boosted by using multiple crops from In total, VGG-16 requires 138M weights and 15.5G MACs the image and an ensemble of multiple trained models (i.e., to process one 224224 input image. VGG has two different the DNN needs to be run several times); these results were models: VGG-16 (described here) and VGG-19. VGG-19 gives used to compete in the ImageNet Challenge. The second row a 0.1% lower top-5 error rate than VGG-16 at the cost of reports the accuracy if only a single crop was used (i.e., the 1.27more MACs. 9 <

> <

> Fig. 13. Decomposing larger filters into smaller filters. Fig. 14. Inception module from GoogleNet [51] with example channel lengths. GoogLeNet[51] goes even deeper with 22 layers. It in- troduced an inception module, shown in Fig. 14, which is composed of parallel connections, whereas previously there was only a single serial connection. Different sized filters (i.e., 1x1, 3x3, 5x5), along with 3x3 max-pooling, are used for each parallel connection and their outputs are concatenated for the module output. Using multiple filter sizes has the effect of processing the input at multiple scales. For improved training speed, GoogLeNet is designed such that the weights ReLU and the activations, which are stored for backpropagation during <> training, could all fit into the GPU memory. In order to reduce the number of weights, 1x1 filters are applied as a ‘bottleneck’ to reduce the number of channels for each filter [52]. The 22 layers consist of three CONV layers, followed by 9 inceptions layers (each of which are two CONV layers deep), and one FC layer. Since its introduction in 2014, GoogleNet (also referred <

> to as Inception) has multiple versions: v1 (described here), v3 7 smaller 1-D filters as shown in Fig. 13(b) to reduce number Fig. 15. Shortcut module from ResNet [15]. Note that ReLU following last of MACs and weights in order to go deeper to 42 layers. CONV layer in short cut is after the addition. In conjunction with batch normalization [47], v3 achieves over 3% lower top-5 error than v1 with 2.5% increase in is used. This is similar to the LSTM networks that are used for computation [53]. Inception-v4 uses residual connections [54], sequential data. ResNet also uses the ‘bottleneck’ approach of described in the next section, for a 0.4% reduction in error. using 1x1 filters to reduce the number of weight parameters.ResNet[15], also known as Residual Net, uses residual As a result, the two layers in the shortcut module are replace d connections to go even deeper (34 layers or more). It was by three layers (1x1, 3x3, 1x1) where the 1x1 reduces and the first entry DNN in ImageNet Challenge that exceeded then increases (restores) the number of weights. ResNet-50human-level accuracy with a top-5 error rate below 5%. One consists of one CONV layer, followed by 16 shortcut layers of the challenges with deep networks is the vanishing gradient (each of which are three CONV layers deep), and one FC during training: as the error backpropagates through the network layer; it requires 25.5M weights and 3.9G MACs per image.the gradient shrinks, which affects the ability to update the There are various versions of ResNet with multiple depths weights in the earlier layers for very deep networks. Residual (e.g.,without bottleneck:18, 34;with bottleneck:50, 101, 152).net introduces a ‘shortcut’ module which contains an identity The ResNet with 152 layers was the winner of the ImageNet connection such that the weight layers (i.e., CONV layers) Challenge requiring 11.3G MACs and 60M weights. Compared can be skipped as shown in Fig. 15. Rather than learning the to ResNet-50, it reduces the top-5 error by around 1% at the function for the weight layersF(x), the shortcut module learns cost of 2.9% more MACs and 2.5% more weights.the residual mapping <>. Initially, <> is zero and the identity connection is taken; then gradually during Several trends can be observed in the popular DNNs shown training, the actual forward connection through the weight layer in Table II. Increasing the depth of the network tends to provide higher accuracy. Controlling for number of weights, a deeper 7 v2 is very similar to v3. network can support a wider range of non-linear functions that are more discriminative and also provides more levels B. Models of hierarchy in the learned representation [15,50,51,55]. Pretrained DNN models can be downloaded from variousThe number of filter shapes continues to vary across layers, websites [56–59] for the various different frameworks. It shouldthus flexibility is still important. Furthermore, most of the be noted that even for the same DNN (e.g., AlexNet) thecomputation has been placed on CONV layers rather than FC accuracy of these models can vary by around 1% to 2%layers. In addition, the number of weights in the FC layers is depending on how the model was trained, and thus the resultsreduced and in most recent networks (since GoogLeNet) the do not always exactly match the original publication.CONV layers also dominate in terms of weights. Thus, the focus of hardware implementations should be on addressing the efficiency of the CONV layers, which in many domains C. Popular Datasets for Classification are increasingly important. It is important to factor in the difficulty of the task when comparing different DNN models. For instance, the task of IV. DNN DEVELOPMENT RESOURCES classifying handwritten digits from the MNIST dataset [62] is much simpler than classifying an object into one of 1000 One of the key factors that has enabled the rapid development classes as is required for the ImageNet dataset [14](Fig. 16). of DNNs is the set of development resources that have been It is expected that the size of the DNNs (i.e., number ofmade available by the research community and industry. These weights) and the number of MACs will be larger for the moreresources are also key to the development of DNN accelerators difficult task than the simpler task and thus require moreby providing characterizations of the workloads and facilitating energy and have lower throughput. For instance, LeNet-5[48]the exploration of trade-offs in model complexity and accuracy. is designed for digit classification, while AlexNet[3], VGG-This section will describe these resources such that those who 16[50], GoogLeNet[51], and ResNet[15] are designed for theare interested in this field can quickly get started. There are many AI tasks that come with publicly availableA. Frameworks For ease of DNN development and to enable sharing of Public datasets are important for comparing the accuracy of trained networks, several deep learning frameworks have been different approaches. The simplest and most common task developed from various sources. These open source libraries is image classification, which involves being given an entire contain software libraries for DNNs. Caffe was made available image, and selecting 1 of N classes that the image most likely in 2014 from UC Berkeley [46]. It supports C, C++, Python belongs to. There is no localization or detection. and MATLAB. Tensorflow was released by Google in 2015, MNISTis a widely used dataset for digit classification and supports C++ and python; it also supports multiple CPUs that was introduced in 1998 [62]. It consists of 2828 pixel and GPUs and has more flexibility than Caffe, with the grayscale images of handwritten digits. There are 10 classes computation expressed as dataflow graphs to manage the (for 10 digits) and 60,000 training images and 10,000 test tensors (multidimensional arrays). Another popular framework images. LeNet-5 was able to achieve an accuracy of 99.05% is Torch, which was developed by Facebook and NYU and when MNIST was first introduced. Since then the accuracy has supports C, C++ and Lua. There are several other frameworks increased to 99.79% using regularization of neural networks such as Theano, MXNet, CNTK, which are described in [60]. with dropconnect [63]. Thus, MNIST is now considered a fairly There are also higher-level libraries that can run on top of easy dataset. the aforementioned frameworks to provide a more universal CIFARis a dataset that consists of 3232 pixel colored experience and faster development. One example of such images of of various objects, which was released in 2009 [64]. libraries is Keras, which is written in Python and supports CIFAR is a subset of the 80 million Tiny Image dataset [65]. Tensorflow, CNTK and Theano. CIFAR-10 is composed of 10 mutually exclusive classes. There The existence of such frameworks are not only a convenient are 50,000 training images (5000 per class) and 10,000 test aid for DNN researchers and application designers, but they images (1000 per class). A two-layer convolutional deep belief are also invaluable for engineering high performance or more network was able to achieve 64.84% accuracy on CIFAR-10 efficient DNN computation engines. In particular, because the when it was first introduced [66]. Since then the accuracy has frameworks make heavy use of a set primitive operations, increased to 96.53% using fractional max pooling [67]. such processing of a CONV layer, they can incorporate use of ImageNetis a large scale image dataset that was first optimized software or hardware accelerators. This acceleration introduced in 2010; the dataset stabilized in 2012 [14]. It is transparent to the user of the framework. Thus, for example, contains images of 256256 pixel in color with 1000 classes. most frameworks can use Nvidia’s cuDNN library for rapid The classes are defined using the WordNet as a backbone to execution on Nvidia GPUs. Similarly, transparent incorporation handle ambiguous word meanings and to combine together of dedicated hardware accelerators can be achieved as was synonyms into the same object category. In otherwords, there done with the Eyeriss chip [61]. is a hierarchy for the ImageNet categories. The 1000 classes Finally, these frameworks are a valuable source of workloads were selected such that there is no overlap in the ImageNet for hardware researchers. They can be used to drive experi- hierarchy. The ImageNet dataset contains many fine-grained mental designs for different workloads, for profiling different categories including 120 different breeds of dogs. There are workloads and for exploring hardware-software trade-offs. 1.3M training images (732 to 1300 per class), 100,000 testing 11 <

> TABLE II SUMMARY OF POPULAR DNN S [3,15,48,50,51]. y ACCURACY IS MEASURED BASED ON TOP -5 ERROR ON IMAGE NET [14]. z THIS VERSION OF LE NET -5 HAS 431 K WEIGHTS FOR THE FILTERS AND REQUIRES 2.3M MAC S PER IMAGE ,AND USES RE LU RATHER THAN SIGMOID . be localized and classified (out of 1000 classes). The DNN outputs the top five categories and top five bounding box locations. There is no penalty for identifying an object that is in the image but not included in the ground truth. For object detection, all objects in the image must be localized and classified (out of 200 classes). The bounding box for all objects in these categories must be labeled. Objects that are not labeled are penalized as are duplicated detections. Fig. 16. MNIST (10 classes, 60k training, 10k testing) [62] vs. ImageNet (1000 classes, 1.3M training, 100k testing)[14] dataset. Beyond ImageNet, there are also other popular image datasets for computer vision tasks. For object detection, there images (100 per class) and 50,000 validation images (50 per is the PASCAL VOC (2005-2012) dataset that contains 11k class). images representing 20 classes (27k object instances, 7k of The accuracy of the ImageNet Challenge are reported using which has detailed segmentation) [68]. For object detection, two metrics: Top-5 and Top-1 error. Top-5 error means that if segmentation and recognition in context, there is the MS COCO any of the top five scoring categories are the correct category, dataset with 2.5M labeled instances in 328k images (91 object it is counted as a correct classification. The Top-1 requires categories) [69]; compared to ImageNet, COCO has fewer that the top scoring category be correct. In 2012, the winner categories but more instances per category, which is useful for of the ImageNet Challenge (AlexNet) was able to achieve an precise 2-D localization. COCO also has more labeled instances accuracy of 83.6% for the top-5 (which is substantially better per image to potentially help with contextual information. than the 73.8% which was second place that year that did not Most recently even larger scale datasets have been made use DNNs); it achieved 61.9% on the top-1 of the validation available. For instance, Google has an Open Images dataset set. In 2017, the highest accuracy was 97.7% for the top-5. with over 9M images [70], spanning 6000 categories. There is In summary of the various image classification datasets, it also a YouTube dataset with 8M videos (0.5M hours of video) is clear that MNIST is a fairly easy dataset, while ImageNet covering 4800 classes [71]. Google also released an audio is a challenging one with a wider coverage of classes. Thus dataset comprised of 632 audio event classes and a collection in terms of evaluating the accuracy of a given DNN, it is of 2M human-labeled 10-second sound clips [72]. These large important to consider that dataset upon which the accuracy is datasets will be evermore important as DNNs become deeper measured. with more weight parameters to train. Undoubtedly, both larger datasets and datasets for new D. Datasets for Other Tasks domains will serve as important resources for profiling and exploring the efficiency of future DNN engines.Since the accuracy of the state-of-the-art DNNs are perform- ing better than human-level accuracy on image classification tasks, the ImageNet Challenge has started to focus on more V. H ARDWARE FOR DNN P ROCESSING difficult tasks such as single-object localization and object Due to the popularity of DNNs, many recent hardware detection. For single-object localization, the target object must platforms have special features that target DNN processing. For 12 instance, the Intel Knights Landing CPU features special vector instructions for deep learning; the Nvidia PASCAL GP100 GPU features 16-bit floating point (FP16) arithmetic support to perform two FP16 operations on a single precision core for faster deep learning computation. Systems have also been built specifically for DNN processing such as Nvidia DGX-1 and Facebook’s Big Basin custom DNN server [73]. DNN inference has also been demonstrated on various embedded System-on- Chips (SoC) such as Nvidia Tegra and Samsung Exynos as well as FPGAs. Accordingly, it’s important to have a good understanding of how the processing is being performed on these platforms, and how application-specific accelerators can <

> be designed for DNNs for further improvement in throughput and energy efficiency. Fig. 17. Highly-parallel compute paradigms. The fundamental component of both the CONV and FC lay- ers are the multiply-and-accumulate (MAC) operations, which can be easily parallelized. In order to achieve high performance, highly-parallel compute paradigms are very commonly used, including both temporal and spatial architectures as shown in <> Fig. 17. The temporal architectures appear mostly in CPUs parallelism such as vectors (SIMD) or parallel threads (SIMT). Such temporal architecture use a centralized control for a large number of ALUs. These ALUs can only fetch data from the memory hierarchy and cannot communicate directly with each other. In contrast, spatial architectures use dataflow processing, i.e., the ALUs form a processing chain so that they can pass data from one to another directly. Sometimes each ALU can have its own control logic and local memory, called a scratchpad or register file. We refer to the ALU with its own local memory as a processing engine (PE). Spatial architectures are commonly used for DNNs in ASIC and FPGA-based designs. In this section, we will discuss the different design strategies for efficient processing on these different platforms, without any impact on accuracy (i.e., all approaches in this section produce bit-wise identical results); specifically, <

> * For temporal architectures such as CPUs and GPUs, we will discuss howcomputational transformson the kernel Fig. 18. Mapping to matrix multiplication for fully connected layers can reduce the number of multiplications to increase throughput. * For spatial architectures used in accelerators, we will discuss howdataflowscan increase data reuse from low andNin Fig. 18(b)); finally, the height of the output feature cost memories in the memory hierarchy toreduce energy map matrix is the number of channels in the output feature consumption. maps (M), and the width is the number of output feature maps (N), where each output feature map of the FC layer has the dimension of 1x1 number of output channels (M). A. Accelerate Kernel Computation on CPU and GPU Platforms The CONV layer in a DNN can also be mapped to a matrix CPUs and GPUs use parallelizaton techniques such as SIMD multiplication using a relaxed form of the Toeplitz matrix as or SIMT to perform the MACs in parallel. All the ALUs share shown in Fig. 19. The downside for using matrix multiplication the same control and memory (register file). On these platforms, for the CONV layers is that there is redundant data in the input both the FC and CONV layers are often mapped to a matrix feature map matrix as highlighted in Fig. 19(a). This can lead multiplication (i.e., the kernel computation). Fig. 18 shows how to either inefficiency in storage, or a complex memory access a matrix multiplication is used for the FC layer. The height of pattern. the filter matrix is the number of filters and the width is the There are software libraries designed for CPUs (e.g., Open- number of weights per filter (input channels (C) width (W) BLAS, Intel MKL, etc.) and GPUs (e.g., cuBLAS, cuDNN, height (H), sinceR=WandS=Hin the FC layer); etc.) that optimize for matrix multiplications. The matrix the height of the input feature maps matrix is the number of multiplication is tiled to the storage hierarchy of these platforms, activations per input feature map <>, and the which are on the order of a few megabytes at the higher levels. <

> Fig. 21. Read and write access per MAC. <

> Fig. 19. Mapping to matrix multiplication for convolutional layers. for a 3x3 filter, respectively, at the cost of reduced numerical stability, increased storage requirements, and specialized The matrix multiplications on these platforms can be further processing depending on the size of the filter. sped up by applying computational transforms to the data to In practice, different algorithms might be used for different reduce the number of multiplications, while still giving the layer shapes and sizes (e.g., FFT for filters greater than 5x5, same bit-wise result. Often this can come at a cost of increased and Winograd for filters 3x3 and below). Existing platform number of additions and a more irregular data access pattern. libraries, such as MKL and cuDNN, dynamically chose the appropriate algorithm for a given shape and size [77, 78].Fast Fourier Transform (FFT) [10,74] is a well known approach, shown in Fig. 20 that reduces the number of multiplications from <> to <> B. Energy-Efficient Dataflow for Accelerators <>, where the output size is <> and the filter size is <>. To For DNNs, the bottleneck for processing is in the memory perform the convolution, we take the FFT of the filter and access. Each MAC requires three memory reads (for filterinput feature map, and then perform the multiplication in weight, fmap activation, and partial sum) and one memorythe frequency domain; we then apply an inverse FFT to the write (for the updated partial sum) as shown in Fig. 21. In theresulting product to recover the output feature map in the worst case, all of the memory accesses have to go through the spatial domain. However, there are several drawbacks to using off-chip DRAM, which will severely impact both throughput FFT: (1) the benefits of FFTs decrease with filter size; (2) the and energy efficiency. For example, in AlexNet, to support itssize of the FFT is dictated by the output feature map size which 724M MACs, nearly 3000M DRAM accesses will be required. is often much larger than the filter; (3) the coefficients in the Furthermore, DRAM accesses require up to several orders offrequency domain are complex. As a result, while FFT reduces magnitude higher energy than computation [79].computation, it requires larger storage capacity and bandwidth. Accelerators, such as spatial architectures as shown inFinally, a popular approach for reducing complexity is to make Fig. 17, provide an opportunity to reduce the energy cost ofthe weights sparse, which will be discussed in SectionVII-B2; data movement by introducing several levels of local memoryusing FFTs makes it difficult for this sparsity to be exploited. hierarchy with different energy cost as shown in Fig. 22. This Several optimizations can be performed on FFT to make it includes a large global buffer with a size of several hundred more effective for DNNs. To reduce the number of operations, kilobytes that connects to DRAM, an inter-PE network that the FFT of the filter can be precomputed and stored. In addition, can pass data directly between the ALUs, and a register file the FFT of the input feature map can be computed once and (RF) within each processing element (PE) with a size of a used to generate multiple channels in the output feature map. few kilobytes or less. The multiple levels of memory hierarchy Finally, since an image contains only real values, its Fourier help to improve energy efficiency by providing low-cost data Transform is symmetric and this can be exploited to reduce accesses. For example, fetching the data from the RF or storage and computation cost. neighbor PEs is going to cost 1 or 2 orders of magnitude Other approaches include Strassen [75] and Winograd [76], lower energy than from DRAM. which rearrange the computation such that the number of Accelerators can be designed to support specialized process- multiplications reduce from <> to <> and by 2.25% ing dataflows that leverage this memory hierarchy. The dataflow 14 <> Fig. 23. Data reuse opportunities in DNNs [80]. <> Fig. 22. Memory hierarchy and data movement energy [80]. decides what data gets read into which level of the memory hierarchy and when are they getting processed. Since there is no randomness in the processing of DNNs, it is possible to design a fixed dataflow that can adapt to the DNN shapes and sizes and optimize for the best energy efficiency. The optimized dataflow minimizes access from the more energy consuming <> levels of the memory hierarchy. Large memories that can store a significant amount of data consume more energy than smaller memories. For instance, DRAM can store gigabytes of data, but consumes two orders of magnitude higher energy per access than a small on-chip memory of a few kilobytes. Thus, every time a piece of data is moved from an expensive level to a Fig. 24. An analogy between the operation of DNN accelerators (texts in lower cost level in terms of energy, we want to reuse that piece black) and that of general-purpose processors (texts in red). Figure adopted from [81]. of data as much as possible to minimize subsequent accesses to the expensive levels. The challenge, however, is that the storage capacity of these low cost memories are limited. Thus program into machine-readable binary codes for executionwe need to explore different dataflows that maximize reuse given the hardware architecture (e.g., x86 or ARM); in theunder these constraints. processing of DNNs, the mapper translates the DNN shapeFor DNNs, we investigate dataflows that exploit three forms and size into a hardware-compatible computation mappingof input data reuse (convolutional, feature map and filter) as for execution given the dataflow. While the compiler usuallyshown in Fig. 23. For convolutional reuse, the same input optimizes for performance, the mapper optimizes for energyfeature map activations and filter weights are used within efficiency.a given channel, just in different combinations for different The following taxonomy (Fig. 25) can be used to classifyweighted sums. For feature map reuse, multiple filters are the DNN dataflows in recent works [82–93] based on their applied to the same feature map, so the input feature map data handling characteristics [80]: activations are used multiple times across filters. Finally, for 1) Weight stationary (WS):The weight stationary dataflow filter reuse, when multiple input feature maps are processed at is designed to minimize the energy consumption of reading once (referred to as a batch), the same filter weights are used weights by maximizing the accesses of weights from the register multiple times across input features maps. file (RF) at the PE (Fig. 25(a)). Each weight is read from If we can harness the three types of data reuse by storing DRAM into the RF of each PE and stays stationary for further the data in the local memory hierarchy and accessing them accesses. The processing runs as many MACs that use the multiple times without going back to the DRAM, it can save same weight as possible while the weight is present in the RF; a significant amount of DRAM accesses. For example, in it maximizes convolutional and filter reuse of weights. The AlexNet, the number of DRAM reads can be reduced by up to inputs and partial sums must move through the spatial array 500in the CONV layers. The local memory can also be used and global buffer. The input fmap activations are broadcast to for partial sum accumulation, so they do not have to reach all PEs and then the partial sums are spatially accumulated DRAM. In the best case, if all data reuse and accumulation across the PE array. can be achieved by the local memory hierarchy, the 3000M One example of previous work that implement weight DRAM accesses in AlexNet can be reduced to only 61M. stationary dataflow is nn-X, or neuFlow [85], which uses The operation of DNN accelerators is analogous to that of eight 2-D convolution engines for processing a 1010 filter. general-purpose processors as illustrated in Fig. 24 [81]. In There are total 100 MAC units, i.e. PEs, per engine with each conventional computer systems, the compiler translates the PE having a weight that stays stationary for processing. The <

> Fig. 26. Variations of output stationary [80].(b) Output Stationary are [89], [88], and [90], respectively. No local reuse (NLR): While small register files are efficient in terms of energy (pJ/bit), they are inefficient in terms Psum <> of area (<>). In order to maximize the storage capacity, and minimize the off-chip memory bandwidth, no local storage Fig. 25. Dataflows for DNNs [80]. is allocated to the PE and instead all that area is allocated to the global buffer to increase its capacity (Fig. 25(c)). The no local reuse dataflow differs from the previous dataflows in input fmap activations are broadcast to all MAC units and the that nothing stays stationary inside the PE array. As a result, partial sums are accumulated across the MAC units. In order to there will be increased traffic on the spatial array and to the accumulate the partial sums correctly, additional delay storage global buffer for all data types. Specifically, it has to multicast elements are required, which are counted into the required size the activations, single-cast the filter weights, and then spatially of local storage. Other weight stationary examples are found accumulate the partial sums across the PE array. in [82–84, 86, 87]. In an example of the no local reuse dataflow from 2) Output stationary (OS):The output stationary dataflow is UCLA [91], the filter weights and input activations are read designed to minimize the energy consumption of reading and from the global buffer, processed by the MAC units with custom writing the partial sums (Fig. 25(b)). It keeps the accumulation adder trees that can complete the accumulation in a single cycle, of partial sums for the same output activation value local in the and the resulting partial sums or output activations are then put RF. In order to keep the accumulation of partial sums stationary back to the global buffer. Another example is DianNao [92], in the RF, one common implementation is to stream the input which also reads input activations and filter weights from activations across the PE array and broadcast the weight to all the buffer, and processes them through the MAC units with PEs in the array. custom adder trees. However, DianNao implements specialized One example that implements the output stationary dataflow registers to keep the partial sums in the PE array, which helps is ShiDianNao [89], where each PE handles the processing for to further reduce the energy consumption of accessing partial each output activation value by fetching the corresponding input sums. Another example of no local reuse dataflow is found activations from neighboring PEs. The PE array implements in [93]. dedicated networks to pass data horizontally and vertically. 4) Row stationary (RS): A row stationary dataflow is Each PE also has data delay registers to keep data around for proposed in [80], which aims to maximize the reuse and the required amount of cycles. At the system level, the global accumulation at the RF level foralltypes of data (weights, buffer streams the input activations and broadcasts the weights pixels, partial sums) for the overall energy efficiency. This into the PE array. The partial sums are accumulated inside differs from WS or OS dataflows, which optimize for only each PE and then get streamed out back to the global buffer. weights and partial sums, respectively. Other examples of output stationary are found in [88, 90]. The row stationary dataflow assigns the processing of a There are multiple possible variants of output stationary as 1-D row convolution into each PE for processing as shown shown in Fig. 26 since the output activations that get processed in Fig. 27. It keeps the row of filter weights stationary inside at the same time can come from different dimensions. For the RF of the PE and then streams the input activations into example, the variantOS A targets the processing of CONV the PE. The PE does the MACs for each sliding window at a layers, and therefore focuses on the processing of output time, which uses just one memory space for the accumulation activations from the same channel at a time in order to of partial sums. Since there are overlaps of input activations maximize data reuse opportunities. The variantOS C targets between different sliding windows, the input activations can the processing of FC layers, and focuses on generating output then be kept in the RF and get reused. By going through all the activations from all different channels, since each channel only sliding windows in the row, it completes the 1-D convolution has one output activation. The variantOS B is something in and maximize the data reuse and local accumulation of data betweenOS A andOS C . Example of variantsOS A ,OS B , and in this row. 16 <

> Fig. 27. 1-D Convolutional reuse within PE for Row Stationary Dataflow [80]. <

> Fig. 29. Multiple rows of different input feature maps, filters and channels are mapped to same PE within array for additional reuse in the Row Stationary <

> Fig. 28. 2-D convolutional reuse within spatial array for Row Stationary shown in Fig. 28. For example, to generate the first row of output activations with a filter having three rows, three 1-D Fig. 30. Mapping optimization takes in hardware and DNNs shape constraints convolutions are required. Therefore, we can use three PEs in to determine optimal energy dataflow [80]. a column, each running one of the three 1-D convolutions. The partial sums are further accumulated vertically across the three PEs to generate the first output row. To generate the second different channels are interleaved, and run through the same PE row of output, we use another column of PEs, where three as a 1-D convolution. The partial sums from different channels rows of input activations are shifted down by one row, and use then naturally get accumulated inside the PE. the same rows of filters to perform the three 1-D convolutions. The number of filters, channels, and fmaps that can be Additional columns of PEs are added until all rows of the processed at the same time is programmable, and there exists an output are completed (i.e., the number of PE columns equals optimal mapping for the best energy efficiency, which depends the number of output rows). on the shape configuration of the DNN as well as the hardware This 2-D array of PEs enables other forms of reuse to reduce resources provided, e.g., the number of PEs and the size of the accesses to the more expensive global buffer. For example, each memory in the hierarchy. Since all of the variables are known filter row is reused across multiple PEs horizontally. Each row before runtime, it is possible to build a compiler (i.e., mapper) of input activations is reused across multiple PEs diagonally. to perform this optimization off-line to configure the hardware And each row of partial sums are further accumulated across for different mappings of the RS dataflow for different DNNs the PEs vertically. Therefore, 2-D convolutional data reuse and as shown in Fig. 30. accumulation are maximized inside the 2-D PE array. One example that implements the row stationary dataflow To address the high-dimensional convolution of the CONV is Eyeriss [94]. It consists of a 14x12 PE array, a 108KB layer (i.e., multiple fmaps, filters, and channels), multiple rows global buffer, ReLU and fmap compression units as shown can be mapped onto the same PE as shown in Fig. 29. The in Fig. 31. The chip communicates with the off-chip DRAM 2-D convolution is mapped to a set of PEs, and the additional using a 64-bit bidirectional data bus to fetch data into the dimensions are handled by interleaving or concatenating the global buffer. The global buffer then streams the data into the additional data. For filter reuse within the PE, different rows PE array for processing. of fmaps are concatenated and run through the same PE In order to support the RS dataflow, two problems need to be as a 1-D convolution. For input fmap reuse within the PE, solved in the hardware design. First, how can the fixed-size PE different filter rows are interleaved and run through the same array accommodate different layer shapes? Second, although PE as a 1-D convolution. Finally, to increase local partial sum the data will be passed in a very specific pattern, it still changes accumulation within the PE, filter rows and fmap rows from with different shape configurations. How can the fixed design needs of each dataflow under the same area constraint. For example, since the no local reuse dataflow does not require any Processing RF in PE, it is allocated with a much larger global buffer. The If map simulation uses the layer configurations from AlexNet with a Buffer batch size of 16. The simulation also takes into account the bits fact that accessing different levels of the memory hierarchy Enc. requires different energy cost. of each dataflow for the CONV layers of AlexNet with a batch size of 16. The WS and OS dataflows have the lowest energy consumption for accessing weights and partial sums, respectively. However, the RS dataflow has the lowest total 13 energy consumption since it optimizes for the overall energy efficiency instead of only for a certain data type. Fig. 33(a) shows the same results with breakdown in terms of memory hierarchy. The RS dataflow consumes the most energy in the RF, since by design most of the accesses have been moved to the lowest level of the memory hierarchy. This helps to achieve the lowest total energy consumption since RF has the lowest energy per access. The NLR dataflow has the lowest Clock Gated energy consumption at the DRAM level, since it has a much larger global buffer and thus higher on-chip storage capacity compared to others. However, most of the data accesses in relatively large energy consumption per access compared to accessing data from RF or inside the PE array. As a result, the overall energy consumption of the NLR dataflow is still fairly high. Overall, RS dataflow uses 1.4% to 2.5% lower energy pass data in different patterns? <

> Fig. 32. Mapping uses replication and folding to maximized utilization of the NLR dataflow is from the global buffer, which still has a PE array [94]. Two mapping strategies can be used to solve the first problem than other dataflows. as shown in Fig. 32. First, replication can be used to map shapes Fig. 34 shows the energy efficiency between different that do not use up the entire PE array. For example, in the dataflows in the FC layers of AlexNet with a batch size of 16. third to fifth layers of AlexNet, each 2-D convolution only uses Since there is not as much data reuse in the FC layers as in a 13x3 PE array. This structure is then replicated four times, the CONV layers, all dataflows spend a significant amount of and runs different channels and filters in each replication. The energy on reading weights. However, RS dataflow still has the second strategy is called folding. For example, in the second lowest energy consumption because it optimizes for the energy layer of AlexNet, it requires a 27x5 PE array to complete the of accessing input activations and partial sums. For the OS2-D convolution. In order to fit it into the 14x12 physical PE dataflows,OSarray, it is folded into two parts, 14x5 and 13x5, and each C now consumes lower energy thanOS A since it is designed for the FC layers. Overall, RS still consumesare vertically mapped into the physical PE array. Since not all 1.3% lower energy compared to other dataflows at the batchPEs are used by the mapping, the unused PEs can be clock size of 16.gated to save energy consumption. A custom multicast network is used to solve the second Fig. 35 shows the RS dataflow design with energy breakdown problem about flexible data delivery. The simplest way to pass in terms of different layers of AlexNet. In the CONV layers, the data to multiple destinations is to broadcast the data to all PEs energy is mostly consumed by the RF, while in the FC layers, and let each PE decide if it has to process the data or not. the energy is mostly consumed by DRAM. However, most However, it is not very energy efficient especially when the of the energy is consumed by the CONV layers, which takes size of PE array is large. Instead, a multicast network is used around 80% of the energy. As recent DNN models go deeper to send data to only the places where it is needed. with more CONV layers, the ratio between number of CONV 5) Energy comparison of different dataflows:To evaluate and FC layers only gets larger. Therefore, moving forward, and compare different dataflows, the same total hardware area significant effort should be placed on energy optimizations for and number of PEs (256) are used in the simulation of a spatial CONV layers. architecture for all dataflows. The local memory (register file) at Finally, up until now, we have been looking at architec- each processing element (PE) is on the order of 0.5 – 1.0kB and tures with relatively limited storage on the order of a few a shared memory (global buffer) is on the order of 100 – 500kB. hundred kilobytes. With much larger storage on the order of The sizes of these memories are selected to be comparable to a few megabytes, additional dataflows can be considered. For a typical accelerator for multimedia processing, such as video example, Fused-Layer looks at dataflow optimizations across coding [95]. The memory sizes are further adjusted for the layers [96]. 18 <> Fig. 35. Energy breakdown across layers of the AlexNet [80]. RF energy dominates in convolutional layers. DRAM energy dominates in the fully connected layer. Convolutional layer dominate energy consumption. In this section, we will discuss how moving compute and data Normalized closer to reduce data movement (i.e., near-data processing) can pixels be achieved using mixed-signal circuit design and advanced memory technologies. Many of these works use analog processing which has the drawback of increased sensitivity to circuit and device non- idealities. Consequentially, the computation is often performed at reduced precision, which can be accounted for during (b) Energy breakdown across data type the training of the DNNs using the techniques discussed in Section VII. Another factor to take into consideration is that Fig. 33. Comparison of energy efficiency between different dataflows in the DNNs are often trained in the digital domain; thus for analog CONV layers of AlexNet with a batch size of 16 [3]: (a) breakdown in terms of storage levels and ALU, (b) breakdown in terms of data types. OS processing, there is an additional overhead cost for analog- A , OS B and OS C are three variants of the OS dataflow that are commonly seen in to-digital conversion (ADC) and digital-to-analog conversion different implementations [80]. (DAC). A. DRAM Advanced memory technology can reduce the access energy for high density memories such as DRAMs. For instance, psums embedded DRAM (eDRAM)brings high density memory on- chip to avoid the high energy cost of switching off-chip pixels capacitance [97]; eDRAM is 2.85higher density than SRAM 0.5 and 32% more energy efficient than DRAM (DDR3) [93]. eDRAM also offers higher bandwidth and lower latency compared to DRAM. In DNN processing, eDRAM can be used DNN Dataflows to store tens of megabytes of weights and activations on-chip to avoid off-chip access, as demonstrated in DaDianNao [93]. off-chip DRAM and can increase the cost of the chip. Rather than integrating DRAM into the chip itself, the DRAM can also be stacked on top of the chip using throughVI. N EAR -D ATA PROCESSING silicon vias (TSV). This technology is often referred to as3-D The previous section highlighted that data movement domi- memory, and has been commercialized in the form of Hybrid nates energy consumption. While spatial architectures distribute Memory Cube (HMC) [98] and High Bandwidth Memory the on-chip memory such that it is closer to the computation (HBM) [99]. 3-D memory delivers an order of magnitude higher (e.g., into the PE), there have also been efforts to bring the bandwidth and reduces access energy by up to 5relative to off-chip high density memory closer to the computation or to existing 2-D DRAMs, as TSV have lower capacitance than integrate the computation into the memory itself; the latter is typical off-chip interconnects. Recent works have explored the often referred to asprocessing-in-memoryorlogic-in-memory. use of HMC for efficient DNN processing in a variety of ways. In embedded systems, there have also been efforts to bring the For instance, Neurocube [100] integrates SIMD processors into computation into the sensor where the data is first collected. the logic die of the HMC to bring the memory and computation 19 voltage as the input, and the current as the output as shown in resistive memory. <

> Fig. 36. Analog computation by (a) SRAM bit-cell and (b) non-volatile Processing with non-volatile resistive memories has several drawbacks as described in [108]. First, it suffers from the reduced precision and ADC/DAC overhead of analog process- ing described earlier. Second, the array size is limited by thecloser together. Tetris [101] explores the use of HMC with wires that connect the resistive devices; specifically, wire energythe Eyeriss spatial architecture and row stationary dataflow. dominates for large arrays (e.g., 1k1k), and the IR drop alongIt proposes allocating more area to computation than on-chip wire can degrade the read accuracy. Third, the write energymemory (i.e., larger PE array and smaller global buffer) in to program the resistive devices can be costly, in some casesorder to exploit the low energy and high throughput properties requiring multiple pulses. Finally, the resistive devices can alsoof the HMC. It also adapts the dataflow to account for the suffer from device-to-device and cycle-to-cycle variations withHMC memory and smaller on-chip memory. Tetris achieves non-linear conductance across the conductance range.a 1.5reduction in energy consumption and 4.1increase There have been several recent works that explore the use ofin throughput over a baseline system with conventional 2-D memristors for DNNs. ISAAC [104] replaces the eDRAM inDRAM. DaDianNao with memristors. To address the limited precision support, ISAAC computes a 16-bit dot product operation with B. SRAM 8 memristors each storing 2-bits; a 1-bit2-bit multiplication Rather than bringing the memory near the compute, recent is performed at each memristor, where a 16-bit input requires work has also investigated bringing the compute into the 16 cycles to complete. In other words, the ISAAC architecture memory. For instance, the multiply and accumulate operation trades off area and time for increased precision. Finally, ISAAC can be directly integrated into the bit-cells of an SRAM arranges its 25.1M memristors in a hierarchical structure to array [102], as shown in Fig. 36(a). In this work, a 5-bit avoid issues with large arrays. PRIME [109] also replaces the DAC is used to drive the word line (WL) to an analog voltage DRAM main memory with memristors; specifically, it uses that represents the feature vector, while the bit-cells store the 256256 memristor arrays that can be configured for 4-bit binary weights1. The bit-cell current (I multi-level cell computation or 1-bit single level cell storage. BC ) is effectively a product of the value of the feature vector and the value of It should be noted that results from ISAAC and PRIME are the weight stored in the bit-cell; the currents from the bit- obtained from simulations. The task of actually fabricating cells within a column add together to discharge the bitline large memristors arrays is still very much a research challenge; (V for instance, [110] uses a fabricated 1212 memristor array BL ). This approach gives 12energy savings compared to reading the 1-bit weights from the SRAM and performing the to demonstrate a linear classifier. computation separately. To counter circuit non-idealities, the DAC accounts for the non-linear bit-line discharge with respect D. Sensors to the WL voltage, and boosting is used to combine the weak In certain applications, such as image processing, the dataclassifiers that are susceptible to device variations to form a movement from the sensor itself can account for a significantstrong classifier [103]. portion of the system energy consumption. Thus there has also been research on performing the computation as close C. Non-volatile Resistive Memories as possible to the sensor. In particular, much of the work focuses on moving the computation into the analog domain toThe multiply and accumulate operation can also be directly avoid using the ADC within the sensor, which accounts for aintegrated into advancednon-volatilehigh density memories significant portion of the sensor power. However, as mentionedby using them as programmable resistive elements, commonly referred to asmemristors[105]. Specifically, a multiplication 8 The resistive devices can be inserted between the cross-point of two wires is performed with the resistor’s conductance as the weight, the and in certain cases can avoid the need for an access transistor. 20 earlier, lower precision is required for analog computation due to circuit non-idealities. In [111], the matrix multiplication is integrated into the ADC, where the most significant bits of the multiplications are performed using switched capacitors in an 8-bit successive approximation format. This is extended in [112] to not only perform the multiplications, but also the accumulations in the analog domain. In this work, it is assumed that 3-bits and 6-bits are sufficient to represent the weights and activations, respectively. This reduces the number of ADC conversions in the sensor by 21. RedEye [113] takes this approach even further by performing the entire convolution layer (including convolution, max pooling and quantization) in the analog domain at the sensor. It should be noted that [111] and [112] report measured results from fabricated test chips, while results in [113] are from simulations. <

> It is also feasible to embed the computation not just before the ADC, but into the sensor itself. For instance, in [114] an Fig. 37. Various methods of quantization (Figures from [117, 118]). Angle Sensitive Pixels sensor is used to compute the gradient of the input, which along with compression, reduces the data the number of bits. The benefits of reduced precision includemovement from the sensor by 10. In addition, since the reduced storage cost and/or reduced computation requirements.first layer of the DNN often outputs a gradient-like feature map, it maybe possible to skip the computations in the first There are several ways to map the data to quantization levels. layer, which further reduces energy consumption as discussed The simplest method is a linear mapping with uniform distance in [115, 116]. between each quantization level (Fig. 37(a)). Another approach is to use a simple mapping function such as alog function (Fig. 37(b)) where the distance between the levels varies; thisVII. C O -DESIGN OF DNN MODELS AND HARDWARE mapping can often be implemented with simple logic such as aIn earlier work, the DNN models were designed to maximize shift. Alternatively, a more complex mapping function can beaccuracy without much consideration of the implementation used where the quantization levels are determined or learnedcomplexity. However, this can lead to designs that are chal- from the data (Fig. 37(c)), e.g., using k-means clustering; forlenging to implement and deploy. To address this, recent this approach, the mapping is usually implemented with a lookwork has shown that DNN models and hardware can be co- up table.designed to jointly maximize accuracy and throughput, while Finally, the quantization can be fixed (i.e., the same methodminimizing energy and cost, which increases the likelihood of of quantization is used for all data types and layers, filters, andadoption. In this section, we will highlight various efforts that channels in the network); or it can be variable (i.e., differenthave been made towards the co-design of DNN models and methods of quantization can be used for weights and activations,hardware. Note that unlike Section V, the techniques discussed and different layers, filters, and channels in the network).in this section can affect the accuracy; thus, the goal is to Reduced precision research initially focused on reducingnot only substantially reduce energy consumption and increase the precision of the weights rather than the activations, sincethroughput, but also to minimize any degradation in accuracy. weights directly increase the storage capacity requirement,The co-design approaches can be loosely grouped into the while the impact of activations on storage capacity depends onfollowing categories: the network architecture and dataflow. However, more recent Reduce precision of operations and operands.This in- works have also started to look at the impact of quantizationcludes going from floating point to fixed point, reducing on activations. Most reduced precision research also focusesthe bitwidth, non-linear quantization and weight sharing. on reducing the precision for inference rather than training Reduce number of operations and model size. This (with some exceptions [88,119,120]) due to the sensitivity ofincludes techniques such as compression, pruning and the gradients to quantization.compact network architectures. The key techniques used in recent work to reduce precision are summarized in Table III; both linear and non-linear A. Reduce Precision quantization applied to weights and activations are explored. Quantization involves mapping data to a smaller set of The impact on accuracy is reported relative to a baseline quantization levels. The ultimate goal is to minimize the error precision of 32-bit floating point, which is the default precision between the reconstructed data from the quantization levels and used on platforms such as GPUs and CPUs. the original data. The number of quantization levels reflects the 1) Linear quantization:The first step of reducing precision precisionand ultimately the number of bits required to represent is usually to convert values and operations from floating point the data (usuallylog 2 of the number of levels); thus,reduced to fixed point. A 32-bit floating point number, as shown in precisionrefers to reducing the number of levels, and thus Fig. 38(a), is represented by <>, wheres product; that output would need to be accumulated with <> bit precision, where M is determined based on the largest filter (b) 8-bit dynamic fixed point examples size <> (<> from Fig. 9(b)), which is in the range of 0 to 16 bits for the popular DNNs described in SectionIII-B. Fig. 38. Various methods of number representations. 1 After accumulation, the precision of the final output activation is typically reduced to N-bits [88,121], as shown in Fig. 39.is the sign bit, e is the 8-bit exponent, andmis the 23-bit The reduced output precision does not have a significant impact mantisa, and covers the range of <>. on accuracy if the distribution of the weights and activationsAn N-bit fixed point number is represented by <> are centered near zero such that the accumulation would not 2f , wheresis the sign bit,mis the (N-1)-bit mantissa, and move only in one direction; this is particularly true when batchfdetermines the location of the decimal point and acts as a normalization is used. scale factor. For instance, for an 8-bit integer, whenf= 0, The reduced precision is not only explored in research,the dynamic range is -128 to 127, whereas whenf= 10, the but has been used in recent commercial platforms for DNN dynamic range is -0.125 to 0.124023438.Dynamicfixed point processing. For instance, Google’s Tensor Processing Unitrepresentation allowsfto vary based on the desired dynamic (TPU) which was announced in May 2016, was designed forrange as shown in Fig. 38(b). This is useful for DNNs, since 8-bit integer arithmetic [123]. Similarly, Nvidia’s PASCAL the dynamic range of the weights and activations can be quite GPU, which was announced in April 2016, also has 8-bitdifferent. In addition, the dynamic range can also vary across \integer instructions for deep learning inference [124]. In generallayers and layer types (e.g., convolutional vs. fully connected). purpose platforms such as CPUs and GPUs, the main benefit Using dynamic fixed point, the bitwidth can be reduced to 8 of using 8-bit computation is an increase in throughput, asbits for the weights and 10 bits for the activations without any four 8-bit operations rather than one 32-bit operation can befine-tuning of the weights [121]; with fine-tuning, both weights performed for a given clock cycle.and activations can reach 8-bits [122]. While general purpose platforms usually support 8-bit,Using 8-bit fixed point has the following impact on energy 16-bit and/or 32-bit operations, it has been shown that theand area [79]: minimum bit precision for DNNs can actually vary in a more An 8-bit fixed point add consumes 3.3% less energy fine grained manner. For instance, the weight and activation (3.8less area) than a 32-bit fixed point add, and 30% precision can vary between 4 and 9 bits for AlexNet across less energy (116less area) than a 32-bit floating point different layers without significant impact on accuracy (i.e., a add. The energy and area of a fixed-point add scales change of less than 1%) [125,126]. This fine-grained variation approximately linearly with the number of bits. can be exploited for increased throughput or reduced energy An 8-bit fixed point multiply consumes 15.5% less energy consumption with specialized hardware. For instance, if bit- (12.4% less area) than a 32-bit fixed point multiply, serial processing is used, where the number of clock cycles to and 18.5% less energy (27.5% less area) than a 32-bit complete an operation is proportional to the bitwidth, adapting floating point multiply. The energy and area of a fixed- to fine-grain variations in bit precision can result in a 2.24% point multiply scales approximately quadratically with the speed up versus 16-bits [125]. Alternatively, a multiplier can number of bits. be designed such that its critical path reduces based on the bit Reducing the precision also reduces the energy and area cost precision as fewer adders are needed to resolve the product; for storage, which is important since memory access and data this can be combined with voltage scaling for a 2.56energy movement dominate energy consumption as described earlier. savings versus 16-bits [126]. While these bit scaling results The energy and area of the memory scale approximately linearly are reported relative to 16-bit, it would be interesting to see with number of bits. It should be noted, however, that changing their impact relative to the maximum precision required across from floating point to fixed point, without reducing bit-width, layers (i.e., 9-bits for [125, 126]). does not reduce the energy or area cost of the memory. The precision can be reduced even more aggressively to a For completeness, it should be noted that the precision of single bit; this area of research is often referred to asbinary nets. the internal values of a fixed-point multiply and accumulate BinaryConnect (BC) [127] introduced the concept of binary (MAC) operation are typically higher than the weights and weights (i.e., -1 and 1), where using a binary weight reduced activations. To guarantee no precision loss, weights and input the multiplication in the MAC to addition and subtraction activations with N-bit fixed-point precision would require an only. This was later extended in Binarized Neural Networks N-bitxN-bit multiplication which generates a 2N-bit output (BNN) [128] that uses binary weightsandactivations, which <

> Fig. 40. Weight sharing hardware. w, where w is the average of the absolute values of the weights in the filter) 9 , keeping the first and last layers at 32-bit floating point precision, and performing normalization before VGG-16 [117]. Furthermore, when weights are quantized to convolution to reduce the dynamic range of the activations. powers of two, the multiplication can be replaced with a bit- With these changes, BWN reduced the accuracy loss to 0.8%, shift [122,135]. 10 Incremental Network Quantization (INQ) while XNOR-Nets reduced the loss to 11%. The loss of XNOR- can be used to further reduce the loss in accuracy by dividing Net can be further reduced by increasing the precision of the the large and small weights into different groups, and then activations to be slightly larger than one bit. For instance, iteratively quantizing and re-training the weights [136]. Quantized Neural Networks (QNN) [119], DoReFa-Net [120], Weight Sharingforces several weights to share a single value. and HWGQ-Net [130] allow the activations to have 2-bits, This reduces the number of unique weights in a filter or a while the weights remain at 1-bit; in HWGQ-Net, this reduces layer. One example is to group the weights by using a hashing the accuracy loss to 5.2%. function and use one value for each group [137]. Alternatively, All the previously described binary nets limit the weights the weights can be grouped by the k-means algorithm [118]. to two values (-wandw); however, there may be benefits Both the shared weights and the indexes indicating which for allowing weights to be zero (i.e., -w, 0,w). Although weight to use at each position of the filter are stored. This this requires an additional bit per weight compared to binary leads to a two step process to fetch the weight: (1) read the weights, the sparsity of the weights can be exploited to reduce weight index; (2) using the weight index, read the shared computation and storage cost, which can potentially cancel weights. This approach can reduce the cost of reading and out the cost of the additional bit. This is explored in Ternary storing the weights if the weight index (log 2 of the number of Weight Nets (TWN) [131] and then extended in Trained Ternary unique weights) is less than the bitwidth of the weight itself. Quantization (TTQ) where a different scale is trained for each For instance, in Deep Compression [118], the number of weight (i.e., -w unique weights per layer is reduced to 256 for convolutional 1 , 0,w2 ) for an accuracy loss of 0.6% [132], assuming 32-bit floating point for the activations. layers and 16 for fully-connected layers in AlexNet, requiring Hardware implementations for binary/ternary nets have 8-bit and 4-bit weight indexes, respectively. Assuming there been explored in recent publications. YodaNN [133] uses areUunique weights and the size of the filters in the layer binary weights, while BRein [134] uses binary weights and is <> from Fig. 9(b), there will be energy savings activations. Binary weights are also used in the compute if reading from a CRSM <> U-bit memory plus aU16- in SRAM work [102] described in Section VI. Finally, the bit memory (as shown in Fig. 40) cost less than reading nominally spike-inspired TrueNorth chip can implement a from a CRSM 16-bit memory. Note that unlike the previous reduced precision neural network with binary activations and quantization methods, the weight sharing approach does not ternary weights using TrueNorth’s quantized weight table [9]. reduce the precision of the MAC computation itself and only These works tend not to support state-of-the-art DNN models reduces the weight storage requirement. (with the exception of YodaNN). 2) Non-linear quantization:The previous works described B. Reduce Number of Operations and Model Size involve linear quantization where the levels are uniformly In addition to reducing the size of each operation or operandspaced out. It has been shown that the distributions of the (weight/activation), there is also a significant amount of researchweights and activations are not uniform [118,135], and thus on methods to reduce the number of operations and modela non-linear quantization can potentially improve accuracy. size. These techniques can be loosely classified as exploitingSpecifically, there have been two popular approaches taken activation statistics, network pruning, network architecturein recent works: (1) log domain quantization; (2) learned design and knowledge distillation.quantization or weight sharing. 1) Exploiting Activation Statistics: As discussed in Sec-Log domain quantizationIf the quantization levels are tionIII-A1, ReLU is a popular form of non-linearity used inassigned based on a logarithmic distribution as shown in DNNs that sets all negative values to zero as shown in Fig. 41(a). Fig 37(b), the weights and activations are more equally As a result, the output activations of the feature maps after the distributed across the different levels and each level is used ReLU are sparse; for instance, the feature maps in AlexNetmore efficiently resulting in less quantization error. For instance, have sparsity between 19% to 63% as shown in Fig. 41(b).using 4 bits in linear quantization results in a 27.8% loss in This sparsity gives ReLU an implementation advantage overaccuracy versus a 5% loss for log base-2 quantization for other non-linearities such as sigmoid, etc. 9 This can also be thought of as a form of weights sharing, where only two 10 Note however that multiplications do not account for a significant portion weights are used per filter. of the total energy. <> TABLE III METHODS TO REDUCE NUMERICAL PRECISION FOR ALEX NET . ACCURACY MEASURED FOR TOP-5 ERROR ON IMAGE NET . a cost of reduced accuracy. 2) Network Pruning:To make network training easier, the networks are usually over-parameterized. Therefore, a large amount of the weights in a network are redundant and can be removed (i.e., set to zero). This process is called network pruning. Aggressive network pruning often requires some fine- tuning of the weights to maintain the original accuracy. This was first proposed in 1989 through a technique called Optimal Brain Damage [140]. The idea was to compute the impact of each weight on the training loss (discussed in SectionII-C), referred to as the weight saliency. The low-saliency weights (Normalized) were removed and the remaining weights were fine-tuned; this process was repeated until the desired weight reduction and accuracy were reached. In 2015, a similar idea was applied to modern DNNs in [141]. <> Rather than using the saliency as a metric, which is too difficult to compute for the large-scaled DNNs, the pruning was simply Fig. 41. Sparsity in activations due to ReLU. based on the magnitude of the weights. Small weights were pruned and the model was fine-tuned to restore the accuracy. Without fine-tuning the weights, about 50% of the weightsThe sparsity can be exploited for energy and area savings could be pruned. With fine-tuning, over 80% of the weightsusing compression, particularly for off-chip DRAM access were pruned. Overall this approach can reduce the numberwhich is expensive. For instance, a simple run length coding of weights in AlexNet by 9and the number of MACsthat involves signaling non-zero values of 16-bits and then runs by 3. Most of the weight reduction comes from the fully-of zeros up to 31 can reduce the external memory bandwidth connected layers (9.9for fully-connected layers versus 2.7of the activations by 2.1and the overall external bandwidth for convolutional layers).(including weights) by 1.5[61]. 11 In addition to compression, the hardware can also be modified such that it skips reading the However, the number of weights alone is not a good metric weights and performing the MAC for zero-valued activations for energy. For instance, in AlexNet, the number of weights to reduce energy cost by 45% [94]. Rather than just gating the in the fully-connected layers is much larger than in the read and MAC computation, the hardware could also skip the convolutional layers; however, the energy of the convolutional cycle to increase the throughput by 1.37%[138]. layers is much higher than the fully-connected layers as shown The activations can be made to be even more sparse by prun- in Fig. 35 [80]. Rather than using the number of weights ing the low-valued activations. For instance, if all activations and MAC operations as proxies for energy, the pruning of with small values are pruned, this can be translated into an the weights can be directly driven by energy itself [142]. An additional 11% speed up [138] or 2power reduction [139] energy evaluation method can be used to estimate the DNN with little impact on accuracy. Aggressively pruning more energy that accounts for the data movement from different activations can provide additional throughput improvement at levels of the memory hierarchy, the number of MACs, and the data sparsity as shown in Fig. 42; this energy estimation tool is available at [143]. The resulting energy values for popular This simple run length compression is within 5-10% of the theoretical entropy limit. DNN models are shown in Fig. 43(a). Energy-aware pruning 24 <

> Fig. 42. Energy estimation methodology from [142], which estimates the energy based on data movement from different levels of the memory hierarchy, <

> Fig. 43. Energy values estimated with methodology in [142]. a time [144]. The CSC format will provide an overall lower memory bandwidth than CSR if the output is smaller than the input, or in the case of DNN, if the number of filters isnot can then be used to prune weights based on energy to reduce significantly larger than the number of weights in the filter the overall energy across all layers by 3.7% for AlexNet, which (<> from Fig. 9(b)). Since this is often true, CSC can is 1.74more efficient than magnitude-based approaches [141] be an effective format for sparse DNN processing. as shown in Fig. 43(b). As mentioned previously, it is well Custom hardware has been explored to efficiently supportknown that AlexNet is over-parameterized. The energy-aware pruned DNN models. Many works aim to perform the process-pruning can also be applied to GoogleNet, which is already a ing without decompressing the weights or activations. EIE [145]small DNN model, for a 1.6energy reduction. performs the sparse matrix-vector multiplication specifically for Recent works have examine how to efficiently support the fully connected layers. It stores the weights in a CSC format processing of sparse weights in hardware. One area of interest along with the start location of each column, which needs to be is how to best store the sparse weights after pruning. Similar to stored since the compressed weights have variable length. When compressing the sparse activations discussed in SectionVII-B1, the input is not zero, the compressed weight column is read and the sparse weights can be compressed to reduce memory access the output is updated. To handle the sparsity, additional logic bandwidth by 20 to 30% [118]. is used to keep track of the location of the output that should When DNN processing is performed as a matrix-vector be updated. SCNN [146] supports processing of convolutional 25 layers in a compressed format. It uses an input stationary weights [154]. It proposes afiremodule that first ‘squeezes’ dataflow to deliver the compressed weights and activations to the network with 1x1 convolution filters and then expands a multiplier array followed by a scatter network to add the it with multiple 1x1 and 3x3 convolution filters. It achieves scattered partial sums. an overall 50% reduction in number of weights compared to Recent works have also explored the use of structured AlexNet, while maintaining the same accuracy. It should be pruning to avoid the need for custom hardware [147,148]. noted, however, that reducing the number of weights does not Rather than pruning individual weights (also referred to as fine- necessarily reduce energy; for instance, SqueezeNet consumes grained pruning), structured pruning involves pruning groups more energy than AlexNet, as shown in Fig. 43(a). of weights (also referred to as coarse-grained pruning). The b) After Training:Tensor decomposition can be used to benefits of structured pruning are (1) the resulting weights can decompose filters in a trained network without impacting the better align with the data-parallel architecture (e.g., SIMD) accuracy. It treats weights in a layer as a 4-D tensor and breaks found in existing general purpose hardware, which results in it into a combination of smaller tensors (i.e., several layers). more efficient processing [149]; (2) it amortizes the overhead Low-rank approximation can then be applied to further increase cost required to signal the location of the non-zero weights the compression rate at the cost of accuracy degradation, which across a group of weights, which improves compression and can be restored by fine-tuning the weights. thus reduces storage cost. These groups of weights can include This approach is demonstrated using Canonical Polyadic (CP) a pair of neighboring weights, an entire row or column of a decomposition, a high-order extension of singular value decom- filter, an entire channel of a filter or the entire filter itself; using position that can be solved by various methods, such as a greedy larger groups tends to result in higher loss in accuracy [150]. algorithm [155] or a non-linear least-square method [156]. 3) Compact Network Architectures:The number of weights Combining CP-decomposition with low-rank approximation and operations can also be reduced by improving the network achieves a 4.5% speed-up on CPUs [156]. However, CP- architecture itself. The trend is to replace a large filter with a decomposition cannot be computed in a numerically stable series of smaller filters, which have fewer weights in total; when way when the dimension of the tensor, which represents the the filters are applied sequentially, they achieve the same overall weights, is larger than two [156]. To alleviate this problem, effective receptive field (i.e., the region the filter uses from input Tucker decomposition is adopted instead in [157]. image to compute an output). This approach can be applied 4) Knowledge Distillation:Using a deep network or av- during the network architecture design (before training) or by eraging the predictions of different models (i.e., ensemble) decomposing the filters of a trained network (after training). gives a better accuracy than using a single shallower network. The latter one avoids the hassle of training networks from However, the computational complexity is also higher. To get scratch. However, it is less flexible than the former one. For the best of both worlds, knowledge distillation transfers the example, existing methods can only decompose a filter in a knowledge learned by the complex model (teacher) to the trained network into a series of filters without non-linearity simpler model (student). The student network can therefore between them. achieve an accuracy that would be unachievable if it was a) Before Training:In recent DNN models, filters with directly trained with the same dataset [158,159]. For example, a smaller width and height are used more frequently because [160] shows how using knowledge distillation can improve the concatenating several of them can emulate a larger filter as speech recognition accuracy of a student net by 2%, which is shown in Fig. 13. For example, one 5x5 convolution can be similar to the accuracy of a teacher net that is composed of replaced with two 3x3 convolutions. Alternatively, one NxN an ensemble of 10 networks. convolution can be decomposed into two 1-D convolutions, one Fig. 45 shows the simplest knowledge distillation 1xN and one Nx1 convolution [53]; this basically imposes method [158]. The softmax layer is commonly used as the a restriction that the 2-D filter must be separable, which is output layer in the image classification networks to generate a common constraint in image processing [151]. Similarly, a the class probabilities from the class scores 12 ; it squashes the 3-D convolution can be replaced by a set of 2-D convolutions class scores into values between 0 and 1 that sum up to 1. (i.e., applied only on one of the input channels) followed by For this knowledge distillation method, soft targets (values 1x1 3-D convolutions as demonstrated in Xception [152] and between 0 and 1) such as the class scores of the teacher DNN MobileNets [153]. The order of the 2-D convolutions and 1x1 (or an ensemble of teacher DNNs) are used instead of the 3-D convolutions can be switched. hard targets (values of either 0 or 1) such as the labels in the 1x1 convolutional layers can also be used to reduce the dataset. The objective is to minimize the squared difference number of channels in the output feature map for a given between the soft targets and the class scores of the student DNN. layer, which reduces the number of filter channels and thus Class scores are used as the soft targets instead of the class computation cost for the filters in the next layer as demonstrated probabilities because small values in the class scores contain in [15,51,52]; this is often referred to as a ‘bottleneck’ as important information that may be eliminated by the softmax. discussed in SectionIII-B. For this purpose, the number of 1x1 Alternatively, class probabilities after the softmax layer can be filters has to be less than the number of channels in the 1x1 used as soft targets if the softmax is configured to generate filter. For example, 32 filters of 1x164 can transform an input softer class probabilities where the smaller values retain more with 64 channels to an output of 32 channels and reduce the information [160]. Finally, the intermediate representations of number of filter channels in the next layer to 32. SqueezeNet uses many 1x1 filters to aggressively reduce the number of 12 Also commonly referred to as logits. robotics. For data analytics, high throughput means that more data can be analyzed in a given amount of time. As the amount of visual data is growing exponentially, high-throughput big data analytics becomes important, particularly if an action needs to be taken based on the analysis (e.g., security or terrorist prevention; medical diagnosis). Try to match Low latencyis necessary for real-time interactive applications. Latency measures the time between when the pixel arrives to a system and when the result is generated. Latency is Simple DNN measured in terms of seconds, while throughput is measured in operations/second. Often high throughput is obtained by batching multiple images/frames together for processing; this Fig. 45. Knowledge distillation matches the class scores of a small DNN to results in multiple frame latency (e.g., at 30 frames per second, an ensemble of large DNNs. a batch of 100 frames results in a 3 second delay). This delay is not acceptable for real-time applications, such as high-speed navigation where it would reduce the time available for coursethe teacher DNN can also be incorporated as the extra hints correction. Thus achieving low latency and high throughputto train the student DNN [161]. Hardware costis in large part dictated by the amount of on-chip storage and the number of cores. Typical embedded processors have limited on-chip storage on the order of a few simultaneously can be a challenge. VIII. B ENCHMARKING METRICS FOR DNN EVALUATION AND COMPARISON As we have seen in this article, there has been a significant hundred kilobytes. Since there is a trade-off between the amount amount of research on efficient processing of DNNs. We should of on-chip memory and the external memory bandwidth, both consider several key metrics to compare the various strengths metrics should be reported. Similarly, there is a correlation and weaknesses of different designs and proposed techniques. between the number of cores and the throughput. In addition, These metrics should cover important attributes such as accu- while many cores can be built on a chip, the number of cores racy/robustness, power/energy consumption, throughput/latency that can actually be used at a given time should be reported. It is and cost. Reporting all these metrics is important in order often unrealistic to assume peak utilization and performance due to provide a complete picture of the trade-offs made by a to limitations of mapping and memory bandwidth. Accordingly, proposed design or technique. We have prepared a website to the power and throughput should be reported for running actual collect these metrics from various publications [162]. DNNs as opposed to only reporting theoretical limits. In terms ofaccuracyandrobustness, it is important that the accuracy be reported on widely-accepted datasets as discussed in Section IV. The difficulty of the dataset and/or task should A. Metrics for DNN Models be considered when measuring the accuracy. For instance, the To evaluate the properties of a given DNN model, we should MNIST dataset for digit recognition is significantly easier than consider the following metrics:the ImageNet dataset. As a result, a DNN that performs well on MNIST may not necessarily perform well on ImageNet. Theaccuracy of the model in terms of the top-5 error Thus it is important that the same dataset and task is used when on datasets such as ImageNet. Also, the type of data comparing the accuracy of different DNN models; currently augmentation used (e.g., multiple crops, ensemble models) ImageNet is preferred since it presents a challenge for DNNs, should be reported. as opposed to MNIST, which can also be addressed with simple Thenetwork architectureof the model should be reported, non-DNN techniques. To demonstrate primarily hardware including number of layers, filter sizes, number of filters innovations, it would be desirable to report results for widely- and number of channels. used DNN models (e.g., AlexNet, GoogLeNet) whose accuracy Thenumber of weightsimpact the storage requirement of and robustness have been well studied and tested. the model and should be reported. If possible, the number Energyandpowerare important when processing DNNs at of non-zero weights should be reported since this reflects the edge in embedded devices with limited battery capacity the theoretical minimum storage requirements. (e.g., smart phones, smart sensors, UAVs, and wearables), or in Thenumber of MACsthat needs to be performed should the cloud in data centers with stringent power ceilings due to be reported as it is somewhat indicative of the number cooling costs, respectively. Edge processing is preferred over of operations and potential throughput of the given DNN. the cloud for certain applications due to latency, privacy or If possible, the number of non-zero MACs should also communication bandwidth limitations. When evaluating the be reported since this reflects the theoretical minimum power and energy consumption, it is important to account compute requirements. for all aspects of the system including the chip and external Table IV shows how these metrics are reported for various memory accesses. well known DNNs. The accuracy is reported for the case where High throughputis necessary to deliver real-time perfor- only a single crop for a single model is used for classification, mance for interactive applications such as navigation and such that the number of weights and MACs in the table are reported in terms of the core area in squared millimeters per multiplier along with process technology. In terms of cost, different platforms will have different implementation-specific metrics. For instance, for an FPGA, (Number of CONV Layers) the specific device should be reported, along with the utilization of resources such as DSP, BRAM, LUT and FF; performance density such as GOPs/slice can also be reported. Stride Each processor should report various specifications for each NZ Weights metric as shown in Table V, using the Eyeriss chip as an example. It is important that all metrics and specifications are accounted for in order fairly evaluate all the design trade-offs. Number of Channels For instance, without the accuracy given for a specific dataset Number of Filters and task, one could run a simple DNN and easily claim low power, high throughput, and low cost – however, the processor might not be usable for a meaningful task; alternatively, without Total NZ MACs reporting the off-chip bandwidth, one could build a processor with only multipliers and easily claim low cost, high throughput, high accuracy, and lowchippower – however, when evaluating systempower, the off-chip memory access would be substantial. Finally, the test setup should also be reported, including whether the results are measured or obtained from simulation and consistent. (NZ) operations significantly reduces the number of MACs In summary, the evaluation process for whether a DNNand weights. Since the number of NZ MACs depends on the system is a viable solution for a given application might go asinput data, we propose using the publicly available 50,000 follows: (1) the accuracy determines if it can perform the givenvalidation images from ImageNet for the computation. Finally, task; (2) the latency and throughput determine if it can run fast there are various methods to reduce the weights in a DNN enough and in real-time; (3) the energy and power consumption (e.g., network pruning in SectionVII-B2). Table IV shows will primarily dictate the form factor of the device where the another example of these DNN model metrics, by comparing processing can operate; (4) the cost, which is primarily dictatedsparse DNNs pruned using [142] to dense DNNs. by the chip area, determines how much one would pay for this solution. <

> TABLE IV METRICS FOR POPULAR DNN MODELS. SPARSITY IS ACCOUNT FOR BY REPORTING NON-ZERO (NZ) WEIGHTS AND MACS. B. Metrics for DNN Hardware To measure the efficiency of the DNN hardware, we should IX. SUMMARY consider the following additional metrics: The use of deep neural networks (DNNs) has seen explosive Thepower and energyconsumption of the design should growth in the past few years. They are currently widely used be reported for various DNN models; the DNN model for many artificial intelligence (AI) applications including specifications should be provided including which layers computer vision, speech recognition and robotics and are often and bit precision are supported by the hardware during delivering better than human accuracy. However, while DNNs measurement. In addition, the amount of off-chip accesses can deliver this outstanding accuracy, it comes at the cost (e.g., DRAM accesses) should be included since it of high computational complexity. Consequently, techniques accounts for a significant portion of the system power; it that enable efficient processing of deep neural network to can be reported in terms of the total amount of data that improveenergy-efficiencyandthroughputwithout sacrificing is read and written off-chip per inference. accuracywith cost-effective hardware are critical to expanding Thelatency and throughputshould be reported in terms the deployment of DNNs in both existing and new domains. of the batch size and the actual run time for various Creating a system for efficient DNN processing should DNN models, which accounts for mapping and memory begin with understanding the current and future applications bandwidth effects. This provides a more useful and and the specific computations required both now and the informative metric than peak throughput. potential evolution of those computations. This article surveys a Thecostof the chip depends on the area efficiency, which number of the current applications, focusing on computer vision accounts for the size and type of memory (e.g., registers applications, the associated algorithms, and the data being used or SRAM) and the amount of control logic. It should be to drive the algorithms. These applications, algorithms and input data are experiencing rapid change. So extrapolating 13 Data augmentation is often used to increase accuracy. This includes using these trends to determine the degree of flexibility desired to multiple crops of an image to account for misalignment; in addition, an handle next generation computations, becomes an important ensemble of multiple models can be used where each model has different weights due to different training settings, such as using different initializations ingredient of any design project. or datasets, or even different network architectures. If multiple crops and models are used, then the number of MACs and weights required would <> TABLE V EXAMPLE BENCHMARK METRICS FOR EYERISS [94]. During the design-space exploration process, it is critical to article both reviews a variety of these techniques and discusses understand and balance the important system metrics. For DNN the frameworks that are available for describing, running and computation these include the accuracy, energy, throughput training networks. and hardware cost. Evaluating these metrics is, of course, Finally, DNNs afford the opportunity to use mixed-signal key, so this article surveys the important components of circuit design and advanced technologies to improve efficiency. a DNN workload. In specific, a DNN workload has two These include using memristors for analog computation and 3-D major components. First, the workload is the form of each stacked memory. Advanced technologies can also can facilitate DNN network including the ‘shape’ of each layer and the moving computation closer to the source by embedding compu- interconnections between layers. These can vary both within tation near or within the sensor and the memories. Of course, all and between applications. Second, the workload consists of of these techniques should also be considered in combination, the specific the data input to the DNN. This data will vary while being careful to understand their interactions and looking with the input set used for training or the data input during for opportunities for joint hardware/algorithm co-optimization. operation for inference. In conclusion, although much work has been done, deep This article also surveys a number of avenues that prior neural networks remain an important area of research with work have taken to optimize DNN processing. Since data many promising applications and opportunities for innovation movement dominates energy consumption, a primary focus at various levels of hardware design. of some recent research has been to reduce data movement while maintaining accuracy, throughput and cost. This means ACKNOWLEDGMENTS selecting architectures with favorable memory hierarchies like Funding provided by DARPA YFA, MIT CICS, and gifts a spatial array, and developing dataflows that increase data from Nvidia and Intel. The authors thank the anonymous reuse at the low-cost levels of the memory hierarchy. We reviewers as well as James Noraky, Mehul Tikekar and have included a taxonomy of dataflows and an analysis of Zhengdong Zhang for providing valuable feedback on this their characteristics. Other work is presented that aims to save paper. space and energy by changing the representation of data values in the DNN. Still other work saves energy and sometimes REFERENCES increases throughput by exploiting the sparsity of weights [1]Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,”Nature, and/or activations. vol. 521, no. 7553, pp. 436–444, May 2015. The DNN domain also affords an excellent opportunity [2]L. Deng, J. Li, J.-T. Huang, K. Yao, D. Yu, F. Seide, M. Seltzer, G. Zweig, X. He, J. Williamset al., “Recent advances in deep for joint hardware/software co-design. For example, various learning for speech research at Microsoft,” inICASSP, 2013. efforts have noted that efficiency can be improved by increasing [3]A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet sparsity (increasing the number of zero values) or optimizing Classification with Deep Convolutional Neural Networks,” in the representation of data by reducing the precision of values NIPS, 2012. or using more complex mappings of the stored value to the [4]C. Chen, A. Seff, A. Kornhauser, and J. Xiao, “Deepdriving: Learning affordance for direct perception in autonomous actual value used for computation. However, to avoid losing driving,” inICCV, 2015. accuracy it is often useful to modify the network or fine-tune the [5]A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. network’s weights to accommodate these changes. Thus, this Blau, and S. Thrun, “Dermatologist-level classification of skin 29 cancer with deep neural networks,”Nature, vol. 542, no. 7639, [25]J. Zhou and O. G. Troyanskaya, “Predicting effects of noncod- pp. 115–118, 2017. ing variants with deep learning-based sequence model,”Nature [6]D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, methods, vol. 12, no. 10, pp. 931–934, 2015. G. van den Driessche, J. Schrittwieser, I. Antonoglou, [26]B. Alipanahi, A. Delong, M. T. Weirauch, and B. J. Frey, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, “Predicting the sequence specificities of dna-and rna-binding J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, proteins by deep learning,”Nature biotechnology, vol. 33, no. 8, K. Kavukcuoglu, T. Graepel, and D. Hassabis, “Mastering the pp. 831–838, 2015. game of Go with deep neural networks and tree search,”Nature, [27]H. Zeng, M. D. Edwards, G. Liu, and D. K. Gifford, “Convolu- vol. 529, no. 7587, pp. 484–489, Jan. 2016. tional neural network architectures for predicting dna–protein [7]F.-F. Li, A. Karpathy, and J. Johnson, “Stanford CS class binding,”Bioinformatics, vol. 32, no. 12, pp. i121–i127, 2016. CS231n: Convolutional Neural Networks for Visual Recogni- [28]M. Jermyn, J. Desroches, J. Mercier, M.-A. Tremblay, K. St- tion,” http://cs231n.stanford.edu/. Arnaud, M.-C. Guiot, K. Petrecca, and F. Leblond, “Neural net- [8]P. A. Merolla, J. V. Arthur, R. Alvarez-Icaza, A. S. Cassidy, works improve brain cancer detection with raman spectroscopy J. Sawada, F. Akopyan, B. L. Jackson, N. Imam, C. Guo, in the presence of operating room light artifacts,”Journal of Y. Nakamuraet al., “A million spiking-neuron integrated circuit Biomedical Optics, vol. 21, no. 9, pp. 094002–094002, 2016. with a scalable communication network and interface,”Science, [29]D. Wang, A. Khosla, R. Gargeya, H. Irshad, and A. H. Beck, vol. 345, no. 6197, pp. 668–673, 2014. “Deep learning for identifying metastatic breast cancer,”arXiv [9]S. K. Esser, P. A. Merolla, J. V. Arthur, A. S. Cassidy, preprint arXiv:1606.05718, 2016. R. Appuswamy, A. Andreopoulos, D. J. Berg, J. L. McKinstry, [30]L. P. Kaelbling, M. L. Littman, and A. W. Moore, “Rein- T. Melano, D. R. Barchet al., “Convolutional networks for forcement learning: A survey,”Journal of artificial intelligence fast, energy-efficient neuromorphic computing,”Proceedings research, vol. 4, pp. 237–285, 1996. of the National Academy of Sciences, 2016. [31]V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, [10]M. Mathieu, M. Henaff, and Y. LeCun, “Fast training of D. Wierstra, and M. Riedmiller, “Playing Atari with Deep convolutional networks through FFTs,” inICLR, 2014. Reinforcement Learning,” inNIPS Deep Learning Workshop, [11]Y. LeCun, L. D. Jackel, B. Boser, J. S. Denker, H. P. Graf, 2013. I. Guyon, D. Henderson, R. E. Howard, and W. Hubbard, [32]S. Levine, C. Finn, T. Darrell, and P. Abbeel, “End-to-end “Handwritten digit recognition: applications of neural network training of deep visuomotor policies,”Journal of Machine chips and automatic learning,”IEEE Commun. Mag., vol. 27, Learning Research, vol. 17, no. 39, pp. 1–40, 2016. no. 11, pp. 41–46, Nov 1989. [33]M. Pfeiffer, M. Schaeuble, J. Nieto, R. Siegwart, and C. Cadena, [12]B. Widrow and M. E. Hoff, “Adaptive switching circuits,” in “From Perception to Decision: A Data-driven Approach to End- 1960 IRE WESCON Convention Record, 1960. to-end Motion Planning for Autonomous Ground Robots,” in [13]B. Widrow, “Thinking about thinking: the discovery of the ICRA, 2017. LMS algorithm,”IEEE Signal Process. Mag., 2005. [34]S. Gupta, J. Davidson, S. Levine, R. Sukthankar, and J. Malik, [14]O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, “Cognitive mapping and planning for visual navigation,” in Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, CVPR, 2017. and L. Fei-Fei, “ImageNet Large Scale Visual Recognition [35]T. Zhang, G. Kahn, S. Levine, and P. Abbeel, “Learning deep Challenge,”International Journal of Computer Vision (IJCV), control policies for autonomous aerial vehicles with mpc-guided vol. 115, no. 3, pp. 211–252, 2015. policy search,” inICRA, 2016. [15]K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning [36]S. Shalev-Shwartz, S. Shammah, and A. Shashua, “Safe, multi- for Image Recognition,” inCVPR, 2016. agent, reinforcement learning for autonomous driving,” inNIPS [16]“Complete Visual Networking Index (VNI) Forecast,” Cisco, Workshop on Learning, Inference and Control of Multi-Agent June 2016. Systems, 2016. [17]J. Woodhouse, “Big, big, big data: higher and higher resolution [37]N. Hemsoth, “The Next Wave of Deep Learning Applications,” video surveillance,” technology.ihs.com, January 2016. Next Platform, September 2016. [18]R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich [38]S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Feature Hierarchies for Accurate Object Detection and Semantic Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997. Segmentation,” inCVPR, 2014. [39]T. N. Sainath, A.-r. Mohamed, B. Kingsbury, and B. Ramab- [19]J. Long, E. Shelhamer, and T. Darrell, “Fully Convolutional hadran, “Deep convolutional neural networks for LVCSR,” in Networks for Semantic Segmentation,” inCVPR, 2015. ICASSP, 2013. [20]K. Simonyan and A. Zisserman, “Two-stream convolutional [40]V. Nair and G. E. Hinton, “Rectified Linear Units Improve networks for action recognition in videos,” inNIPS, 2014. Restricted Boltzmann Machines,” inICML, 2010. [21]G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, [41]A. L. Maas, A. Y. Hannun, and A. Y. Ng, “Rectifier nonlin- A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainathet al., “Deep earities improve neural network acoustic models,” inICML, neural networks for acoustic modeling in speech recognition: 2013. The shared views of four research groups,”IEEE Signal Process. [42]K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into Mag., vol. 29, no. 6, pp. 82–97, 2012. rectifiers: Surpassing human-level performance on imagenet [22]R. Collobert, J. Weston, L. Bottou, M. Karlen, K. Kavukcuoglu, classification,” inICCV, 2015. and P. Kuksa, “Natural language processing (almost) from [43]D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and scratch,”Journal of Machine Learning Research, vol. 12, no. Accurate Deep Network Learning by Exponential Linear Units Aug, pp. 2493–2537, 2011. (ELUs),”ICLR, 2016. [23]A. van den Oord, S. Dieleman, H. Zen, K. Simonyan, [44]X. Zhang, J. Trmal, D. Povey, and S. Khudanpur, “Improving O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior, and deep neural network acoustic models using generalized maxout K. Kavukcuoglu, “Wavenet: A generative model for raw audio,” networks,” inICASSP, 2014. CoRR abs/1609.03499, 2016. [45]Y. Zhang, M. Pezeshki, P. Brakel, S. Zhang, , C. Laurent, [24]H. Y. Xiong, B. Alipanahi, L. J. Lee, H. Bretschneider, Y. Bengio, and A. Courville, “Towards End-to-End Speech D. Merico, R. K. Yuen, Y. Hua, S. Gueroussov, H. S. Najafabadi, Recognition with Deep Convolutional Neural Networks,” in T. R. Hugheset al., “The human splicing code reveals new Interspeech, 2016. insights into the genetic determinants of disease,”Science, vol. [46] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Gir- 347, no. 6218, p. 1254806, 2015. shick, S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for fast feature embedding,” inACM International [75]J. Cong and B. Xiao, “Minimizing computation in convolutional Conference on Multimedia, 2014. neural networks,” inICANN, 2014. [47]S. Ioffe and C. Szegedy, “Batch normalization: Accelerating [76]A. Lavin and S. Gray, “Fast algorithms for convolutional neural deep network training by reducing internal covariate shift,” in networks,” inCVPR, 2016. ICML, 2015. [77]“Intel Math Kernel Library,” https://software.intel.com/en-us/ [48]Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient- mkl. based learning applied to document recognition,”Proc. IEEE, [78]S. Chetlur, C. Woolley, P. Vandermersch, J. Cohen, J. Tran, vol. 86, no. 11, pp. 2278–2324, Nov 1998. B. Catanzaro, and E. Shelhamer, “cuDNN: Efficient Primitives [49]P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and for Deep Learning,”arXiv preprint arXiv:1410.0759, 2014. Y. LeCun, “OverFeat: Integrated Recognition, Localization and [79]M. Horowitz, “Computing’s energy problem (and what we can Detection using Convolutional Networks,” inICLR, 2014. do about it),” inISSCC, 2014. [50]K. Simonyan and A. Zisserman, “Very Deep Convolutional [80]Y.-H. Chen, J. Emer, and V. Sze, “Eyeriss: A Spatial Archi- Networks for Large-Scale Image Recognition,” inICLR, 2015. tecture for Energy-Efficient Dataflow for Convolutional Neural [51]C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, Networks,” inISCA, 2016. D. Erhan, V. Vanhoucke, and A. Rabinovich, “Going Deeper [81]——, “Using Dataflow to Optimize Energy Efficiency of Deep With Convolutions,” inCVPR, 2015. Neural Network Accelerators,”IEEE Micro’s Top Picks from the [52]M. Lin, Q. Chen, and S. Yan, “Network in Network,” inICLR, Computer Architecture Conferences, vol. 37, no. 3, May-June 2014. 2017. [53]C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, [82]M. Sankaradas, V. Jakkula, S. Cadambi, S. Chakradhar, I. Dur- “Rethinking the inception architecture for computer vision,” in danovic, E. Cosatto, and H. P. Graf, “A Massively Parallel CVPR, 2016. Coprocessor for Convolutional Neural Networks,” inASAP, [54]C. Szegedy, S. Ioffe, V. Vanhoucke, and A. Alemi, “Inception- 2009. v4, Inception-ResNet and the Impact of Residual Connections [83]V. Sriram, D. Cox, K. H. Tsoi, and W. Luk, “Towards an on Learning,” inAAAI, 2017. embedded biologically-inspired machine vision processor,” in [55]G. Urban, K. J. Geras, S. E. Kahou, O. Aslan, S. Wang, FPT, 2010. R. Caruana, A. Mohamed, M. Philipose, and M. Richardson, [84]S. Chakradhar, M. Sankaradas, V. Jakkula, and S. Cadambi, “Do Deep Convolutional Nets Really Need to be Deep and “A Dynamically Configurable Coprocessor for Convolutional Convolutional?”ICLR, 2017. Neural Networks,” inISCA, 2010. [56]“Caffe LeNet MNIST,” http://caffe.berkeleyvision.org/gathered/ [85]V. Gokhale, J. Jin, A. Dundar, B. Martini, and E. Culurciello, examples/mnist.html. “A 240 G-ops/s Mobile Coprocessor for Deep Neural Networks,” [57]“Caffe Model Zoo,” http://caffe.berkeleyvision.org/modelzoo. inCVPR Workshop, 2014. html. [86]S. Park, K. Bong, D. Shin, J. Lee, S. Choi, and H.-J. Yoo, “A [58]“Matconvnet Pretrained Models,” http://www.vlfeat.org/ 1.93TOPS/W scalable deep learning/inference processor with matconvnet/pretrained/. tetra-parallel MIMD architecture for big-data applications,” in [59]“TensorFlow-Slim image classification library,” https://github. ISSCC, 2015. com/tensorflow/models/tree/master/slim. [87]L. Cavigelli, D. Gschwend, C. Mayer, S. Willi, B. Muheim, and [60]“Deep Learning Frameworks,” https://developer.nvidia.com/ L. Benini, “Origami: A Convolutional Network Accelerator,” deep-learning-frameworks. inGLVLSI, 2015. [61]Y.-H. Chen, T. Krishna, J. Emer, and V. Sze, “Eyeriss: An [88]S. Gupta, A. Agrawal, K. Gopalakrishnan, and P. Narayanan, Energy-Efficient Reconfigurable Accelerator for Deep Convolu- “Deep Learning with Limited Numerical Precision,” inICML, tional Neural Networks,”IEEE J. Solid-State Circuits, vol. 51, 2015. no. 1, 2017. [89]Z. Du, R. Fasthuber, T. Chen, P. Ienne, L. Li, T. Luo, [62]C. J. B. Yann LeCun, Corinna Cortes, “THE MNIST X. Feng, Y. Chen, and O. Temam, “ShiDianNao: Shifting DATABASE of handwritten digits,” http://yann.lecun.com/exdb/ Vision Processing Closer to the Sensor,” inISCA, 2015. mnist/. [90]M. Peemen, A. A. A. Setio, B. Mesman, and H. Corporaal, [63]L. Wan, M. Zeiler, S. Zhang, Y. L. Cun, and R. Fergus, “Memory-centric accelerator design for Convolutional Neural “Regularization of neural networks using dropconnect,” inICML, Networks,” inICCD, 2013. 2013. [91]C. Zhang, P. Li, G. Sun, Y. Guan, B. Xiao, and J. Cong, “Opti- [64]A. Krizhevsky, V. Nair, and G. Hinton, “The CIFAR-10 dataset,” mizing FPGA-based Accelerator Design for Deep Convolutional https://www.cs.toronto.edu/ kriz/cifar.html. Neural Networks,” inFPGA, 2015. [65]A. Torralba, R. Fergus, and W. T. Freeman, “80 million tiny [92]T. Chen, Z. Du, N. Sun, J. Wang, C. Wu, Y. Chen, and images: A large data set for nonparametric object and scene O. Temam, “DianNao: A Small-footprint High-throughput recognition,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 30, Accelerator for Ubiquitous Machine-learning,” inASPLOS, no. 11, pp. 1958–1970, 2008. 2014. [66]A. Krizhevsky and G. Hinton, “Convolutional deep belief [93]Y. Chen, T. Luo, S. Liu, S. Zhang, L. He, J. Wang, L. Li, networks on cifar-10,”Unpublished manuscript, vol. 40, 2010. T. Chen, Z. Xu, N. Sun, and O. Temam, “DaDianNao: A [67]B. Graham, “Fractional max-pooling,” arXiv preprint Machine-Learning Supercomputer,” inMICRO, 2014. arXiv:1412.6071, 2014. [94]Y.-H. Chen, T. Krishna, J. Emer, and V. Sze, “Eyeriss: An [68]“Pascal VOC data sets,” http://host.robots.ox.ac.uk/pascal/ Energy-Efficient Reconfigurable Accelerator for Deep Convo- VOC/. lutional Neural Networks,” inISSCC, 2016. [69]“Microsoft Common Objects in Context (COCO) dataset,” http: [95]V. Sze, M. Budagavi, and G. J. Sullivan, “High Efficiency Video //mscoco.org/. Coding (HEVC): Algorithms and Architectures,” inIntegrated [70]“Google Open Images,” https://github.com/openimages/dataset. Circuit and Systems. Springer, 2014, pp. 1–375. [71]“YouTube-8M,” https://research.google.com/youtube8m/. [96]M. Alwani, H. Chen, M. Ferdman, and P. Milder, “Fused-layer [72]“AudioSet,” https://research.google.com/audioset/index.html. CNN accelerators,” inMICRO, 2016. [73]S. Condon, “Facebook unveils Big Basin, new server geared [97]D. Keitel-Schulz and N. Wehn, “Embedded DRAM develop- for deep learning,” ZDNet, March 2017. ment: Technology, physical design, and application issues,” [74] C. Dubout and F. Fleuret, “Exact acceleration of linear object IEEE Des. Test. Comput., vol. 18, no. 3, pp. 7–15, 2001. detectors,” inECCV, 2012. [98]J. Jeddeloh and B. Keeth, “Hybrid memory cube new DRAM architecture increases density and performance,” inSymp. on and modularized RTL compilation of Convolutional Neural VLSI, 2012. Networks onto FPGA,” inFPL, 2016. [99]J. Standard, “High bandwidth memory (HBM) DRAM,” [122]P. Gysel, M. Motamedi, and S. Ghiasi, “Hardware-oriented JESD235, 2013. Approximation of Convolutional Neural Networks,” inICLR, [100]D. Kim, J. Kung, S. Chai, S. Yalamanchili, and S. Mukhopad- 2016. hyay, “Neurocube: A programmable digital neuromorphic [123]S. Higginbotham, “Google Takes Unconventional Route with architecture with high-density 3D memory,” inISCA, 2016. Homegrown Machine Learning Chips,” Next Platform, May [101]M. Gao, J. Pu, X. Yang, M. Horowitz, and C. Kozyrakis, 2016. “TETRIS: Scalable and Efficient Neural Network Acceleration [124]T. P. Morgan, “Nvidia Pushes Deep Learning Inference With with 3D Memory,” inASPLOS, 2017. New Pascal GPUs,” Next Platform, September 2016. [102]J. Zhang, Z. Wang, and N. Verma, “A machine-learning [125]P. Judd, J. Albericio, T. Hetherington, T. M. Aamodt, and classifier implemented in a standard 6T SRAM array,” inSymp. A. Moshovos, “Stripes: Bit-serial deep neural network comput- on VLSI, 2016. ing,” inMICRO, 2016. [103]Z. Wang, R. Schapire, and N. Verma, “Error-adaptive classifier [126]B. Moons and M. Verhelst, “A 0.3–2.6 TOPS/W precision- boosting (EACB): Exploiting data-driven training for highly scalable processor for real-time large-scale ConvNets,” inSymp. fault-tolerant hardware,” inICASSP, 2014. on VLSI, 2016. [104]A. Shafiee, A. Nag, N. Muralimanohar, R. Balasubramonian, [127]M. Courbariaux, Y. Bengio, and J.-P. David, “Binaryconnect: J. P. Strachan, M. Hu, R. S. Williams, and V. Srikumar, “ISAAC: Training deep neural networks with binary weights during A Convolutional Neural Network Accelerator with In-Situ propagations,” inNIPS, 2015. Analog Arithmetic in Crossbars,” inISCA, 2016. [128]M. Courbariaux and Y. Bengio, “Binarynet: Training deep [105]L. Chua, “Memristor-the missing circuit element,”IEEE Trans. neural networks with weights and activations constrained to+ Circuit Theory, vol. 18, no. 5, pp. 507–519, 1971. 1 or-1,”arXiv preprint arXiv:1602.02830, 2016. [106]L. Wilson, “International technology roadmap for semiconduc- [129]M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi, “XNOR- tors (ITRS),”Semiconductor Industry Association, 2013. Net: ImageNet Classification Using Binary Convolutional [107]Lu, Darsen, “Tutorial on Emerging Memory Devices,” 2016. Neural Networks,” inECCV, 2016. [108]S. B. Eryilmaz, S. Joshi, E. Neftci, W. Wan, G. Cauwenberghs, [130]Z. Cai, X. He, J. Sun, and N. Vasconcelos, “Deep learning with and H.-S. P. Wong, “Neuromorphic architectures with electronic low precision by half-wave gaussian quantization,” inCVPR, synapses,” inISQED, 2016. 2017. [109]P. Chi, S. Li, Z. Qi, P. Gu, C. Xu, T. Zhang, J. Zhao, Y. Liu, [131]F. Li and B. Liu, “Ternary weight networks,” inNIPS Workshop Y. Wang, and Y. Xie, “PRIME: A Novel Processing-In-Memory on Efficient Methods for Deep Neural Networks, 2016. Architecture for Neural Network Computation in ReRAM-based [132]C. Zhu, S. Han, H. Mao, and W. J. Dally, “Trained Ternary Main Memory,” inISCA, 2016. Quantization,”ICLR, 2017. [110]M. Prezioso, F. Merrikh-Bayat, B. Hoskins, G. Adam, K. K. [133]R. Andri, L. Cavigelli, D. Rossi, and L. Benini, “YodaNN: An Likharev, and D. B. Strukov, “Training and operation of Ultra-Low Power Convolutional Neural Network Accelerator an integrated neuromorphic network based on metal-oxide Based on Binary Weights,” inISVLSI, 2016. memristors,”Nature, vol. 521, no. 7550, pp. 61–64, 2015. [134]K. Ando, K. Ueyoshi, K. Orimo, H. Yonekawa, S. Sato, [111]J. Zhang, Z. Wang, and N. Verma, “A matrix-multiplying ADC H. Nakahara, M. Ikebe, T. Asai, S. Takamaeda-Yamazaki, and implementing a machine-learning classifier directly with data M. Kuroda, T.and Motomura, “BRein Memory: A 13-Layer conversion,” inISSCC, 2015. 4.2 K Neuron/0.8 M Synapse Binary/Ternary Reconfigurable [112]E. H. Lee and S. S. Wong, “A 2.5 GHz 7.7 TOPS/W switched- In-Memory Deep Neural Network Accelerator in 65nm CMOS,” capacitor matrix multiplier with co-designed local memory in inSymp. on VLSI, 2017. 40nm,” inISSCC, 2016. [135]D. Miyashita, E. H. Lee, and B. Murmann, “Convolutional [113]R. LiKamWa, Y. Hou, J. Gao, M. Polansky, and L. Zhong, Neural Networks using Logarithmic Data Representation,” “RedEye: analog ConvNet image sensor architecture for contin- arXiv preprint arXiv:1603.01025, 2016. uous mobile vision,” inISCA, 2016. [136]A. Zhou, A. Yao, Y. Guo, L. Xu, and Y. Chen, “Incremental [114]A. Wang, S. Sivaramakrishnan, and A. Molnar, “A 180nm Network Quantization: Towards Lossless CNNs with Low- CMOS image sensor with on-chip optoelectronic image com- precision Weights,” inICLR, 2017. pression,” inCICC, 2012. [137]W. Chen, J. T. Wilson, S. Tyree, K. Q. Weinberger, and Y. Chen, [115]H. Chen, S. Jayasuriya, J. Yang, J. Stephen, S. Sivaramakrish- “Compressing Neural Networks with the Hashing Trick,” in nan, A. Veeraraghavan, and A. Molnar, “ASP Vision: Optically ICML, 2015. Computing the First Layer of Convolutional Neural Networks [138]J. Albericio, P. Judd, T. Hetherington, T. Aamodt, N. E. Jerger, using Angle Sensitive Pixels,” inCVPR, 2016. and A. Moshovos, “Cnvlutin: ineffectual-neuron-free deep [116]A. Suleiman and V. Sze, “Energy-efficient HOG-based object neural network computing,” inISCA, 2016. detection at 1080HD 60 fps with multi-scale support,” inSiPS, [139]B. Reagen, P. Whatmough, R. Adolf, S. Rama, H. Lee, S. K. 2014. Lee, J. M. Hernandez-Lobato, G.-Y. Wei, and D. Brooks,´ [117]E. H. Lee, D. Miyashita, E. Chai, B. Murmann, and S. S. Wong, “Minerva: Enabling low-power, highly-accurate deep neural “Lognet: Energy-Efficient Neural Networks Using Logrithmic network accelerators,” inISCA, 2016. Computations,” inICASSP, 2017. [140]Y. LeCun, J. S. Denker, and S. A. Solla, “Optimal Brain [118]S. Han, H. Mao, and W. J. Dally, “Deep Compression: Damage,” inNIPS, 1990. Compressing Deep Neural Networks with Pruning, Trained [141]S. Han, J. Pool, J. Tran, and W. J. Dally, “Learning both weights Quantization and Huffman Coding,” inICLR, 2016. and connections for efficient neural networks,” inNIPS, 2015. [119] I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Ben- [142]T.-J. Yang, Y.-H. Chen, and V. Sze, “Designing Energy-Efficient gio, “Quantized neural networks: Training neural networks Convolutional Neural Networks using Energy-Aware Pruning,” with low precision weights and activations,”arXiv preprint inCVPR, 2017. arXiv:1609.07061, 2016. [143]“DNN Energy Estimation,” http://eyeriss.mit.edu/energy.html. [120]S. Zhou, Y. Wu, Z. Ni, X. Zhou, H. Wen, and Y. Zou, “DoReFa- [144]R. Dorrance, F. Ren, and D. Markovic, “A scalable sparse´ Net: Training low bitwidth convolutional neural networks with matrix-vector multiplication kernel for energy-efficient sparse- low bitwidth gradients,”arXiv preprint arXiv:1606.06160, 2016. blas on FPGAs,” inISFPGA, 2014. [121]Y. Ma, N. Suda, Y. Cao, J.-S. Seo, and S. Vrudhula, “Scalable [145]S. Han, X. Liu, H. Mao, J. Pu, A. Pedram, M. A. Horowitz, and W. J. Dally, “EIE: efficient inference engine on compressed deep neural network,” inISCA, 2016. [146]A. Parashar, M. Rhu, A. Mukkara, A. Puglielli, R. Venkatesan, B. Khailany, J. Emer, S. W. Keckler, and W. J. Dally, “Scnn: An accelerator for compressed-sparse convolutional neural networks,” inISCA, 2017. [147]W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li, “Learning structured sparsity in deep neural networks,” inNIPS, 2016. [148]S. Anwar, K. Hwang, and W. Sung, “Structured pruning of deep convolutional neural networks,”ACM Journal of Emerging Technologies in Computing Systems, vol. 13, no. 3, p. 32, 2017. [149]J. Yu, A. Lukefahr, D. Palframan, G. Dasika, R. Das, and S. Mahlke, “Scalpel: Customizing dnn pruning to the underlying hardware parallelism,” inISCA, 2017. [150]H. Mao, S. Han, J. Pool, W. Li, X. Liu, Y. Wang, and W. J. Dally, “Exploring the regularity of sparse structure in convolutional neural networks,” inCVPR Workshop on Tensor Methods In Computer Vision, 2017. [151]J. S. Lim, “Two-dimensional signal and image processing,” Englewood Cliffs, NJ, Prentice Hall, 1990, 710 p., 1990. [152]F. Chollet, “Xception: Deep Learning With Depthwise Separa- ble Convolutions,”CVPR, 2017. [153]A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convolutional neural networks for mobile vision applications,” arXiv preprint arXiv:1704.04861, 2017. [154]F. N. Iandola, M. W. Moskewicz, K. Ashraf, S. Han, W. J. Dally, and K. Keutzer, “SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and<1MB model size,”ICLR, 2017. [155]E. Denton, W. Zaremba, J. Bruna, Y. LeCun, and R. Fergus, “Exploiting Linear Structure Within Convolutional Networks for Efficient Evaluation,” inNIPS, 2014. [156]V. Lebedev, Y. Ganin, M. Rakhuba1, I. Oseledets, and V. Lem- pitsky, “Speeding-Up Convolutional Neural Networks Using Fine-tuned CP-Decomposition,”ICLR, 2015. [157]Y.-D. Kim, E. Park, S. Yoo, T. Choi, L. Yang, and D. Shin, “Compression of Deep Convolutional Neural Networks for Fast and Low Power Mobile Applications,” inICLR, 2016. [158]C. Bucilu, R. Caruana, and A. Niculescu-Mizil, “Model Compression,” inSIGKDD, 2006. [159]L. Ba and R. Caurana, “Do Deep Nets Really Need to be Deep?”NIPS, 2014. [160]G. Hinton, O. Vinyals, and J. Dean, “Distilling the Knowledge in a Neural Network,” inNIPS Deep Learning Workshop, 2014. [161]A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y. Bengio, “Fitnets: Hints for Thin Deep Nets,”ICLR, 2015. [162]“Benchmarking DNN Processors,” http://eyeriss.mit.edu/benchmarking.html. < <> <> <> <> <> EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks Abstract Convolutional Neural Networks (ConvNets) are commonly developed at a fixed resource budget, and then scaled up for better accuracy if more resources are available. In this paper, we systematically study model scaling and identify that carefully balancing network depth, width, and resolution can lead to better performance. Based on this observation, we propose a new scaling method that uniformly scales all dimensions of depth/width/resolution using a simple yet highly effective compound coefficient. We demonstrate the effectiveness of this method on scaling up MobileNets and ResNet. To go even further, we use neural architecture search to design a new baseline network and scale it up to obtain a family of models, called EfficientNets, which achieve much better accuracy and efficiency than previous ConvNets. In particular, our EfficientNet-B7 achieves state-of-the-art 84.4% top-1 / 97.1% top-5 accuracy on ImageNet, while being 8.4x smaller and 6.1x faster on inference than the best existing ConvNet. Our EfficientNets also transfer well and achieve state-of-the-art accuracy on CIFAR-100 (91.7%), Flowers (98.8%), and 3 other transfer learning datasets, with an order of magnitude fewer parameters. Source code is at https: //github.com/tensorflow/tpu/tree/master/models/official/efficientnet. 1. Introduction Scaling up ConvNets is widely used to achieve better accuracy. For example, ResNet (He et al., 2016) can be scaled up from ResNet-18 to ResNet-200 by using more layers; Recently, GPipe (Huang et al., 2018) achieved 84.3% Ima. geNet top-1 accuracy by scaling up a baseline model four <

> Figure 1. Model Size vs. ImageNet Accuracy. All numbers are for single-crop, single-model. Our EfficientNets significantly out.perform other ConvNets. In particular, EfficientNet-B7 achieves new state-of-the-art 84.4% top-1 accuracy but being 8.4x smaller and 6.1x faster than GPipe. EfficientNet-B1 is 7.6x smaller and 5.7x faster than ResNet-152. Details are in Table 2 and 4. time larger. However, the process of scaling up ConvNets has never been well understood and there are currently many ways to do it. The most common way is to scale up Con.vNets by their depth (He et al., 2016) or width (Zagoruyko & Komodakis, 2016). Another less common, but increasingly popular, method is to scale up models by image resolution (Huang et al., 2018). In previous work, it is common to scale only one of the three dimensions � depth, width, and image size. Though it is possible to scale two or three dimensions arbitrarily, arbitrary scaling requires tedious manual tuning and still often yields sub-optimal accuracy and efficiency. In this paper, we want to study and rethink the process of scaling up ConvNets. In particular, we investigate the central question: is there a principled method to scale up ConvNets that can achieve better accuracy and efficiency? Our empirical study shows that it is critical to balance all dimensions of network width/depth/resolution, and surpris.ingly such balance can be achieved by simply scaling each of them with constant ratio. Based on this observation, we propose a simple yet effective compound scaling method. Unlike conventional practice that arbitrary scales these fac.tors, our method uniformly scales network width, depth, <

> Figure 2. Model Scaling. (a) is a baseline network example; (b)-(d) are conventional scaling that only increases one dimension of network width, depth, or resolution. (e) is our proposed compound scaling method that uniformly scales all three dimensions with a fixed ratio. and resolution with a set of fixed scaling coefficients. For example, if we want to use 2N times more computational resources, then we can simply increase the network depth by <>, width by <> , and image size by <> are constant coefficients determined by a small grid search on the original small model. Figure 2 illustrates the difference between our scaling method and conventional methods. Intuitively, the compound scaling method makes sense be.cause if the input image is bigger, then the network needs more layers to increase the receptive field and more channels to capture more fine-grained patterns on the bigger image. In fact, previous theoretical (Raghu et al., 2017; Lu et al., 2018) and empirical results (Zagoruyko & Komodakis, 2016) both show that there exists certain relationship between network width and depth, but to our best knowledge, we are the first to empirically quantify the relationship among all three dimensions of network width, depth, and resolution. We demonstrate that our scaling method work well on exist.ing MobileNets (Howard et al., 2017; Sandler et al., 2018) and ResNet (He et al., 2016). Notably, the effectiveness of model scaling heavily depends on the baseline network; to go even further, we use neural architecture search (Zoph & Le, 2017; Tan et al., 2019) to develop a new baseline network, and scale it up to obtain a family of models, called EfficientNets. Figure 1 summarizes the ImageNet performance, where our EfficientNets significantly outperform other ConvNets. In particular, our EfficientNet-B7 surpasses the best existing GPipe accuracy (Huang et al., 2018), but using 8.4x fewer parameters and running 6.1x faster on inference. Compared to the widely used ResNet-50 (He et al., 2016), our EfficientNet-B4 improves the top-1 accuracy from 76.3% to 83.0% (+6.7%) with similar FLOPS. Besides ImageNet, EfficientNets also transfer well and achieve state-of-the-art accuracy on 5 out of 8 widely used datasets, while reducing parameters by up to 21x than existing ConvNets. 2. Related Work ConvNet Accuracy: Since AlexNet (Krizhevsky et al., 2012) won the 2012 ImageNet competition, ConvNets have become increasingly more accurate by going bigger: while the 2014 ImageNet winner GoogleNet (Szegedy et al., 2015) achieves 74.8% top-1 accuracy with about 6.8M parameters, the 2017 ImageNet winner SENet (Hu et al., 2018) achieves 82.7% top-1 accuracy with 145M parameters. Recently, GPipe (Huang et al., 2018) further pushes the state-of-the-art ImageNet top-1 validation accuracy to 84.3% using 557M parameters: it is so big that it can only be trained with a specialized pipeline parallelism library by partitioning the network and spreading each part to a different accelerator. While these models are mainly designed for ImageNet, recent studies have shown better ImageNet models also per.form better across a variety of transfer learning datasets (Kornblith et al., 2019), and other computer vision tasks such as object detection (He et al., 2016; Tan et al., 2019). Although higher accuracy is critical for many applications, we have already hit the hardware memory limit, and thus further accuracy gain needs better efficiency. ConvNet efficiency: Deep ConvNets are often over-parameterized. Model compression (Han et al., 2016; He et al., 2018; Yang et al., 2018) is a common way to re.duce model size by trading accuracy for efficiency. As mo.bile phones become ubiquitous, it is also common to hand.craft efficient mobile-size ConvNets, such as SqueezeNets (Iandola et al., 2016; Gholami et al., 2018), MobileNets (Howard et al., 2017; Sandler et al., 2018), and ShuffleNets (Zhang et al., 2018; Ma et al., 2018). Recently, neural architecture search becomes increasingly popular in designing efficient mobile-size ConvNets (Tan et al., 2019; Cai et al., 2019), and achieves even better efficiency than hand-crafted mobile ConvNets by extensively tuning the network width, depth, convolution kernel types and sizes. However, it is unclear how to apply these techniques for larger models that have much larger design space and much more expensive tuning cost. In this paper, we aim to study model efficiency for super large ConvNets that surpass state-of-the-art accuracy. To achieve this goal, we resort to model scaling. Model Scaling: There are many ways to scale a Con.vNet for different resource constraints: ResNet (He et al., 2016) can be scaled down (e.g., ResNet-18) or up (e.g., ResNet-200) by adjusting network depth (#layers), while WideResNet (Zagoruyko & Komodakis, 2016) and Mo.bileNets (Howard et al., 2017) can be scaled by network width (#channels). It is also well-recognized that bigger input image size will help accuracy with the overhead of more FLOPS. Although prior studies (Raghu et al., 2017; Lin & Jegelka, 2018; Sharir & Shashua, 2018; Lu et al., 2018) have shown that network depth and width are both important for ConvNets� expressive power, it still remains an open question of how to effectively scale a ConvNet to achieve better efficiency and accuracy. Our work systematically and empirically studies ConvNet scaling for all three dimensions of network width, depth, and resolutions. 3. Compound Model Scaling In this section, we will formulate the scaling problem, study different approaches, and propose our new scaling method. 3.1. Problem Formulation A ConvNet Layer i can be defined as a function: <>, where Fi is the operator, Yi is output tensor, Xi is input tensor, with tensor shape <>, where H_i and W_i are spatial dimension and C_i is the channel dimension. A ConvNet N can be represented by a list of composed lay- <> practice, ConvNet layers are often partitioned into multiple stages and all layers in each stage share the same architecture: for example, ResNet (He et al., 2016) has �ve stages, and all layers in each stage has the same convolutional type except the first layer performs down-sampling. Therefore, we can define a ConvNet as: <> <> where <> denotes layer F_i is repeated L_i times in stage i, <> denotes the shape of input tensor X of layer 1For the sake of simplicity, we omit batch dimension. i. Figure 2(a) illustrate a representative ConvNet, where the spatial dimension is gradually shrunk but the channel dimension is expanded over layers, for example, from initial input shape h224, 224, 3i to final output shape h7, 7, 512i. Unlike regular ConvNet designs that mostly focus on find.ing the best layer architecture Fi, model scaling tries to expand the network length (Li), width (Ci), and/or resolution (Hi,Wi) without changing Fi predefined in the baseline network. By �xing Fi, model scaling simplifies the design problem for new resource constraints, but it still remains a large design space to explore different <> for each layer. In order to further reduce the design space, we restrict that all layers must be scaled uniformly with constant ratio. Our target is to maximize the model accuracy for any given resource constraints, which can be formulated as an optimization problem: <> (2) where <> are coefficients for scaling network width, depth, and resolution; <> are predefined parameters in baseline network (see Table 1 as an example). 3.2. Scaling Dimensions The main difficulty of problem 2 is that the optimal d, w, r depend on each other and the values change under different resource constraints. Due to this difficulty, conventional methods mostly scale ConvNets in one of these dimensions: Depth (d): Scaling network depth is the most common way used by many ConvNets (He et al., 2016; Huang et al., 2017; Szegedy et al., 2015; 2016). The intuition is that deeper ConvNet can capture richer and more complex features, and generalize well on new tasks. However, deeper networks are also more difficult to train due to the vanishing gradient problem (Zagoruyko & Komodakis, 2016). Although several techniques, such as skip connections (He et al., 2016) and batch normalization (Ioffe & Szegedy, 2015), alleviate the training problem, the accuracy gain of very deep network diminishes: for example, ResNet-1000 has similar accuracy as ResNet-101 even though it has much more layers. Figure 3 (middle) shows our empirical study on scaling a baseline model with different depth coefficient d, further suggesting the diminishing accuracy return for very deep ConvNets. Width (w): Scaling network width is commonly used for small size models (Howard et al., 2017; Sandler et al., 2018; <

> Figure 3. Scaling Up a Baseline Model with Different Network Width (w), Depth (d), and Resolution (r) coefficients. Bigger networks with larger width, depth, or resolution tend to achieve higher accuracy, but the accuracy gain quickly saturate after reaching 80%, demonstrating the limitation of single dimension scaling. Baseline network is described in Table 1. Tan et al., 2019)2. As discussed in (Zagoruyko & Komodakis, 2016), wider networks tend to be able to capture more fine-grained features and are easier to train. However, extremely wide but shallow networks tend to have dif�cul.ties in capturing higher level features. Our empirical results in Figure 3 (left) show that the accuracy quickly saturates when networks become much wider with larger w. Resolution (r): With higher resolution input images, Con.vNets can potentially capture more fine-grained patterns. Starting from 224x224 in early ConvNets, modern Con.vNets tend to use 299x299 (Szegedy et al., 2016) or 331x331 (Zoph et al., 2018) for better accuracy. Recently, GPipe (Huang et al., 2018) achieves state-of-the-art ImageNet ac.curacy with 480x480 resolution. Higher resolutions, such as 600x600, are also widely used in object detection ConvNets (He et al., 2017; Lin et al., 2017). Figure 3 (right) shows the results of scaling network resolutions, where indeed higher resolutions improve accuracy, but the accuracy gain dimin.ishes for very high resolutions (r =1.0 denotes resolution 224x224 and r =2.5 denotes resolution 560x560). The above analyses lead us to the first observation: Observation 1 � Scaling up any dimension of network width, depth, or resolution improves accuracy, but the accuracy gain diminishes for bigger models. 3.3. Compound Scaling We empirically observe that different scaling dimensions are not independent. Intuitively, for higher resolution images, we should increase network depth, such that the larger receptive fields can help capture similar features that include more pixels in bigger images. Correspondingly, we should also increase network width when resolution is higher, in In some literature, scaling number of channels is called depth multiplier, which means the same as our width coefficient w. <

> Figure 4. Scaling Network Width for Different Baseline Net.works. Each dot in a line denotes a model with different width coefficient (w). All baseline networks are from Table 1. The first baseline network <> has 18 convolutional layers with resolution 224x224, while the last baseline <> has 36 layers with resolution 299x299. order to capture more fine-grained patterns with more pixels in high resolution images. These intuitions suggest that we need to coordinate and balance different scaling dimensions rather than conventional single-dimension scaling. To validate our intuitions, we compare width scaling under different network depths and resolutions, as shown in Figure 4. If we only scale network width w without changing depth <<(d=1.0)>> and resolution <<(r=1.0)>>, the accuracy saturates quickly. With deeper (d=2.0) and higher resolution <<(r=2.0)>>, width scaling achieves much better accuracy under the same FLOPS cost. These results lead us to the second observation: Observation 2 In order to pursue better accuracy and efficiency, it is critical to balance all dimensions of network width, depth, and resolution during ConvNet scaling. In fact, a few prior work (Zoph et al., 2018; Real et al., 2019) have already tried to arbitrarily balance network width and depth, but they all require tedious manual tuning. In this paper, we propose a new compound scaling method, which use a compound coefficient . to uniformly scales network width, depth, and resolution in a principled way: <> (3) where <> are constants that can be determined by a small grid search. Intuitively, . is a user-specified coefficient that controls how many more resources are available for model scaling, while <> specify how to assign these extra resources to network width, depth, and resolution respectively. Notably, the FLOPS of a regular convolution op is proportional to <> i.e., doubling network depth will double FLOPS, but doubling network width or resolution will increase FLOPS by four times. Since convolution ops usually dominate the computation cost in ConvNets, scaling a ConvNet with equation 3 will approximately in. crease total FLOPS by <> In this paper, we constraint <> such that for any new <>, the total FLOPS will approximately3 increase by 2. 4. EfficientNet Architecture Since model scaling does not change layer operators F_i in baseline network, having a good baseline network is also critical. We will evaluate our scaling method using existing ConvNets, but in order to better demonstrate the effectiveness of our scaling method, we have also developed a new mobile-size baseline, called EfficientNet. Inspired by (Tan et al., 2019), we develop our baseline net.work by leveraging a multi-objective neural architecture search that optimizes both accuracy and FLOPS. Specifically, we use the same search space as (Tan et al., 2019), and use <> as the optimization goal, where <> and <> denote the accuracy and FLOPS of model m, T is the target FLOPS and w=-0.07 is a hyperparameter for controlling the trade-off between accuracy and FLOPS. Unlike (Tan et al., 2019; Cai et al., 2019), here we optimize FLOPS rather than latency since we are not targeting any specific hardware de.vice. Our search produces an efficient network, which we name EfficientNet-B0. Since we use the same search space as (Tan et al., 2019), the architecture is similar to <>. FLOPS may differ from theoretical value due to rounding. Table 1. EfficientNet-B0 baseline network <> Each row describes a stage i with L_i layers, with input resolution <> and output channels C_i. Notations are adopted from equation 2. <> Net, except our EfficientNet-B0 is slightly bigger due to the larger FLOPS target (our FLOPS target is 400M). Ta.ble 1 shows the architecture of EfficientNet-B0. Its main building block is mobile inverted bottleneck MBConv (Sandler et al., 2018; Tan et al., 2019), to which we also add squeeze-and-excitation optimization (Hu et al., 2018). Starting from the baseline EfficientNet-B0, we apply our compound scaling method to scale it up with two steps: STEP 1: we first <> assuming twice more re.sources available, and do a small grid search of <> based on Equation 2 and 3. In particular, we find the best values for EfficientNet-B0 are <>, under constraint of <>. STEP 2: we then <> as constants and scale up baseline network with different . using Equation 3, to obtain EfficientNet-B1 to B7 (Details in Table 2). Notably, it is possible to achieve even better performance by searching for <> directly around a large model, but the search cost becomes prohibitively more expensive on larger models. Our method solves this issue by only doing search once on the small baseline network (step 1), and then use the same scaling coefficients for all other models (step 2). 5. Experiments In this section, we will first evaluate our scaling method on existing ConvNets and the new proposed EfficientNets. 5.1. Scaling Up MobileNets and ResNets As a proof of concept, we first apply our scaling method to the widely-used MobileNets (Howard et al., 2017; Sandler et al., 2018) and ResNet (He et al., 2016). Table 3 shows the ImageNet results of scaling them in different ways. Compared to other single-dimension scaling methods, our compound scaling method improves the accuracy on all these models, suggesting the effectiveness of our proposed scaling method for general existing ConvNets. Table 2. EfficientNet Performance Results on ImageNet (Russakovsky et al., 2015). All EfficientNet models are scaled from our baseline EfficientNet-B0 using different compound coefficient . in Equation 3. ConvNets with similar top-1/top-5 accuracy are grouped together for efficiency comparison. Our scaled EfficientNet models consistently reduce parameters and FLOPS by an order of magnitude (up to 8.4x parameter reduction and up to 16x FLOPS reduction) than existing ConvNets. <

> We omit ensemble and multi-crop models (Hu et al., 2018), or models pretrained on 3.5B Instagram images (Mahajan et al., 2018). Table 3. Scaling Up MobileNets and ResNet. <> Table 5. EfficientNet Performance Results on Transfer Learning Datasets. Our scaled EfficientNet models achieve new state-of-the.art accuracy for 5 out of 8 datasets, with 9.6x fewer parameters on average. Comparison to best public-available results Comparison to best reported results Model Accuracy. <> Figure 6. Model Parameters vs. Transfer Learning Accuracy weight decay 1e-5; initial learning rate 0.256 that decays by 0.97 every 2.4 epochs. <

> We also use swish activation (Ramachandran et al., 2018; Elfwing et al., 2018), fixed Au.to Augment policy (Cubuk et al., 2019), and stochastic depth (Huang et al., 2016) with survival probability 0.8. As commonly known that bigger models need more regularization, we linearly increase dropout (Srivastava et al., 2014) ratio from 0.2 for EfficientNet-B0 to 0.5 for EfficientNet-B7. Table 2 shows the performance of all EfficientNet models that are scaled from the same baseline EfficientNet-B0. Our EfficientNet models generally use an order of magnitude fewer parameters and FLOPS than other ConvNets with similar accuracy. In particular, our EfficientNet-B7 achieves 84.4% top1 / 97.1% top-5 accuracy with 66M parameters and 37B FLOPS, being more accurate but 8.4x smaller than the previous best GPipe (Huang et al., 2018). All models are pretrained on ImageNet and fine tuned on new datasets. Figure 1 and Figure 5 illustrates the parameters-accuracy and FLOPS-accuracy curve for representative ConvNets, where our scaled EfficientNet models achieve better accuracy with much fewer parameters and FLOPS than other ConvNets. Notably, our EfficientNet models are not only small, but also computational cheaper. For example, our EfficientNet-B3 achieves higher accuracy than ResNeXt.101 (Xie et al., 2017) using 18x fewer FLOPS. To validate the computational cost, we have also measured the inference latency for a few representative CovNets on a real CPU as shown in Table 4, where we report average latency over 20 runs. Our EfficientNet-B1 runs 5.7x faster than the widely used ResNet-152 (He et al., 2016), while EfficientNet-B7 runs about 6.1x faster than GPipe (Huang et al., 2018), suggesting our EfficientNets are indeed fast on real hardware. Figure 7. Class Activation Map (CAM) (Zhou et al., 2016) for Models with different scaling methods-Our compound scaling method allows the scaled model (last column) to focus on more relevant regions with more object details. Model details are in Table 7. <

> Table 6. Transfer Learning Datasets. <

> 5.3. Transfer Learning Results for EfficientNet We have also evaluated our EfficientNet on a list of commonly used transfer learning datasets, as shown in Table 6. We borrow the same training settings from (Kornblith et al., 2019) and (Huang et al., 2018), which take ImageNet pretrained checkpoints and fine tune on new datasets. Table 5 shows the transfer learning performance: (1) Com.pared to public available models, such as NASNet-A (Zoph et al., 2018) and Inception-v4 (Szegedy et al., 2017), our EfficientNet models achieve better accuracy with 4.7x average (up to 21x) parameter reduction. (2) Compared to state-of-the-art models, including DAT (Ngiam et al., 2018) that dynamically synthesizes training data and GPipe (Huang et al., 2018) that is trained with specialized pipeline parallelism, our EfficientNet models still surpass their accuracy in 5 out of 8 datasets, but using 9.6x fewer parameters Figure 6 compares the accuracy-parameters curve for a variety of models. In general, our EfficientNets consistently achieve better accuracy with an order of magnitude fewer parameters than existing models, including ResNet (He et al., 2016), DenseNet (Huang et al., 2017), Inception (Szegedy et al., 2017), and NASNet (Zoph et al., 2018). 6. Discussion Figure 8. Scaling Up EfficientNet-B0 with Different Methods. Table 7. Scaled Models Used in Figure 7. <

> To disentangle the contribution of our proposed scaling method from the EfficientNet architecture, Figure 8 com.pares the ImageNet performance of different scaling methods for the same EfficientNet-B0 baseline network. In general, all scaling methods improve accuracy with the cost of more FLOPS, but our compound scaling method can further improve accuracy, by up to 2.5%, than other single-dimension scaling methods, suggesting the importance of our proposed compound scaling. In order to further understand why our compound scaling method is better than others, Figure 7 compares the class activation map (Zhou et al., 2016) for a few representative models with different scaling methods. All these models are scaled from the same baseline, and their statistics are shown in Table 7. Images are randomly picked from ImageNet validation set. As shown in the figure, the model with com.pound scaling tends to focus on more relevant regions with more object details, while other models are either lack of object details or unable to capture all objects in the images. 7. Conclusion In this paper, we systematically study ConvNet scaling and identify that carefully balancing network width, depth, and resolution is an important but missing piece, preventing us from better accuracy and efficiency. To address this issue, we propose a simple and highly effective compound scaling method, which enables us to easily scale up a baseline Con.vNet to any target resource constraints in a more principled way, while maintaining model efficiency. Powered by this compound scaling method, we demonstrate that a mobile-size EfficientNet model can be scaled up very effectively, surpassing state-of-the-art accuracy with an order of magnitude fewer parameters and FLOPS, on both ImageNet and five commonly used transfer learning datasets. Acknowledgements We thank Ruoming Pang, Vijay Vasudevan, Alok Aggarwal, Barret Zoph, Hongkun Yu, Xiaodan Song, Samy Bengio, Jeff Dean, and Google Brain team for their help. References Berg, T., Liu, J., Woo Lee, S., Alexander, M. L., Jacobs, D. W., and Belhumeur, P. N. Birdsnap: Large-scale fine-grained visual categorization of birds. CVPR, pp. 2011�2018, 2014. Bossard, L., Guillaumin, M., and Van Gool, L. Food-101� mining discriminative components with random forests. ECCV, pp. 446�461, 2014. Cai, H., Zhu, L., and Han, S. Proxylessnas: Direct neural architecture search on target task and hardware. ICLR, 2019. Chollet, F. Xception: Deep learning with depthwise separa.ble convolutions. CVPR, pp. 1610�02357, 2017. Cubuk, E. D., Zoph, B., Mane, D., Vasudevan, V., and Le, Q. V. Autoaugment: Learning augmentation policies from data. CVPR, 2019. Elfwing, S., Uchibe, E., and Doya, K. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural Networks, 107:3�11, 2018. Gholami, A., Kwon, K., Wu, B., Tai, Z., Yue, X., Jin, P., Zhao, S., and Keutzer, K. Squeezenext: Hardware-aware neural network design. ECV Workshop at CVPR�18, 2018. Han, S., Mao, H., and Dally, W. J. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. ICLR, 2016. He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. CVPR, pp. 770�778, 2016. He, K., Gkioxari, G., Dollar,� P., and Girshick, R. Mask r-cnn. ICCV, pp. 2980�2988, 2017. He, Y., Lin, J., Liu, Z., Wang, H., Li, L.-J., and Han, S. Amc: Automl for model compression and acceleration on mobile devices. ECCV, 2018. Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., and Adam, H. Mobilenets: efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. Hu, J., Shen, L., and Sun, G. Squeeze-and-excitation net.works. CVPR, 2018. Huang, G., Sun, Y., Liu, Z., Sedra, D., and Weinberger, K. Q. Deep networks with stochastic depth. ECCV, pp. 646�661, 2016. Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger, K. Q. Densely connected convolutional networks. CVPR, 2017. Huang, Y., Cheng, Y., Chen, D., Lee, H., Ngiam, J., Le, Q. V., and Chen, Z. Gpipe: efficient training of giant neural networks using pipeline parallelism. arXiv preprint arXiv:1808.07233, 2018. Iandola, F. N., Han, S., Moskewicz, M. W., Ashraf, K., Dally, W. J., and Keutzer, K. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. Ioffe, S. and Szegedy, C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. ICML, pp. 448�456, 2015. Kornblith, S., Shlens, J., and Le, Q. V. Do better imagenet models transfer better? CVPR, 2019. Krause, J., Deng, J., Stark, M., and Fei-Fei, L. Collecting a large-scale dataset of fine-grained cars. Second Workshop on Fine-Grained Visual Categorizatio, 2013. Krizhevsky, A. and Hinton, G. Learning multiple layers of features from tiny images. Technical Report, 2009. Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classi�cation with deep convolutional neural networks. In NIPS, pp. 1097�1105, 2012. Lin, H. and Jegelka, S. Resnet with one-neuron hidden layers is a universal approximator. NeurIPS, pp. 6172� 6181, 2018. Lin, T.-Y., Dollar,� P., Girshick, R., He, K., Hariharan, B., and Belongie, S. Feature pyramid networks for object detection. CVPR, 2017. Liu, C., Zoph, B., Shlens, J., Hua, W., Li, L.-J., Fei-Fei, L., Yuille, A., Huang, J., and Murphy, K. Progressive neural architecture search. ECCV, 2018. Lu, Z., Pu, H., Wang, F., Hu, Z., and Wang, L. The expres.sive power of neural networks: A view from the width. NeurIPS, 2018. Ma, N., Zhang, X., Zheng, H.-T., and Sun, J. Shuf�enet v2: Practical guidelines for efficient cnn architecture design. ECCV, 2018. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., and van der Maaten, L. Explor.ing the limits of weakly supervised pretraining. arXiv preprint arXiv:1805.00932, 2018. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., and Vedaldi, A. Fine-grained visual classi�cation of aircraft. arXiv preprint arXiv:1306.5151, 2013. Ngiam, J., Peng, D., Vasudevan, V., Kornblith, S., Le, Q. V., and Pang, R. Domain adaptive transfer learning with spe.cialist models. arXiv preprint arXiv:1811.07056, 2018. Nilsback, M.-E. and Zisserman, A. Automated �ower clas.si�cation over a large number of classes. ICVGIP, pp. 722�729, 2008. Parkhi, O. M., Vedaldi, A., Zisserman, A., and Jawahar, C. Cats and dogs. CVPR, pp. 3498�3505, 2012. Raghu, M., Poole, B., Kleinberg, J., Ganguli, S., and Sohl-Dickstein, J. On the expressive power of deep neural networks. ICML, 2017. Ramachandran, P., Zoph, B., and Le, Q. V. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2018. Real, E., Aggarwal, A., Huang, Y., and Le, Q. V. Regu.larized evolution for image classi�er architecture search. AAAI, 2019. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. Imagenet large scale visual recognition chal.lenge. International Journal of Computer Vision, 115(3): 211�252, 2015. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and Chen, L.-C. Mobilenetv2: Inverted residuals and linear bottlenecks. CVPR, 2018. Sharir, O. and Shashua, A. On the expressive power of overlapping architectures of deep learning. ICLR, 2018. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. Dropout: a simple way to prevent neural networks from over�tting. The Journal of Machine Learning Research, 15(1):1929�1958, 2014. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. Going deeper with convolutions. CVPR, pp. 1�9, 2015. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna, Z. Rethinking the inception architecture for computer vision. CVPR, pp. 2818�2826, 2016. Szegedy, C., Ioffe, S., Vanhoucke, V., and Alemi, A. A. Inception-v4, inception-resnet and the impact of residual connections on learning. AAAI, 4:12, 2017. Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M., Howard, A., and Le, Q. V. MnasNet: Platform-aware neural architecture search for mobile. CVPR, 2019. Xie, S., Girshick, R., Doll�ar, P., Tu, Z., and He, K. Aggre.gated residual transformations for deep neural networks. CVPR, pp. 5987�5995, 2017. Yang, T.-J., Howard, A., Chen, B., Zhang, X., Go, A., Sze, V., and Adam, H. Netadapt: Platform-aware neural net.work adaptation for mobile applications. ECCV, 2018. Zagoruyko, S. and Komodakis, N. Wide residual networks. BMVC, 2016. Zhang, X., Li, Z., Loy, C. C., and Lin, D. Polynet: A pursuit of structural diversity in very deep networks. CVPR, pp. 3900�3908, 2017. Zhang, X., Zhou, X., Lin, M., and Sun, J. Shuf�enet: An ex.tremely efficient convolutional neural network for mobile devices. CVPR, 2018. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., and Torralba, A. Learning deep features for discriminative localization. CVPR, pp. 2921�2929, 2016. Zoph, B. and Le, Q. V. Neural architecture search with reinforcement learning. ICLR, 2017. Zoph, B., Vasudevan, V., Shlens, J., and Le, Q. V. Learning transferable architectures for scalable image recognition. CVPR, 2018. <> <> <> <> <> <> Energy and Policy Considerations for Deep Learning in NLP Emma Strubell Ananya Ganesh Andrew McCallum College of Information and Computer Sciences University of Massachusetts Amherst {strubell, aganesh, Abstract Recent progress in hardware and methodology for training neural networks has ushered in a new generation of large networks trained on abundant data. These models have obtained notable gains in accuracy across many NLP tasks. However, these accuracy improvements depend on the availability of exception.ally large computational resources that necessitate similarly substantial energy consumption. As a result these models are costly to train and develop, both financially, due to the cost of hardware and electricity or cloud compute time, and environmentally, due to the car.bon footprint required to fuel modern tensor processing hardware. In this paper we bring this issue to the attention of NLP researchers by quantifying the approximate financial and environmental costs of training a variety of recently successful neural network models for NLP. Based on these findings, we propose actionable recommendations to reduce costs and improve equity in NLP research and practice. 1 Introduction Advances in techniques and hardware for train.ing deep neural networks have recently enabled impressive accuracy improvements across many fundamental NLP tasks (Bahdanau et al., 2015; Luong et al., 2015; Dozat and Manning, 2017; Vaswani et al., 2017), with the most computationally-hungry models obtaining the highest scores (Peters et al., 2018; Devlin et al., 2019; Radford et al., 2019; So et al., 2019). As a result, training a state-of-the-art model now re.quires substantial computational resources which demand considerable energy, along with the associated financial and environmental costs. Research and development of new models multiplies these costs by thousands of times by requiring re.training to experiment with model architectures and hyperparameters. Whereas a decade ago most <

> Table 1: Estimated CO2 emissions from training com.mon NLP models, compared to familiar consumption. NLP models could be trained and developed on a commodity laptop or server, many now require multiple instances of specialized hardware such as GPUs or TPUs, therefore limiting access to these highly accurate models on the basis of finances. Even when these expensive computational resources are available, model training also incurs a substantial cost to the environment due to the energy required to power this hardware for weeks or months at a time. Though some of this energy may come from renewable or carbon credit-offset resources, the high energy demands of these models are still a concern since (1) energy is not currently derived from carbon-neural sources in many locations, and (2) when renewable energy is available, it is still limited to the equipment we have to pro.duce and store it, and energy spent training a neural network might better be allocated to heating a family's home. It is estimated that we must cut carbon emissions by half over the next decade to deter escalating rates of natural disaster, and based on the estimated CO2 emissions listed in Table 1, 1Sources: (1) Air travel and per-capita consumption: https://bit.ly/2Hw0xWc; (2) car lifetime: https://bit.ly/2Qbr0w1. model training and development likely make up a substantial portion of the greenhouse gas emissions attributed to many NLP researchers. To heighten the awareness of the NLP community to this issue and promote mindful practice and policy, we characterize the dollar cost and carbon emissions that result from training the neural net.works at the core of many state-of-the-art NLP models. We do this by estimating the kilowatts of energy required to train a variety of popular off-the-shelf NLP models, which can be converted to approximate carbon emissions and electricity costs. To estimate the even greater resources re.quired to transfer an existing model to a new task or develop new models, we perform a case study of the full computational resources required for the development and tuning of a recent state-of-the-art NLP pipeline (Strubell et al., 2018). We conclude with recommendations to the community based on our findings, namely: (1) Time to retrain and sensitivity to hyperparameters should be reported for NLP machine learning models; (2) academic Researchers need equitable access to computational resources; and (3) researchers should prioritize developing efficient models and hardware. 2 Methods To quantify the computational and environmental cost of training deep neural network models for NLP, we perform an analysis of the energy required to train a variety of popular off-the-shelf NLP models, as well as a case study of the complete sum of resources required to develop LISA (Strubell et al., 2018), a state-of-the-art NLP model from EMNLP 2018, including all tuning and experimentation. We measure energy use as follows. We train the models described in 2.1 using the default settings provided, and sample GPU and CPU power con.sumption during training. Each model was trained for a maximum of 1 day. We train all models on a single NVIDIA Titan X GPU, with the exception of ELMo which was trained on 3 NVIDIA GTX 1080 Ti GPUs. While training, we repeatedly query the NVIDIA System Management Interface to sample the GPU power consumption and report the average over all samples. To sample CPU power consumption, we use Intel's Running Average Power Limit interface. <> Table 2: Percent energy sourced from: Renewable (e.g. hydro, solar, wind), natural gas, coal and nuclear for the top 3 cloud compute providers (Cook et al., 2017), compared to the United States,4 China5 and Germany (Burger, 2019). We estimate the total time expected for models to train to completion using training times and hardware reported in the original papers. We then calculate the power consumption in kilowatt-hours (kWh) as follows. Let pc be the average power draw (in watts) from all CPU sockets during train.ing, let pr be the average power draw from all DRAM (main memory) sockets, let pg be the aver.age power draw of a GPU during training, and let g be the number of GPUs used to train. We esti.mate total power consumption as combined GPU, CPU and DRAM consumption, then multiply this by Power Usage Effectiveness (PUE), which ac.counts for the additional energy required to sup.port the compute infrastructure (mainly cooling). We use a PUE coefficient of 1.58, the 2018 global average for data centers (Ascierto, 2018). Then the total power pt required at a given instance during training is given by: <> (1) The U.S. Environmental Protection Agency (EPA) provides average CO2 produced (in pounds per kilowatt-hour) for power consumed in the U.S. (EPA, 2018), which we use to convert power to estimated CO2 emissions: <> (2) This conversion takes into account the relative pro.portions of different energy sources (primarily nat.ural gas, coal, nuclear and renewable) consumed to produce energy in the United States. Table 2 lists the relative energy sources for China, Ger.many and the United States compared to the top three cloud service providers. The U.S. break.down of energy is comparable to that of the most popular cloud compute service, Amazon Web Ser.vices, so we believe this conversion to provide a reasonable estimate of CO2 emissions per kilowatt hour of compute energy used. 2.1 Models We analyze four models, the computational requirements of which we describe below. All models have code freely available online, which we used out-of-the-box. For more details on the models themselves, please refer to the original papers. Transformer. The Transformer model (Vaswani et al., 2017) is an encoder-decoder architecture primarily recognized for efficient and accurate ma.chine translation. The encoder and decoder each consist of 6 stacked layers of multi-head self-attention. Vaswani et al. (2017) report that the Transformer base model (65M parameters) was trained on 8 NVIDIA P100 GPUs for 12 hours, and the Transformer big model (213M parameters) was trained for 3.5 days (84 hours; 300k steps). This model is also the basis for recent work on neural architecture search (NAS) for ma.chine translation and language modeling (So et al., 2019), and the NLP pipeline that we study in more detail in 4.2 (Strubell et al., 2018). So et al. (2019) report that their full architecture search ran for a total of 979M training steps, and that their base model requires 10 hours to train for 300k steps on one TPUv2 core. This equates to 32,623 hours of TPU or 274,120 hours on 8 P100 GPUs. ELMo. The ELMo model (Peters et al., 2018) is based on stacked LSTMs and provides rich word representations in context by pre-training on a large amount of data using a language model.ing objective. Replacing context-independent pre.trained word embeddings with ELMo has been shown to increase performance on downstream tasks such as named entity recognition, semantic role labeling, and coreference. Peters et al. (2018) report that ELMo was trained on 3 NVIDIA GTX 1080 GPUs for 2 weeks (336 hours). BERT. The BERT model (Devlin et al., 2019) provides a Transformer-based architecture for build.ing contextual representations similar to ELMo, but trained with a different language modeling objective. BERT substantially improves accuracy on tasks requiring sentence-level representations such as question answering and natural language inference. Devlin et al. (2019) report that the BERT base model (110M parameters) was trained on 16 TPU chips for 4 days (96 hours). NVIDIA reports that they can train a BERT model in 3.3 days (79.2 hours) using 4 DGX-2H servers, totaling 64 Tesla V100 GPUs (Forster et al., 2019). GPT-2. This model is the latest edition of OpenAI's GPT general-purpose token encoder, also based on Transformer-style self-attention and trained with a language modeling objective (Rad.ford et al., 2019). By training a very large model on massive data, Radford et al. (2019) show high zero-shot performance on question answering and language modeling benchmarks. The large model described in Radford et al. (2019) has 1542M parameters and is reported to require 1 week (168 hours) of training on 32 TPUv3 chips. 6 3 Related work There is some precedent for work characterizing the computational requirements of training and inference in modern neural network architectures in the computer vision community. Li et al. (2016) present a detailed study of the energy use required for training and inference in popular convolutional models for image classification in computer vision, including fine-grained analysis comparing different neural network layer types. Canziani et al. (2016) assess image classification model accuracy as a function of model size and gigaflops required during inference. They also measure average power draw required during inference on GPUs as a function of batch size. Neither work analyzes the recurrent and self-attention models that have become commonplace in NLP, nor do they extrapolate power to estimates of carbon and dol.lar cost of training. Analysis of hyperparameter tuning has been performed in the context of improved algorithms for hyperparameter search (Bergstra et al., 2011; Bergstra and Bengio, 2012; Snoek et al., 2012). To our knowledge there exists to date no analysis of the computation required for R&D and hyperparameter tuning of neural network models in NLP. 6Via the authors on Reddit. 7GPU lower bound computed using pre-emptible <> U.S. resources priced at <>, upper bound uses on-demand U.S. resources priced at <>. We similarly use pre-emptible (<>) and on-demand (<>) pricing as lower and upper bounds for TPU v2/3; cheaper bulk contracts are available. <> Table 3: Estimated cost of training a model in terms of CO2 emissions (lbs) and cloud compute cost (USD).7 Power and carbon footprint are omitted for TPUs due to lack of public information on power draw for this hardware. 4 Experimental results 4.1 Cost of training Table 3 lists CO2 emissions and estimated cost of training the models described in 2.1. Of note is that TPUs are more cost-efficient than GPUs on workloads that make sense for that hardware (e.g. BERT). We also see that models emit substantial carbon emissions; training BERT on GPU is roughly equivalent to a trans-American fight. So et al. (2019) report that NAS achieves a new state-of-the-art BLEU score of 29.7 for English to Ger.man machine translation, an increase of just 0.1 BLEU at the cost of at least $150k in on-demand compute time and non-trivial carbon emissions. 4.2 Cost of development: Case study To quantify the computational requirements of R&D for a new model we study the logs of all training required to develop Linguistically-Informed Self-Attention (Strubell et al., 2018), a multi-task model that performs part-of-speech tagging, labeled dependency parsing, predicate detection and semantic role labeling. This model makes for an interesting case study as a representative NLP pipeline and as a Best Long Paper at EMNLP. Model training associated with the project spanned a period of 172 days (approx. 6 months). During that time 123 small hyperparameter grid searches were performed, resulting in 4789 jobs in total. Jobs varied in length ranging from a minimum of 3 minutes, indicating a crash, to a maximum of 9 days, with an average job length of 52 hours. All training was done on a combination of NVIDIA Titan X (72%) and M40 (28%) GPUs.8 The sum GPU time required for the project totaled 9998 days (27 years). This averages to <> Table 4: Estimated cost in terms of cloud compute and electricity for training: (1) a single model (2) a single tune and (3) all models trained during R&D. about 60 GPUs running constantly throughout the 6 month duration of the project. Table 4 lists upper and lower bounds of the estimated cost in terms of Google Cloud compute and raw electricity re.quired to develop and deploy this model.9 We see that while training a single model is relatively inexpensive, the cost of tuning a model for a new dataset, which we estimate here to require 24 jobs, or performing the full R&D required to develop this model, quickly becomes extremely expensive. 5 Conclusions Authors should report training time and sensitivity to hyperparameters. Our experiments suggest that it would be beneficial to directly compare different models to per.form a cost-bene�t (accuracy) analysis. To ad.dress this, when proposing a model that is meant to be re-trained for downstream use, such as re.training on a new domain or fine-tuning on a new task, authors should report training time and computational resources required, as well as model sensitivity to hyperparameters. This will enable direct comparison across models, allowing subsequent consumers of these models to accurately assess whether the required computational resources We approximate cloud compute cost using P100 pricing. 9Based on average U.S cost of electricity of $0.12/kWh. are compatible with their setting. More explicit characterization of tuning time could also reveal inconsistencies in time spent tuning baseline models compared to proposed contributions. Realizing this will require: (1) a standard, hardware-independent measurement of training time, such as gigaflops required to convergence, and (2) a standard measurement of model sensitivity to data and hyperparameters, such as variance with respect to hyperparameters searched. Academic researchers need equitable access to computation resources. Recent advances in available compute come at a high price not attainable to all who desire access. Most of the models studied in this paper were developed outside academia; recent improvements in state-of-the-art accuracy are possible thanks to industry access to large-scale compute. Limiting this style of research to industry labs hurts the NLP research community in many ways. First, it stifles creativity. Researchers with good ideas but without access to large-scale compute will simply not be able to execute their ideas, instead constrained to focus on different problems. Second, it prohibits certain types of Research on the basis of access to financial resources. This even more deeply promotes the already problematic rich get richer cycle of research funding, where groups that are already successful and thus well-funded tend to receive more funding due to their existing accomplishments. Third, the prohibitive start-up cost of building in-house resources forces resource-poor groups to rely on cloud compute services such as AWS, Google Cloud and Microsoft Azure. While these services provide valuable, flexible, and often relatively environmentally friendly compute resources, it is more cost effective for academic researchers, who often work for nonprofit educational institutions and whose research is funded by government entities, to pool resources to build shared compute centers at the level of funding agencies, such as the U.S. National Science Foundation. For example, an off-the-shelf GPU server containing 8 NVIDIA 1080 Ti GPUs and supporting hardware can be purchased for approximately $20,000 USD. At that cost, the hardware required to develop the model in our case study (approximately 58 GPUs for 172 days) would cost $145,000 USD plus electricity, about half the estimated cost to use on-demand cloud GPUs. Unlike money spent on cloud compute, however, that invested in centralized resources would continue to pay off as resources are shared across many projects. A government-funded academic compute cloud would provide equitable access to all researchers. Researchers should prioritize computationally efficient hardware and algorithms. We recommend a concerted effort by industry and academia to promote research of more computationally efficient algorithms, as well as hardware that requires less energy. An effort can also be made in terms of software. There is already a precedent for NLP software packages prioritizing efficient models. An additional avenue through which NLP and machine learning software developers could aid in reducing the energy associated with model tuning is by providing easy.to-use APIs implementing more efficient alternatives to brute-force grid search for hyperparameter tuning, e.g. random or Bayesian hyperparameter search techniques (Bergstra et al., 2011; Bergstra and Bengio, 2012; Snoek et al., 2012). While software packages implementing these techniques do exist,10 they are rarely employed in practice for tuning NLP models. This is likely because their interoperability with popular deep learning frameworks such as PyTorch and TensorFlow is not optimized, i.e. there are not simple examples of how to tune TensorFlow Estimators using Bayesian search. Integrating these tools into the work�ows with which NLP researchers and practitioners are already familiar could have notable im.pact on the cost of developing and tuning in NLP. Acknowledgements We are grateful to Sherief Farouk and the anonymous reviewers for helpful feedback on earlier drafts. This work was supported in part by the Centers for Data Science and Intelligent Information Retrieval, the Chan-Zuckerberg Initiative under the Scientific Knowledge Base Construction project, the IBM Cognitive Horizons Network agreement no. W1668553, and National Science Foundation grant no. IIS-1514053. Any opinions, findings and conclusions or recommendations ex.pressed in this material are those of the authors and do not necessarily reflect those of the sponsor. For example, the Hyperopt Python library. References Rhonda Ascierto. 2018. Uptime Institute Global Data Center Survey. Technical report, Uptime Institute. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben.gio. 2015. Neural Machine Translation by Jointly Learning to Align and Translate. In 3rd Inter.national Conference for Learning Representations (ICLR), San Diego, California, USA. James Bergstra and Yoshua Bengio. 2012. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13(Feb):281�305. James S Bergstra, R�emi Bardenet, Yoshua Bengio, and Bal�azs K�egl. 2011. Algorithms for hyper-parameter optimization. In Advances in neural information processing systems, pages 2546�2554. Bruno Burger. 2019. Net Public Electricity Generation in Germany in 2018. Technical report, Fraunhofer Institute for Solar Energy Systems ISE. Alfredo Canziani, Adam Paszke, and Eugenio Culur.ciello. 2016. An analysis of deep neural network models for practical applications. Gary Cook, Jude Lee, Tamina Tsai, Ada Kongn, John Deans, Brian Johnson, Elizabeth Jardim, and Brian Johnson. 2017. Clicking Clean: Who is winning the race to build a green internet? Technical report, Greenpeace. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Un.derstanding. In NAACL. Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency pars.ing. In ICLR. EPA. 2018. Emissions & Generation Resource Inte.grated Database (eGRID). Technical report, U.S. Environmental Protection Agency. Christopher Forster, Thor Johnsen, Swetha Man.dava, Sharath Turuvekere Sreenivas, Deyu Fu, Julie Bernauer, Allison Gray, Sharan Chetlur, and Raul Puri. 2019. BERT Meets GPUs. Technical report, NVIDIA AI. Da Li, Xinbo Chen, Michela Becchi, and Ziliang Zong. 2016. Evaluating the energy ef�ciency of deep con.volutional neural networks on cpus and gpus. 2016 IEEE International Conferences on Big Data and Cloud Computing (BDCloud), Social Computing and Networking (SocialCom), Sustainable Comput.ing and Communications (SustainCom) (BDCloud.SocialCom-SustainCom), pages 477�484. Thang Luong, Hieu Pham, and Christopher D. Man.ning. 2015. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412�1421. Associa.tion for Computational Linguistics. Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word rep.resentations. In NAACL. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. Jasper Snoek, Hugo Larochelle, and Ryan P Adams. 2012. Practical bayesian optimization of machine learning algorithms. In Advances in neural informa.tion processing systems, pages 2951�2959. David R. So, Chen Liang, and Quoc V. Le. 2019. The evolved transformer. In Proceedings of the 36th International Conference on Machine Learning (ICML). Emma Strubell, Patrick Verga, Daniel Andor, David Weiss, and Andrew McCallum. 2018. Linguistically-Informed Self-Attention for Se.mantic Role Labeling. In Conference on Empir.ical Methods in Natural Language Processing (EMNLP), Brussels, Belgium. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In 31st Conference on Neural Information Processing Systems (NIPS). <> <> <> <> <> <> Finite-Element Neural Networks for Solving Differential Equations Pradeep Ramuhalli, Member, IEEE, Lalita Udpa, Senior Member, IEEE, and Satish S. Udpa, Fellow, IEEE Abstract The solution of partial differential equations (PDE) arises in a wide variety of engineering problems. Solutions to most practical problems use numerical analysis techniques such as finite-element or finite-difference methods. The drawbacks of these approaches include computational costs associated with the modeling of complex geometries. This paper proposes a finite-element neural network (FENN) obtained by embedding a finite-element model in a neural network architecture that enables fast and ac.curate solution of the forward problem. Results of applying the FENN to several simple electromagnetic forward and inverse problems are presented. Initial results indicate that the FENN performance as a forward model is comparable to that of the conventional finite-element method (FEM). The FENN can also be used in an iterative approach to solve inverse problems associated with the PDE. Results showing the ability of the FENN to solve the in.verse problem given the measured signal are also presented. The parallel nature of the FENN also makes it an attractive solution for parallel implementation in hardware and software. I. INTRODUCTION Solutions of differential equations arise in a wide variety of engineering applications in electromagnetics, signal processing, computational fluid dynamics, etc. These equations are typically solved using either analytical or numerical methods. Analytical solution methods are however feasible only for simple geometries, which limits their applicability. In most practical problems with complex boundary conditions, numerical analysis methods are required in order to obtain a reasonable solution. An example is the solution of Maxwell's equations in electromagnetics. Solutions to Maxwell's equations are used in a variety of applications for calculating the interaction of electromagnetic (EM) fields with different types of media. Very often, the solution to differential equations is necessary for solving the corresponding inverse problems. Inverse problems in general are ill-posed, lacking continuous dependence of the measurements on the input. This has resulted in the development of a variety of solution techniques ranging from simple calibration procedures to other direct (analytical) and iterative approaches [1]. Iterative methods typically employ a forward model that simulates the underlying physical process (Fig. 1) [2]. An initial estimate of the solution of the inverse problem (represented by in Fig. 1) is applied to the forward model, Manuscript received January 17, 2004; revised April 2, 2005. <

> Fig. 1. Iterative inversion method for solving inverse problems. resulting in the corresponding solution to the forward problem <> Although finite-element methods (FEMs) [3], [4] are extremely popular for solving differential equations, their major drawback is computational complexity. This problem becomes more acute when three-dimensional (3-D) finite-element models are used in an iterative algorithm for solving the inverse problem. Recently, several authors have suggested the use of neural networks (MLP or RBF networks [5]) for solving differential equations [6][9]. In these techniques, a neural network is trained using a large database containing the input data and the solution of the differential equation. The neural network during generalization learns the mapping corresponding to the PDE. Alternatively, in [10], the solution to a differential equation is written as a constant term, and an adjustable term with parameters that need to be determined. A neural network is used to determine the optimal values of the parameters. This approach is applicable only to problems with regular boundaries. An extension of the approach to problems with irregular boundaries is given in [11]. Other neural network based differential equation solvers use multilayer perceptron networks or variations on the MLP to approximate the unknown function in a PDE [12][14]. A combination of the PDE and boundary conditions is used to construct an objective function that is minimized during the training process. A major limitation of these approaches is that the network architecture is selected somewhat arbitrarily. A second drawback is that the performance of the neural networks depends on the data used in training and testing. As long the test data is similar to the training data, the network can interpolate between the training data points to obtain a reasonable prediction. However, when the test signal is no longer similar to the training data, the network is forced to extrapolate and the performance degrades. One way around this difficulty is to ensure that the training data.base has a diverse set of signals. However, this is difficult to ensure in practice. Alternatively, we have to design neural net.works that are capable of extrapolation. Extrapolation methods are discussed extensively in literature [15][18], but the design of an extrapolation neural network involves several issues particularly for ensuring that the error in the network prediction stays within reasonable bounds during the extrapolation procedure. An ideal solution to this problem would be to combine the power of numerical models with the computational speed of neural networks, i.e., to embed a numerical model in a neural network structure. One such finite-element neural network (FENN) formulation has been reported by Takeuchi and Kosugi [19]. This approach, based on error minimization, derives the neural network using the energy functional resulting from the finite-element formulation. Other reports of FENN combinations are either similar to the Takeuchi method [20], [21] or use Hopfield neural networks to solve the forward problem [22], [23]. Kalkkuhl et al. [24] provide a description of a FEM-based approach to NARX modeling that may be interpreted both as a local model network, as well as a single layer feedforward network. A slightly different approach to merging numerical methods and neural networks is given in [25], where the finite-difference time domain (FDTD) method is cast in a neural network framework for the purpose of solving electromagnetic forward problems. The related problem of mesh generation in finite-element models has also been tackled using neural networks (for instance, [26]). Generally, these networks are designed to solve the forward problem, and must be modified to solve inverse problems. This paper proposes a new approach that embeds a finite-element model commonly used in the solution of differential equations in a neural network. The network, called the FENN, can solve the forward problem and can also be used in an iterative algorithm to solve inverse problems. The primary advantage of this approach is that the FEM is represented in a parallel form. Thus, it has the potential to alleviate the computational cost associated with using the FEM in an iterative algorithm for solving inverse problems. More importantly, the FENN does not need any training, and the computation of the weights is a one-time process. The proposed approach is also different in that the neural network architecture developed can be used to solve the forward and inverse problems. The structure of the neural network is also simpler than those reported in the literature, making it easier to implement in parallel in both hardware and software. The rest of this paper is organized as follows. Section II briefly describes the FEM, and derives the proposed FENN. In this paper, we focus on the problem of solving typical equations encountered in electromagnetic nondestructive evaluation (NDE). However, the same concepts can be easily applied to solve differential equations encountered in other fields. Sections III, IV and V present the application of the FENN to solving forward and inverse problems, along with initial results. A discussion of the advantages and disadvantages of the proposed FENN architecture is given in Section IV. Finally, Section V draws conclusions from the results and presents ideas for future work. II. THE FENN This section briefly describes the FEM and proposes its reformulation into a parallel neural network structure. Details about the FEM can be found in [3] and [4]. A. The FEM Consider a typical boundary value problem with the governing differential equation <> (1) where <> is a differential operator, <> is the applied source or forcing function, and is the unknown quantity. This differential equation can be solved in conjunction with boundary conditions on the boundary enclosing the domain The variational formulation used in finite-element analysis determines the unknown by minimizing the functional [3], [4] (2) with respect to the trial function The minimization procedure starts by dividing into small subdomains called elements (Fig. 2) and representing in each element by means of basis functions defined over the element (3) where is the unknown solution in element <> (3) is the basis function associated with node in element , is the value of the unknown quantity at node and is the total number of nodes associated with element <> In general, the basis functions (also referred to as interpolation functions or shape functions) can be linear, quadratic, or of higher order. Typically, finite-element models use either linear or polynomial spline basis functions. The functional within an element is expressed as <> (4) By substituting (3) in (4), we obtain the discrete version of the functional within each element <> (5) where is the transpose of a matrix, mental matrix with elements is the ele. <> (6) and is an vector with elements <> (7) Combining the values in (5) for each of the elements (8) where is the global matrix derived from the terms of the elemental matrices for different elements, and is the total number of nodes, also called the stiffness matrix, is a sparse, banded matrix. Equation (8) is the discrete version of the functional and can be minimized with respect to the nodal parameters by taking the derivative of with respect to <> and setting it equal to zero, which results in the matrix equation <> (9) Boundary conditions for these problems are usually of two types: natural boundary conditions and essential boundary conditions. Essential boundary conditions (also referred to as Dirichlet boundary conditions) impose constraints on the value of the unknown at several nodes. Natural boundary conditions (of which Neumann boundary conditions are a special case) impose constraints on the change in across a boundary. Dirichlet boundary conditions are imposed on the functional minimization (9), by deleting the rows and columns of the matrix corresponding to the nodes on the Dirichlet boundary and modifying in (9). Natural boundary conditions are applied in the FEM by adding an additional term to the functional. These boundary conditions are then incorporated into the functional and are satisfied automatically during the solution procedure. As an example, consider the natural boundary condition represented by the following equation [3] on <> (10) where <> represents the Neumann boundary, is its outward normal unit vector, is some constant, and , <>, and are known parameters associated with the boundary. Assuming that the boundary is made up of segments, we can define boundary matrices and with elements <> (11) where <>are basis functions defined over segment and is the length of the segment. The elements of <> are added to the elements of that correspond to the nodes on the boundary. Similarly, the elements of <> are added to the corresponding elements of <> The global matrix (9) is thus modified as follows before solving for <> (12) <

> Fig. 3. FEM domain discretization using two elements and four nodes. This process ensures that natural boundary conditions are implicitly and automatically satisfied during the FEM solution procedure. B. The FENN This section describes how the finite-element model can be converted into a parallel network form. We focus on solving typical inverse problems arising in electromagnetic NDE, but the basic idea is applicable to other areas as well. NDE inverse problems can be formulated as the problem of finding the material properties (such as the conductivity or the permeability) within the domain of the problem. Since the domain is discretized in the FEM method by a large number of elements, the problem can be posed as one of finding the material properties in each of these elements. These properties are usually embedded in the differential operator <> or equivalently, in the global matrix <> Thus, in order to be able to iteratively estimate these properties from the measurements, the material properties need to be separated out from <> This separation is easier to achieve at the element matrix level. For nodes <> and in element > (13) where <> is the parameter representing the material property in element <> and <> represents the differential operator at the <

> Fig. 4. FENN. element level without embedded in it. Substituting (13) into the functional, we get <> (14) If we define <> (15) where <> (16) <> (17) Equation (17) expresses the functional explicitly in terms of <> The assumption that is constant within each element is implicit in this expression. This assumption is usually satisfied in problems in NDE where each element in the FEM mesh is defined within the confines of a domain, and at no time does a single element cross domain boundaries. Furthermore, each element is small enough that minor variations in within an element may be ignored. Equation (17) can be easily converted into a parallel network form. The neural network comprises an input, output and hidden layer. In the general case with <> elements and <> nodes in the FEM mesh, the input layer with network inputs takes the values in each element as input. The hidden layer has neurons arranged in groups of neurons, corresponding to the members of the global <> matrix . The output of each group of hidden layer neurons is the corresponding row vector of . The weights from the input to the hidden layer are set to the appropriate values of . Each neuron in the hidden layer acts as a summation unit, (equivalent to a summation followed by a linear activation function [5]). The outputs of the hidden layer neurons are the elements of the global matrix as given in (15). Each group of hidden neurons is connected to one output neuron (giving a total of output neurons) by a set of weights with each element of representing the nodal values. Note that the set of weights between the first group of hidden neurons and the first output neuron are the same as the set of weights between the second group of hidden neurons and the second output neuron (as well as between successive groups of hidden neurons and the corresponding output neuron). Each output neuron is also a summation unit followed by a linear activation function, and the output of each neuron is equal to <> (18) where the second part of (18) is obtained by using (15). As an example, the FENN architecture for a two-element, four-node FEM mesh (Fig. 3) is shown in Fig. 4. In this case, the FENN has two input neurons, 16 hidden layer neurons and four output neurons. The �gure illustrates the grouping of the hidden layer neurons, as well as the similarity inherent in the weights that connect each group of hidden layer neurons to the corresponding output neuron. To simplify the �gure, the weights between the network input and hidden layer neurons are depicted by means of vectors (for , 2, 3, 4 and , 2), where the individual weight values <> are defined as in (16). 1) Boundary Conditions in the FENN: Note that the elements of <> and in (11) do not depend on the material properties <> and need to be added appropriately to the global matrix and the source vector as shown in (12). <

> Fig. 5. Geometry of mesh for 1-D FEM. <

> Fig. 6. Flowchart (with example) for designing the FENN for a general PDE. Equation (12) thus implies that natural boundary conditions can be ap-layer neurons. These weights will be referred to as the clamped plied in the FENN as bias inputs to the hidden layer neurons weights, while the remaining weights will be referred to as the that are a part of the boundary, and the corresponding output free weights. An example of these weights is presented later. neurons. Dirichlet boundary conditions are applied by clamping The FENN architecture was derived without consideration of the corresponding weights between the hidden layer and output the dimensionality of the problem at hand, and thus can be used for 1-, 2-, 3-, or higher dimensional problems. The number of nodes and elements in the FEM mesh dictates the number of neurons in the different layers. The weights between the input and hidden layer change depending on node-element connectivity information. The major drawback of the FENN is the number of neurons and weights necessary. However, the memory requirements can be reduced considerably, since most of the weights between the input and hidden layer are zero. These weights, and the corresponding connections, can be discarded. Similarly, most of the elements of the matrix are also zero (is a banded matrix). The corresponding neurons in the hidden layer can also be discarded, reducing memory and computation requirements considerably. Furthermore, the weights between each group of hidden layer neurons and the output layer are the same . Weight-sharing approaches can be used here to further reduce the storage requirements. C. A 1-D Example Consider the 1-D equation <> (19) on the boundary <> defined by <> and are constants depending on the material and is the applied source. Laplace's equation and Poisson's equation are special cases of this equation. The FENN formulation for this problem starts by discretizing the domain of interest with <> elements and nodes. In one dimension, each element is defined by two nodes (Fig. 5). define basis functions <> and <> over each element <> and let is the value of <> on node <> in element <> An example of the basis functions is shown in Fig. 5. For these basis functions, i.e., <> (20) the element matrices are given by [3] <> (21) <> (22) Here, <> is the length of element <> The global matrix is then constructed by selectively adding the element matrices based on the nodes that form an element. Specifically, is a sparse tridiagonal matrix, and its nonzero elements are given by <> (23) Fig. 7. Shielded microstrip geometry. (a) Complete problem description. (b) Problem description using symmetry considerations. The network implementation of (23) can be derived as fol.lows. If <> and <> values at each element are the inputs to the network, <> and <> form the weights between the input and hidden layers. The network thus uses input neurons and hidden neurons. The values of <> at each of the nodes are assigned as weights between the hidden and output layers, and the source is the desired output of this network (corresponding to the output neurons). Dirichlet boundary conditions on are applied as explained earlier. D. General Case Fig. 6 shows a flowchart of the general scheme for converting a differential equation into the FENN structure. An example in two dimensions is also provided next to the flowchart. We start with the differential equation and the boundary conditions and formulate the FEM using the variational method. This in.volves discretizing the domain of interest with elements and nodes, selecting basis functions, writing the functional for each element and obtaining the element matrices and the source vector. The example presented uses the FEM mesh shown in Fig. 3, with elements, and <> nodes, and linear basis functions. The unknown solution to the differential equation is represented by its values at each of the nodes in the finite-element mesh <> The element matrices are then separated into two parts, with one part dependent on the material properties <> and while the other is independent of them. The FENN is then designed to have input neurons, hidden neurons, and output neurons, where <> is the number of material property parameters. In the example under consideration, <>, since we have two material property parameters ( and ). The first group of input neurons takes in the values while the second group takes in the values in each element. The weights from the input to the hidden layer are set to the appropriate values of <> In the example, since nodes 1, 2, and 3 are part of element 1 (see Fig. 3), the weights from the first input node to the first group of four neurons in the hidden layer are given by <> (24) The last weight is zero since node 4 is not a part of element 1. Each group of hidden neurons is connected to one output neuron (giving a total of output neurons) by a set of weights <> with each element of representing the nodal values. The output of each neuron in the output layer is equal to <

> Fig. 8. Forward problem solutions for shielded microstrip problem show the contours of constant potential for: (a) FEM solution and (b) FENN solution. (c) Error between (a) and (b). The x-and y-axes show the nodes in the FEM discretization of the domain, and the z-axis in (c) shows the error at each of these nodes in volts. III. FORWARD AND INVERSE PROBLEM FORMULATION USING FENN. where is the output of the FENN based approach, then, for the gradients of the error with respect to the free hidden layer weights is given by the FENN architecture and algorithm lends itself to solving <> (27) both the forward and inverse problems. The forward problem involves determining the weights given the material parameters Equation (27) can be used to solve the forward problem. Similarly, the applied source to solve the inverse problem, while the inverse problem the gradients of the error involves determining and (input of the FENN) are necessary, and approach can be used to solve both these problems. Suppose we are given by define the error at the output of the FENN as <

> TABLE I SUMMARY OF PERFORMANCE OF THE FENN ALGORITHM FOR VARIOUS PDES For the forward problem, such an approach is equivalent to the iterative approaches used to solve for the unknown nodal values in the FEM [4]. IV. RESULTS A. Forward Model Results The FENN was tested using both 1-and 2-D versions of Poisson�s equation <> (30) where represents the material property, and is the applied source. For instance, in electromagnetics may represent the permittivity while represents the charge density. As the first example, consider the following 2-D equation <> (31) with boundary conditions and <> on <> (32) on <> (33) This is the governing equation for the shielded microstrip trans.mission line problem shown in Fig. 7. The forward problem computes the electric potential due to the shielded microstrip shown in Fig. 7(a). The potentials are zero on the shielding con.ductor. Since the geometry is symmetric, we can solve the equiv.alent problem shown in Fig. 7(b), by applying the homogeneous Neumann condition on the plane of symmetry. The inner con.ductor (microstrip) is held at a constant potential of volts. Finally, we also assume that the material inside the shielding conductor has a permittivity , where K is a constant. The permittivity in this case corresponds to the material property . Specifically, and . The homogeneous Neu.mann boundary condition is equivalent to setting . The microstrip and the shielding conductor correspond to the Dirichlet boundary, with <> on the microstrip and on the outer boundary [Fig. 7(b)]. Finally, there is no source term in this example (the source term would correspond to a charge distribution in the domain of interest), i.e., <> In this ex.ample, we assume that volts. Further, we assume that the domain of interest is The solution to the forward problem is presented in Fig. 8, with the FEM solution using 11 nodes in each direction shown in Fig. 8(a) and the corresponding FENN solution in Fig. 8(b). These �gures show contours of constant potential. The error be.tween the FEM and FENN solutions is presented in Fig. 8(c). As seen from the �gure, the FENN is seen to match the FEM solu.tion accurately, with the peak error at any node on the order of Several other examples were also used to test the FENN and the results are summarized in Table I. Column 1 shows the PDE used to evaluate the FENN performance, while column 2 shows the boundary conditions used. The analytic solution to the problem is indicated in Column 3. The FENN structure and the number of iterations for convergence using a gradient de.scent approach are indicated in Columns 4 and 5, respectively. The FENN structure, as explained earlier, has an are the number of elements and nodes in the FEM mesh, respectively, and is the number of hidden neurons, and corresponds to the number of nonzero elements in the FEM global matrix Finally, Columns 6 and 7 present the sum-squared error (SSE) and the maximum error in the solution, respectively, where the errors are computed with respect to the analytical solution. These results indicate that the FENN is capable of accurately deter.mining the potential One advantage of the FENN approach is that the computation of the input-hidden layer weights is a one-time process, as long as the differential equation does not change. The only changes necessary to solve the different problems are changes in the input and the desired output. B. Inverse Model Results The FENN was also used to solve several simple inverse problems based on (30). In all cases, the objective was to determine <

> Fig. 9. FENN inversion results for Poisson's equation with initial solutions (a) the value of <> and <> for given values of <> and The <> first example is a 1-D problem that involves determining given and <> for the differential equation <> (34) with boundary conditions <> and <>. The analytical solution to this inverse problem is <> and <> (35) As seen from (35), the problem has an infinite number of solutions and we expect the solution procedure to converge to one of these solutions depending on the initial value. Fig. 9(a) and (b) shows two solutions to this inverse problem for two different initializations (shown using triangles). In both cases, the FENN solution (in stars) is seen to match the analytical solution (squares). The SSE in both cases was on the order of <> In order to obtain a unique solution, we need to constrain the value of at the boundary as well. Consider the same differen. tial equation as (34), but with and specified as follows: and (36) The analytical solution for this equation is .To solve this problem, we set and clamp the value of at and as follows: , . The results of the constrained inversion obtained using 11 nodes and 10 elements in the corresponding finite-element mesh are shown in Fig. 10. Fig. 10(a) shows the comparison between the analytical solution (solid line with squares) and the FENN result (solid line with stars). The initial value of is shown in the figure as a dashed line. Fig. 10(b) shows the comparison between the actual and desired forcing function at the FENN output. This result indicates that the SSE in the forcing function, as well as the SSE in the inversion result, is fairly large (0.0148 and 0.0197, respectively). The reason for this was traced back to the mesh discretization. Fig. 11 shows the SSE in the output of the FENN and the SSE in the inverse problem solution as a function of FEM discretization. It is seen that increasing the discretization significantly improves the solution. Similar results were observed for other problems. V. DISCUSSION AND CONCLUSION The FENN is closely related to the finite-element model used to solve differential equations. The FENN architecture has a weight structure that allows both the forward and inverse problems to be solved using simple gradient-based algorithms. Initial results indicate that the proposed FENN algorithm is capable of accurately solving both the forward and inverse problems. In addition, the forward problem solution from the FENN is seen to exactly match the FEM solution, indicating that the FENN represents the finite-element model exactly in a parallel configuration. The major advantage of the FENN is that it represents the finite-element model in a parallel form, enabling parallel implementation in either hardware or software. Further, computing gradients in the FENN is very simple. This is an advantage in solving both forward and inverse problems using gradient-based methods. The gradients can also be computed in parallel and the lack of nonlinearities in the neuron activation functions makes the computation of gradients simpler. A major advantage of this approach for solving inverse problems is that it avoids inverting the global matrix in each iteration. The FENN also does not require any training, since most of its weights can be computed in advance and stored. The weights depend on the governing differential equation and its associated boundary conditions, and as long as these two factors do not change, the weights do not change. This is especially an advantage in solving inverse problems in electromagnetic NDE. This approach also reduces the computational effort associated with the network. Future work will concentrate on applying the FENN to 3-D electromagnetic NDE problems. The robustness of the approach will also be tested, since the ability of these approaches to in.vert practical noisy measurements is important. Furthermore, the use of better optimization algorithms, like conjugate gradient methods, is expected to improve the solution speed. In addition, parallel implementation of the FENN in both hardware and software is under investigation. The approach described in this paper is very general in that it can be applied to a variety of inverse problems in fields other than electromagnetic NDE. Some of these other applications will also be investigated to show the general nature of the proposed method. REFERENCES [1] L. Udpa and S. S. Udpa, �Application of signal processing and pattern recognition techniques to inverse problems in NDE,� Int. J. Appl. Elec.tromagn. Mechan., vol. 8, pp. 99�117, 1997. [2] M. Yan, M. Afzal, S. Udpa, S. Mandayam, Y. Sun, L. Udpa, and P. Sacks, �Iterative algorithms for electromagnetic NDE signal inversion,� in ENDE �97, Reggio Calabria, Italy, Sep. 14�16, 1997. [3] J. Jin, The Finite Element Method in Electromagnetics. New York: Wiley, 1993. [4] P. Zhou, Numerical Analysis of Electromagnetic Fields. Berlin, Ger.many: Springer-Verlag, 1993. [5] S. Haykin, Neural Networks: A Comprehensive Foundation. Upper Saddle River, NJ: Prentice-Hall, 1994. [6] C. A. Jensen et al., �Inversion of feedforward neural networks: algo.rithms and applications,� Proc. IEEE, vol. 87, no. 9, pp. 1536�1549, 1999. [7] P. Ramuhalli, L. Udpa, and S. Udpa, �Neural network algorithm for elec.tromagnetic NDE signal inversion,� in ENDE 2000, Budapest, Hungary, Jun. 2000. [8] C. H. Barbosa, A. C. Bruno, M. Vellasco, M. Pacheco, J. P. Wikswo Jr., and A. P. Ewing, �Automation of SQUID nondestructive evaluation of steel plates by neural networks,� IEEE Trans. Appl. Supercond., vol. 9, no. 2, pp. 3475�3478, 1999. [9] W. Qing, S. Xueqin, Y. Qingxin, and Y. Weili, �Using wavelet neural net.works for the optimal design of electromagnetic devices,� IEEE Trans. Magn., vol. 33, no. 2, pp. 1928�1930, 1997. [10] I. E. Lagaris, A. C. Likas, and D. I. Fotiadis, �Arti�cial neural networks for solving ordinary and partial differential equations,� IEEE Trans. Neural Netw., vol. 9, no. 5, pp. 987�1000, 1998. [11] I. E. Lagaris, A. C. Likas, and D. G. Papageorgiou, �Neural-network methods for boundary value problems with irregular boundaries,� IEEE Trans. Neural Netw., vol. 11, no. 5, pp. 1041�1049, 2000. [12] B. P. Van Milligen, V. Tribaldos, and J. A. Jimenez, �Neural network differential equation and plasma equilibrium solver,� Phys. Rev. Lett., vol. 75, no. 20, pp. 3594�3597, 1995. [13] M. W. M. G. Dissanayake and N. Phan-Thien, �Neural-network-based approximations for solving partial differential equations,� Commun. Numer. Meth. Eng., vol. 10, pp. 195�201, 1994. [14] R. Masuoka, �Neural networks learning differential data,� IEICE Trans. Inform. Syst., vol. E83-D, no. 6, pp. 1291�1300, 2000. [15] D. C. Youla, �Generalized image restoration by the method of alternating orthogonal projections,� IEEE Trans. Circuits Syst., vol. CAS-25, no. 9, pp. 694�702, 1978. [16] D. C. Youla and H. Webb, �Image restoration by the method of convex projections: part I�theory,� IEEE Trans. Med. Imag., vol. MI-1, no. 2, pp. 81�94, 1982. [17] A. Lent and H. Tuy, �An iterative method for the extrapolation of band-limited functions,� J. Math. Analysis and Applicat., vol. 83, pp. 554�565, 1981. [18] W. Chen, �A new extrapolation algorithm for band-limited signals using the regularization method,� IEEE Trans. Signal Process., vol. 41, no. 3, pp. 1048�1060, 1993. [19] J. Takeuchi and Y. Kosugi, �Neural network representation of the finite element method,� Neural Netw., vol. 7, no. 2, pp. 389�395, 1994. [20] R. Sikora, J. Sikora, E. Cardelli, and T. Chady, �Arti�cial neural net.work application for material evaluation by electromagnetic methods,� in Proc. Int. Joint Conf. Neural Networks, vol. 6, 1999, pp. 4027�4032. [21] G. Xu, G. Littlefair, R. Penson, and R. Callan, �Application of FE-based neural networks to dynamic problems,� in Proc. Int. Conf. Neural Infor.mation Processing, vol. 3, 1999, pp. 1039�1044. [22] F. Guo, P. Zhang, F. Wang, X. Ma, and G. Qiu, �Finite element anal.ysis-based Hop�eld neural network model for solving nonlinear elec.tromagnetic �eld problems,� in Proc. Int. Joint Conf. Neural Networks, vol. 6, 1999, pp. 4399�4403. [23] H. Lee and I. S. Kang, �Neural algorithm for solving differential equations,� J. Computat. Phys., vol. 91, pp. 110�131, 1990. [24] J. Kalkkuhl, K. J. Hunt, and H. Fritz, �FEM-based neural-network approach to nonlinear modeling with application to longitudinal vehicle dynamics control,� IEEE Trans. Neural Netw., vol. 10, no. 4, pp. 885�897, 1999. [25] R. K. Mishra and P. S. Hall, �NFDTD concept,� IEEE Trans. Neural Netw., vol. 16, no. 2, pp. 484�490, 2005. [26] D. G. Triantafyllidis and D. P. Labridis, �A finite-element mesh gener.ator based on growing neural networks,� IEEE Trans. Neural Netw., vol. 13, no. 6, pp. 1482�1496, 2002. <> <> <> <> <> <> Floating Point Operations in Matrix-Vector Calculus (Version 1.3) Raphael Hunger Technical Report 2007 Technische Universit�t Mchen Associate Institute for Signal Processing Univ.-Prof. Dr.-Ing. Wolfgang Utschick History Version 1.00: October 2005 -Initial version Version 1.01: 2006 -Rewrite of sesquilinear form with a reduced amount of FLOPs -Several Typos fixed concerning the number of FLOPS required for the Cholesky decomposition Version 1.2: November 2006 -Conditions for the existence of the standard <> Cholesky decomposition specified (positive definiteness) -Outer product version of <> Cholesky decomposition removed -FLOPs required in Gaxpy version of <> Cholesky decomposition updated -<> Cholesky decomposition added -Matrix-matrix product LC added with L triangular -Matrix-matrix product <>C added with L triangular and <> not known a priori -Inverse L. 11 of a lower triangular matrix with ones on the main diagonal added Version 1.3: September 2007 -First globally accessible document version ToDo: (unknown when) -QR-Decomposition -LR-Decomposition Please report any bug and suggestion to hunger@tum.de Contents 1. Introduction 4 2. Flop Counting 5 2.1 MatrixProducts .................................... 5 2.1.1 Scalar-Vector Multiplication .a ....................... 5 2.1.2 Scalar-Matrix Multiplication .A ...................... 5 2.1.3 Inner Product aHb ofTwo Vectors ...................... 5 2.1.4 Outer Product ac H ofTwo Vectors ...................... 5 2.1.5 Matrix-Vector Product Ab .......................... 6 2.1.6 Matrix-Matrix Product AC ......................... 6 2.1.7 Matrix Diagonal Matrix Product AD .................... 6 2.1.8 Matrix-Matrix Product LD ......................... 6 2.1.9 Matrix-Matrix Product L1D ......................... 6 2.1.10 Matrix-Matrix Product LC with L Lower Triangular ............ 6 2.1.11 Gram AHA of A ............................... 6 2.1.12 Squared Frobenius Norm kAk2F = tr(AHA) ................ 7 2.1.13 Sesquilinear Form cHAb ........................... 7 2.1.14 Hermitian Form aHRa ............................ 7 2.1.15 Gram LHL of a Lower Triangular Matrix L ................. 7 2.2 Decompositions.................................... 8 2.2.1 Cholesky Decomposition R = <> (GaxpyVersion) ........... 8 2.2.2 Cholesky Decomposition R = L1DL1H ................... 10 2.3 Inverses ofMatrices .................................. 11 2.3.1 Inverse <> of a Lower Triangular Matrix L ................ 11 2.3.2 Inverse L. 11 of a Lower Triangular Matrix L1 with Ones on the Main Diagonal..................................... 12 2.3.3 Inverse R.1 of a Positive definite Matrix R ................. 13 2.4 Solving Systems of Equations ............................ 13 2.4.1 Product <>C with <> not known a priori. ................ 13 3. Overview 14 Appendix 15 Bibliography 16 1. Introduction For the design of efficient und low-complexity algorithms in many signal-processing tasks, a de.tailed analysis of the required number of floating-point operations (FLOPs) is often inevitable. Most frequently, matrix operations are involved, such as matrix-matrix products and inverses of matrices. Structures like Hermiteness or triangularity for example can be exploited to reduce the number of needed FLOPs and will be discussed here. In this technical report, we derive expressions for the number of multiplications and summations that a majority of signal processing algorithms in mobile communications bring with them. Acknowledgments: The author would like to thank Dipl.-Ing. David A. Schmidt and Dipl.-Ing. Guido Dietl for the fruitful discussions on this topic. 2. Flop Counting In this chapter, we offer expressions for the number of complex multiplications and summations required for several matrix-vector operations. A floating-point operation (FLOP) is assumed to be either a complex multiplication or a complex summation here, despite the fact that a complex multiplication requires 4 real multiplications and 2 real summations whereas a complex summations consists of only 2 real summations, making a multiplication more expensive than a summation. However, we count each operation as one FLOP. Throughout this report, we assume <> to be a scalar, the vectors <>, and <> to have dimension N, N, and M, respectively. The matrices <>, and <> are assumed to have no special structure, whereas <> is Hermitian and <> is diagonal. L is a lower triangular <> matrix, en denotes the unit vector with a 1 in the n-th row and zeros elsewhere. Its dimensionality is chosen such that the respective matrix-vector product exists. Finally, [A]a,b denotes the element in the a-th row and b-th column of A, <> selects the submatrix of A consisting of rows a to b and columns c to d. 0a.b is the a . b zero matrix. Transposition, Hermitian transposition, conjugate, and real-part operator are denoted by <>, and <>, respectively, and require no FLOP. 2.1 Matrix Products Frequently arising matrix products and the amount of FLOPs required for their computation will be discussed in this section. 2.1.1 Scalar-Vector Multiplication <> A simple multiplication .a of a vector a with a scalar <> requires N multiplications and no summation. 2.1.2 Scalar-Matrix Multiplication <> Extending the result from Subsection 2.1.1 to a scalar matrix multiplication <> requires NM multiplications and again no summation. 2.1.3 Inner Product aHb of Two Vectors An inner product aHb requires N multiplications and <> summations, i.e., <> FLOPs. 2.1.4 Outer Product <> of Two Vectors An outer product acH requires NM multiplications and no summation. 2. Flop Counting 2.1.5 Matrix-Vector Product <> Computing Ab corresponds to applying the inner product rule <> from Subsection 2.1.3 M times. Obviously, <> and <> represents the i-th row of A. Hence, its computation costs MN multiplications and <> summations, i.e., <> FLOPs. 2.1.6 Matrix-Matrix Product <> Repeated application of the matrix-vector rule Aci from Subsection 2.1.5 with ci being the i-th column of C yields the overall matrix-matrix product AC. Since <>, the matrix-matrix product has the L-fold complexity of the matrix-vector product. Thus, it needs MNL multiplications and summations, altogether <> FLOPs. 2.1.7 Matrix Diagonal Matrix Product AD If the right hand side matrix D of the matrix product AD is diagonal, the computational load reduces to M multiplications for each of the N columns of A, since the n-th column of A is scaled by the n-th main diagonal element of D. Thus, MN multiplications in total are required for the computation of AD, no summations are needed. 2.1.8 Matrix-Matrix Product LD When multiplying a lower triangular matrix L by a diagonal matrix D, column n of the matrix product requires <> multiplications and no summations. With <>, we get <> multiplications. 2.1.9 Matrix-Matrix Product L1D When multiplying a lower triangular matrix L1 with ones on the main diagonal by a diagonal matrix D, column n of the matrix product requires <<<>>> multiplications and no summations. With <>, we get <> multiplications. 2.1.10 Matrix-Matrix Product LC with L Lower Triangular Computing the product of a lower triangular matrix <> and <> is done column-wise. The nth element in each column of LC requires n multiplications and <<<>>> summations, so the complete column needs <> multiplications and <> summations. The complete matrix-matrix product is obtained from computing L columns. We have <> multiplications and <> summations, yielding a total amount of <> FLOPs. 2.1.11 Gram <> of A In contrast to the general matrix product from Subsection 2.1.6, we can make use of the Hermitian structure of the product <>. Hence, the strictly lower triangular part of <> need not be computed, since it corresponds to the Hermitian of the strictly upper triangular part. For this reason, we have to compute only the N main diagonal entries of <> and the <> upper <> off-diagonal elements, so only <> different entries have to be evaluated. Each element requires an inner product step from Subsection 2.1.3 costing M multiplications and <> summations. Therefore, <> multiplications and <> summations are needed, making up a total amount of <> FLOPs. 2.1 Matrix Products 2.1.12 Squared Frobenius Norm <> The squared Hilbert-Schmidt norm <> follows from summing up the MN squared entries from A. We therefore have MN multiplications and <> summations, yielding a total of <> FLOPs. 2.1.13 Sesquilinear Form <> The sesquilinear form cHAb should be evaluated by computing the matrix-vector product Ab in a first step and then multiplying with the row vector cH from the left hand side. The matrix vector product requires MN multiplications and <> summations, whereas the inner product needs M multiplications and <> summations. Altogether, <> multiplications and <> summations have to be computed for the sesquilinear form <>, yielding a total number of <> flops. 2.1.14 Hermitian Form a <> With the Hermitian matrix <>, the product <> can be expressed as <> with <>, and <>. The first sum accumulates the weighted main diagonal entries and requires 2N multiplications and <> summations. The second part of (2.1) accumulates all weighted off-diagonal entries from A. The last two summations sum up 2 terms2. Consequently, the second part of (2.1) requires <> summations and <> products. Finally, the two parts have to be added accounting for an additional summation and yielding an overall amount of <> products and <> summations, corresponding to <> FLOPs. 2.1.15 Gram <> of a Lower Triangular Matrix L During the computation of the inverse of a positive definite matrix, the Gram matrix of a lower triangular matrix occurs when Cholesky decomposition is applied. Again, we make use of the Hermitian structure of the Gram <>, so only the main diagonal entries and the upper right off-diagonal entries of the product have to be evaluated. The a-th main-diagonal entry can be expressed >. We made use of (A1) in the Appendix for the computation of the last sum accumulating subsequent integers. We do not exploit the fact that only real-valued summands are accumulated as we only account for complex flops. The scaling with the factor 2 does not require a FLOP, as it can be implemented by a simple bit shift. Clearly, if <>, we have to subtract one summation from the calculation since no off-diagonal entries exist. 2. Flop Counting <> (2.2) with <>, requiring <> multiplications and <> summations. Hence, all main diagonal elements need <> multiplications and <> summations. The upper right off-diagonal entry <> in row a and column b with <> reads as <>, (2.3) again accounting for <> multiplications and <> summations. These two expressions have to be summed up over all <> and <>, and for the number of multiplications, we find <> (2.4) Again, we made use of (A1) for the sum of subsequent integers and (A2) for the sum of subsequent squared integers. For the number of summations, we evaluate <> Computing all necessary elements of the Gram LHL thereby requires <> multiplications and <> summations. Altogether, <> FLOPs result. The same result of course holds for the Gram of two upper triangular matrices. 2.2 Decompositions 2.2.1 Cholesky Decomposition <> (Gaxpy Version) Instead of computing the inverse of a positive definite matrix R directly, it is more efficient to start with the Cholesky decomposition <> and then invert the lower triangular matrix L and compute its Gram. In this section, we count the number of FLOPs necessary for the Cholesky decomposition. 2.2 Decompositions The implementation of the Generalized Ax plus y (Gaxpy) version of the Cholesky decomposition, which overwrites the lower triangular part of the positive definite matrix R is listed in Algorithm 2.1, see [1]. Note that R needs to be positive definite for the <> decomposition! Algorithm 2.1 Algorithm for the Gaxpy version of the Cholesky decomposition. <> The computation of the first column of L in Line 1 of Algorithm 2.1 requires <> multiplications, a single square-root operation, and no summations. Column <> takes a matrix vector product of dimension <> which is subtracted from another <> dimensional vector involving <> summations, see Line 3. Finally, < multiplications6 and a single square-root operation are necessary in Line 4. In short, row n with <> needs <> multiplications, .<> summations (see Subsection 2.1.5), and one square root operation, which we classify as an additional FLOP. Summing up the multiplications for rows <>, we obtain <> The number of summations for rows <> reads as <> (2.6) <> (2.7) The first element need not be computed twice, since the result of the division is the square root of the denominator. Again, the first element need not be computed twice, since the result of the division is the square root of the denominator. 2. Flop Counting Algorithm 2.2 Algorithm for the Cholesky decomposition <> <> and finally, <> square-root operations are needed for the <> rows. Including the <> multiplications for column <> and the additional square root operation, <> multiplications, <> summations, and N square-root operations occur, <> FLOPs in total. 2.2.2 Cholesky Decomposition <> The main advantage of the <> decomposition compared to the standard <> decomposition is that no square root operations are needed, which may require more than one FLOP depending on the given hardware platform. Another bene�t of the <> decomposition is that it does not require a positive definite matrix R, the only two conditions for the unique existence are that R is Hermitian and all but the last principle minor (i.e., the determinant) of R need to be different from zero [2]. Hence, R may also be rank de�cient to a certain degree. If R is not positive semidefinite, then D may contain negative main diagonal entries. The outcome of the decomposition is a lower triangular matrix L1 with ones on the main diagonal and a diagonal matrix D. Algorithm 2.2 overwrites the strictly lower left part of the matrix R with the strictly lower part of L1 (i.e., without the ones on the main diagonal) and overwrites the main diagonal of R with the main diagonal of D. It is taken from [1] and slightly modi�ed, such that is also applicable to complex matrices (see the conjugate in Line 4) and no existing scalar should be re-computed (see case distinction in Line 4 for i =1). Line 1 needs <> multiplications. Lines 3 to 5 require <> multiplications and are executed for <>, yielding <> multiplications. Line 6 takes <> multiplications and <> summations, again with n =2,...,N, yielding n=2(<>) = 2 multiplications and the same amount of summations. Line 7 does not require any FLOP. In Line 8, the matrix-vector product needs <> multiplications, and additional <> multiplications arise when the complete numerator is divided by the denominator. Hence, we have <> multiplications. For <> we get <> multiplications. The number of summations in Line 8 is <> for the matrix vector product and <> for the subtraction in the numerator. Together, we have <> summations. With <> summations. Summing up, this algorithm requires <> multiplications, and <> summations, yielding a total amount of <> FLOPs. (Note that this formula is also valid for N =1.) 2.3 Inverses of Matrices 2.3.1 Inverse <> of a Lower Triangular Matrix L Let <> denote the inverse of a lower triangular matrix L. Then, X is again lower triangular which means that <> for <>. The following equation holds: <>. (2.8) Via forward substitution, above system can easily be solved. Row <> from (2.8) can be expressed as <>, (2.9) with <> denoting the Kronecker delta which vanishes for <>, and <>. Starting from <>, the xb,n are computed successively, and we find <> (2.10) with all <> having been computed in previous steps. Hence, if <> and a single multiplication is required, no summations are needed. For <> multiplications and <> summations are required, as the Kronecker-delta vanishes. All main diagonal entries can be computed by means of N multiplications The lower left off-diagonal entries Actually, it is a division rather than a multiplication. 2. Flop Counting require <> (2.11) multiplications, and <> (2.12) summations. Including the N multiplications for the main-diagonal entries, <> multiplications and <> summations have to be implemented, yielding a total amount <> FLOPs. 2.3.2 Inverse <> of a Lower Triangular Matrix L1 with Ones on the Main Diagonal The inverse of a lower triangular matrix L1 turns out to require N2 FLOPs less than the inverse of L with arbitrary nonzero diagonal elements. Let X denote the inverse of L1. Clearly, X is again a lower triangular matrix with ones on the main diagonal. We can exploit this fact in order to compute only the unknown entries. The mth row and nth column of the system of equations <> with <> reads as <> or, equivalently, <> Hence, X is computed via forward substitution. To compute <>, we need <> multiplications and <> summations. Remember that <>. The total number of multiplications/summations is obtained from <>) (2.13) We only have to consider <>, since the equations resulting from m> FLOPs are needed. 2.3.3 Inverse R.1 of a Positive definite Matrix R The inverse of a matrix can for example be computed via Gaussian-elimination [1]. However, this approach is computationally expensive and does not exploit the Hermitian structure of R. Instead, it is more efficient to start with the Cholesky decomposition of <> (see Subsection 2.2.1), invert the lower triangular matrix L (see Subsection 2.3.1), and then build the Gram <> of <> (see Subsection 2.1.15). Summing up the respective number of operations, this procedure requires <> multiplications, <> summations, and N square-root operations, which yields a total amount of <> FLOPs. 2.4.1 Product <> with <> not known a priori. A naive way of computing the solution <> of the equation <> is to find <> first and afterwards multiply it by C. This approach needs <> FLOPs as shown in Sections 2.3.1 and 2.1.10. However, doing so is very expensive since we are not interested in the inverse of L in general. Hence, there must be a computationally cheaper variant. Again, forward substitution plays a key role. It is easy to see, that X can be computed column-wise. Let <> and <>. Then, from <>, we get for the element xb,a in row b and column a of X: <> Its computation requires b multiplications and <> summations. A complete column of X can therefore the computed with<> multiplications and <> summations. The complete matrix X with L columns thus needs <> FLOPs, so the forward substitution saves <> FLOPs compared to the direction inversion of L and a subsequent matrix matrix product. Interestingly, computing <> with <> unknown is as expensive as computing LC, see Section 2.1.10. 3. Overview <> and <> are arbitrary matrices.<> is a diagonal matrix, <> is lower triangular, <> is lower triangular with ones on the main diagonal, <>, and <> is positive definite. <

> Appendix A frequently occurring summation in FLOP counting is the sum of subsequent integers. By complete induction, we find <> (A1) Above result can easily be verified by recognizing that the sum of the n-th and the <> summand is equal to <>, and we have <> such pairs. Another sum of relevance is the sum of subsequent squared integers. Again, via complete induction, we find <> (A2) Bibliography [1] G. H. Golub and C.F. Van Loan, Matrix Computations, Johns Hopkins University Press, 1991. [2] Kh.D. Ikramov and N.V. Savel�eva, �Conditionally definite Matrices, Journal of Mathematical Sciences, vol. 98, no. 1, pp. 150, 2000. < <> > <> <> <> Green AI Roy Schwartz Jesse Dodge Noah A. Smith Oren Etzioni Allen Institute for AI, Seattle, Washington, USA Carnegie Mellon University, Pittsburgh, Pennsylvania, USA University of Washington, Seattle, Washington, USA Abstract The computations required for deep learning research have been doubling every few months, resulting in an estimated 300,000x increase from 2012 to 2018 [2]. These computations have a surprisingly large carbon footprint [40]. Ironically, deep learning was inspired by the human brain, which is remarkably energy efficient. Moreover, the financial cost of the computations can make it difficult for academics, students, and researchers, in particular those from emerging economies, to engage in deep learning research. This position paper advocates a practical solution by making efficiency an evaluation criterion for research along- side accuracy and related measures. In addition, we propose reporting the financial cost or “price tag” of developing, training, and running models to provide baselines for the investigation of increasingly efficient methods. Our goal is to make AI both greener and more inclusive—enabling any inspired undergraduate with a laptop to write high-quality research papers. Green AI is an emerging focus at the Allen Institute for AI. 1 Introduction and Motivation Since 2012, the field of artificial intelligence has reported remarkable progress on a broad range of capabilities in- cluding object recognition, game playing, machine translation, and more [36]. This progress has been achieved by increasingly large and computationally-intensive deep learning models. 1 Figure 1 reproduced from [2] plots training cost increase over time for state-of-the-art deep learning models starting with AlexNet in 2012 [20] to AlphaZero in 2017 [38]. The chart shows an overall increase of 300,000x, with training cost doubling every few months. An even sharper trend can be observed in NLP word embedding approaches by looking at ELMo [29] followed by BERT [8], openGPT-2 [30], and XLNet [48]. An important paper [40] has estimated the carbon footprint of several NLP models and argued that this trend is both environmentally unfriendly (which we refer to as Red AI ) and expensive, raising barriers to participation in NLP research. This trend is driven by the strong focus of the AI community on obtaining “state-of-the-art” results, 2 as exemplified by the rising popularity of leaderboards [46, 45], which typically report accuracy measures but omit any mention of cost or efficiency (see, for example,leaderboards.allenai.org). Despite the clear benefits of improving model accuracy in AI, the focus on this single metric ignores the economic, environmental, or social cost of reaching the reported accuracy. We advocate increasing research activity in Green AI —AI research that is more environmentally friendly and inclusive. We emphasize that Red AI research has been yielding valuable contributions to the field of AI, but it’s been overly dominant. We want to shift the balance towards the Green AI option—to ensure that any inspired undergraduate with a laptop has the opportunity to write high-quality papers that could be accepted at premier research conferences. 1 For brevity, we refer to AI throughout this paper, but our focus is on AI research that relies on deep learning methods. 2 Meaning, in practice, that a system’s accuracy on some benchmark is greater than any previously reported system’s accuracy. <

> Figure 1: The amount of compute used to train deep learning models has increased 300,000x in 6 years. Figure taken from [2]. Specifically, we propose making efficiency a more common evaluation criterion for AI papers alongside accuracy and related measures. AI research can be computationally expensive in a number of ways, but each provides opportunities for efficient improvements; for example, papers could be required to plot accuracy as a function of computational cost and of training set size, providing a baseline for more data-efficient research in the future. Reporting the computational price tag of finding, training, and running models is a key Green AI practice (see Equation 1). In addition to providing transparency, price tags are baselines that other researchers could improve on. Our empirical analysis in Figure 2 suggests that the AI research community has paid relatively little attention to computational efficiency. In fact, as Figure 1 illustrates, the computational cost of research is increasing exponentially, at a pace that far exceeds Moore’s Law [28]. Red AI is on the rise despite the well-known diminishing returns of increased cost (e.g., Figure 3). This paper identifies key factors that contribute to Red AI and advocates the introduction of a simple, easy-to-compute efficiency metric that could help make some AI research greener, more inclusive, and perhaps more cognitively plausible. Green AI is part of a broader, long-standing interest in environmentally-friendly scientific research (e.g., see the journalGreen Chemistry). Computer science, in particular, has a long history of investigating sustainable and energy-efficient computing (e.g., see the journalSustainable Computing: Informatics and Systems). The remainder of this paper is organized as follows. Section 2 analyzes practices that move deep-learning research into the realm of Red AI . Section 3 discusses our proposals for Green AI. Section 4 considers related work, and we conclude with a discussion of directions for future research. 2 Red AI Red AI refers to AI research that seeks to obtain state-of-the-art results in accuracy (or related measures) through the use of massive computational power—essentially “buying” stronger results. Yet the relationship between model performance and model complexity (measured as number of parameters or inference time) has long been understood to be at best logarithmic; for a linear gain in performance, an exponentially larger model is required [18]. Similar trends exist with increasing the quantity of training data [41, 13] and the number of experiments [9]. In each of these cases, diminishing returns come at increased computational cost. This section analyzes the factors contributing to Red AI and shows how it is resulting in diminishing returns over time (see Figure 3). We note again that Red AI work is valuable, and in fact, much of it contributes to what we know <

> Figure 2: AI papers tend to target accuracy rather than efficiency. The figure shows the proportion of papers that target accuracy, efficiency, both or other from a sample of 60 papers from top AI conferences. by pushing the boundaries of AI. Our exposition here is meant to highlight areas where computational expense is high, and to present each as an opportunity for developing more efficient techniques. To demonstrate the prevalence of Red AI , we sampled 60 papers from top AI conferences (ACL, 3 NeurIPS, 4 and CVPR 5 ). For each paper we noted whether the authors claim their main contribution to be (a) an improvement to accuracy or some related measure, (b) an improvement to efficiency, (c) both, or (d) other. As shown in Figure 2, in all conferences we considered, a large majority of the papers target accuracy (90% of ACL papers, 80% of NeurIPS papers and 75% of CVPR papers). Moreover, for both empirical AI conferences (ACL and CVPR) only a small portion (10% and 20% respectively) argue for a new efficiency result. 6 This highlights the focus of the AI community on measures of performance such as accuracy, at the expense of measures of efficiency such as speed or model size. In this paper we argue that a larger weight should be given to the latter. To better understand the different ways in which AI research can be red, consider an AI result reported in a scientific paper. This result typically includes a model trained on a training dataset and evaluated on a test dataset. The process of developing that model often involves multiple experiments to tune its hyperparameters. When considering the different factors that increase the computational and environmental cost of producing such a result, three factors come to mind: the cost of executing the model on a single (E)xample (either during training or at inference time); the size of the training (D)ataset, which controls the number of times the model is executed during training, and the number of (H)yperparameter experiments, which controls how many times the model is trained during model development. The total cost of producing a (R)esult in machine learning increases linearly with each of these quantities. This cost can be estimated as follows: <> Equation 1: The equation of Red AI : The cost of an AI (R)esult grows linearly with the cost of processing a single (E)xample, the size of the training (D)ataset and the number of (H)yperparameter experiments. Equation 1 is a simplification (e.g., different hyperparameter assignments can lead to different costs for processing a single example). It also ignores other factors such as the number of training epochs. Nonetheless, it illustrates three quantities that are each an important factor in the total cost of generating a result. Below, we consider each quantity separately. Interestingly, many NeurIPS papers included convergence rates or regret bounds which describe performance as a function of examples or iterations, thus targeting efficiency (55%). This indicates an increased awareness of the importance of this concept, at least in theoretical analyses. . Expensive processing of one example Our focus is on neural models, where it is common for each training step to require inference, so we discuss training and inference cost together as “processing” an example. Some works have used increasingly expensive models which require great amounts of resources, and as a result, in these models, performing inference can require a lot of computation, and training even more so. For instance, Google’s BERT-large [8] contains roughly 350 million parameters. openAI’s openGPT2-XL model [30] contains 1.5 billion parameters. AI2, our home organization, recently released Grover [49], also containing 1.5 billion parameters. In the computer vision community, a similar trend is observed (Figure 1). Such large models have high costs for processing each example, which leads to large training costs. BERT-large was trained on 64 TPU chips for 4 days. Grover was trained on 256 TPU chips for two weeks, at an estimated cost of $25,000. XLNet had a similar architecture to BERT-large, but used a more expensive objective function (in addition to an order of magnitude more data), and was trained on 512 TPU chips for 2.5 days. 7 It is impossible to reproduce the best BERT-large results 8 or XLNet results 9 using a single GPU. Specialized models can have even more extreme costs, such as AlphaGo, the best version of which required 1,920 CPUs and 280 GPUs to play a single game of Go [37] at a cost of over $1,000 per hour. 10 When examining variants of a single model (e.g., BERT-small and BERT-large) we see that larger models can have stronger performance, which is a valuable scientific contribution. However, this implies the financial and environmental cost of increasingly large AI models will not decrease soon, as the pace of model growth far exceeds the resulting increase in model performance [16]. As a result, more and more resources are going to be required to keep improving AI models by simply making them larger. Processing many examples Another way state-of-the-art performance has recently been progressing in AI is by successively increasing the amount of training data models are trained on. BERT-large had top performance in 2018 across many NLP tasks after training on 3 billion word-pieces. XLNet recently outperformed BERT after training on 32 billion word-pieces, including part of Common Crawl; openGPT-2-XL trained on 40 billion words; FAIR’s RoBERTa [23] was trained on 160GB of text, roughly 40 billion word-pieces, requiring around 25,000 GPU hours to train. In computer vision, researchers from Facebook [25] pretrained an image classification model on 3.5 billion images from Instagram, three orders of magnitude larger than existing labelled image datasets such as Open Images. 11 The use of massive data creates barriers for many researchers for reproducing the results of these models, or training their own models on the same setup (especially as training for multiple epochs is standard). For example, the June 2019 Common Crawl contains 242 TB of uncompressed data, 12 so even storing the data is expensive. Finally, as in the case of model size, relying on more data to improve performance is notoriously expensive because of the diminishing return of adding more data [41]. For instance, Figure 3, taken from [25], shows a logarithmic relation between the object recognition top-1 accuracy and the number of training examples. Massive number of experiments Some projects have poured large amounts of computation into tuning hyperparameters or searching over neural architectures, well beyond the reach of most researchers. For instance, researchers from Google [51] trained over 12,800 neural networks in their neural architecture search to improve performance on object detection and language modeling. With a fixed architecture, researchers from DeepMind [26] evaluated 1,500 hyperparameter assignments to demonstrate that an LSTM language model [15] can reach state-of-the-art perplexity results. Despite the value of this result in showing that the performance of an LSTM does not plateau after only a few hyperparameter trials, fully exploring the potential of other competitive models for a fair comparison is prohibitively expensive. 7 Some estimates for the cost of this process reach $250,000 (twitter.com/eturner303/status/1143174828804857856). 8 Seehttps://github.com/google-research/bert 9 Seehttps://github.com/zihangdai/xlnet 10 Recent versions of AlphaGo are far more efficient [39]. 11 https://opensource.google.com/projects/open-images-dataset 12 http://commoncrawl.org/2019/07/ <

> Figure 3: Diminishing returns of training on more data: object detection accuracy increases linearly as the number of training examples increases exponentially [25]. The topic of massive number of experiments is not as well studied as the first two discussed above. In fact, the number of experiments performed during model construction is often under reported. Nonetheless, evidence for a logarithmic relation exists here as well, between the number of experiments and performance gains [9]. Discussion The benefits of pouring more resources into models are certainly of interest to the AI community. Indeed, there is value in pushing the limits of model size, dataset size, and the hyperparameter search space. Currently, despite the massive amount of resources put into recent AI models, such investment still pays off in terms of downstream performance (albeit at an increasingly lower rate). Finding the point of saturation (if such exists) is an important question for the future of AI. Our goal in this paper is to raise awareness of the cost of Red AI , as well as encourage the AI community to recognize the value of work by researchers that take a different path, optimizing efficiency rather than accuracy. Next we turn to discuss concrete measures for making AI more green. 3 Green AI The term Green AI refers to AI research that yields novel results without increasing computational cost, and ideally reducing it. Whereas Red AI has resulted in rapidly escalating computational (and thus carbon) costs, Green AI has the opposite effect. If measures of efficiency are widely accepted as important evaluation metrics for research alongside accuracy, then researchers will have the option of focusing on the efficiency of their models with positive impact on both the environment and inclusiveness. This section reviews several measures of efficiency that could be reported and optimized, and advocates one particular measure—FPO—which we argue should be reported when AI research findings are published. 3.1 Measures of Efficiency To measure efficiency, we suggest reporting the amount of work required to generate a result in AI, that is, the amount of work required to train a model, and if applicable, the sum of works for all hyperparameter tuning experiments. As the cost of an experiment decomposes into the cost of a processing a single example, the size of the dataset, and the number of experiments (Equation 1), reducing the amount of work in each of these steps will result in AI that is more green. When reporting the amount of work done by a model, we want to measure a quantity that allows for a fair comparison between different models. As a result, this measure should ideally be stable across different labs, at different times, and using different hardware. Carbon emission Carbon emission is appealing as it is a quantity we want to directly minimize. Nonetheless it is impractical to measure the exact amount of carbon released by training or executing a model, and accordingly— generating an AI result, as this amount depends highly on the local electricity infrastructure. As a result, it is not comparable between researchers in different locations or even the same location at different times. Electricity usage Electricity usage is correlated with carbon emission while being time- and location-agnostic. Moreover, GPUs often report the amount of electricity each of their cores consume at each time point, which facilitates the estimation of the total amount of electricity consumed by generating an AI result. Nonetheless, this measure is hardware dependent, and as a result does not allow for a fair comparison between different models. Elapsed real time The total running time for generating an AI result is a natural measure for efficiency, as all other things being equal, a faster model is doing less computational work. Nonetheless, this measure is highly influenced by factors such as the underlying hardware, other jobs running on the same machine, and the number of cores used. These factors hinder the comparison between different models, as well as the decoupling of modeling contributions from hardware improvements. Number of parameters Another common measure of efficiency is the number of parameters (learnable or total) used by the model. As with run time, this measure is correlated with the amount of work. Unlike the other measures described above, it does not depend on the underlying hardware. Moreover, this measure also highly correlates with the amount of memory consumed by the model. Nonetheless, different algorithms make different use of their parameters, for instance by making the model deeper vs. wider. As a result, different models with a similar number of parameters often perform different amounts of work. FPO As a concrete measure, we suggest reporting the total number of floating point operations (FPO) required to generate a result. 13 FPO provides an estimate to the amount of work performed by a computational process. It is computed analytically by defining a cost to two base operations, ADD and MUL . Based on these operations, the FPO cost of any machine learning abstract operation (e.g., a tanh operation, a matrix multiplication, a convolution operation, or the BERT model) can be computed as a recursive function of these two operations. FPO has been used in the past to quantify the energy footprint of a model [27, 43, 12, 42], but is not widely adopted in AI. FPO has several appealing properties. First, it directly computes the amount of work done by the running machine when executing a specific instance of a model, and is thus tied to the amount of energy consumed. Second, FPO is agnostic to the hardware on which the model is run. This facilitates fair comparisons between different approaches, unlike the measures described above. Third, FPO is strongly correlated with the running time of the model [4]. Unlike asymptotic runtime, FPO also considers the amount of work done at each time step. Several packages exist for computing FPO in various neural network libraries, 14 though none of them contains all the building blocks required to construct all modern AI models. We encourage the builders of neural network libraries to implement such functionality directly. 13 Floating point operations are often referred to as FLOP(s), though this term is not uniquely defined [12]. To avoid confusion, we use the term FPO. 14 E.g.,https://github.com/Swall0w/torchstat;https://github.com/Lyken17/pytorch-OpCounter <

> Figure 4: Increase in FPO results in diminishing return for object detection top-1 accuracy. Plots (bottom to top): model parameters (in million), FPO (in billions), top-1 accuracy on ImageNet. (4a): Different models: AlexNet [20], ResNet [14], ResNext [47], DPN107 [5], SENet154 [17]. (4b): Comparison of different sizes (measured by the number of layers) of the ResNet model [14]. Discussion Efficient machine learning approaches have received attention in the research community, but are generally not motivated by being green. For example, a significant amount of work in the computer vision community has addressed efficient inference, which is necessary for real-time processing of images for applications like self-driving cars [24, 31, 22], or for placing models on devices such as mobile phones [16, 34]. Most of these approaches target efficient model inference [32, 50, 12], 15 and thus only minimize the cost of processing a single example, while ignoring the other two red practices discussed in Section 2. 16 The above examples indicate that the path to making AI green depends on how it is used. When developing a new model, much of the research process involves training many model variants on a training set and performing inference on a small development set. In such a setting, more efficient training procedures can lead to greater savings, while in a production setting more efficient inference can be more important. We advocate for a holistic view of computational savings which doesn’t sacrifice in some areas to make advances in others. FPO has some limitations. First, it targets the electricity consumption of a model, while ignoring other potential limiting factors for researchers such as the memory consumption by the model, which can often lead to additional energy and monetary costs [24]. Second, the amount of work done by a model largely depends on the model implementation, as two different implementations of the same model could result in very different amounts of processing work. Due to the focus on the modeling contribution, the AI community has traditionally ignored the quality or efficiency of models’ implementation. We argue that the time to reverse this norm has come, and that exceptionally good implementations that lead to efficient models should be credited by the AI community. 3.2 FPO Cost of Existing Models To demonstrate the importance of reporting the amount of work, we present FPO costs for several existing models. A few trends are observable. First, as discussed in Section 2, models get more expensive with time, but the increase in FPO does not lead to similar performance gains. For instance, an increase of almost 35% in FPO between ResNet and ResNext (second and third points in graph) resulted in a 0.5% top-1 accuracy improvement. Similar patterns are observed when considering the effect of other increases in model work. Second, the number of model parameters does not tell the whole story: AlexNet (first point in the graph) actually has more parameters than ResNet (second point), but dramatically less FPO, and also much lower accuracy. Figure 4b shows the same analysis for a single object recognition model, ResNet [14], while comparing different versions of the model with different number of layers. This creates a controlled comparison between the different models, as they are identical in architecture, except for their size (and accordingly, their FPO cost). Once again, we notice the same trend: the large increase in FPO cost does not translate to a large increase in performance. 14 Figure 4a shows the number of parameters and FPO of several leading object recognition models, as well as their performance on the ImageNet dataset [6]. 15 Some very recent work also targeted efficient training [7]. 16 In fact, creating smaller models often results in longer running time, so mitigating the different trends might be at odds [44]. 17 We consider this exclusive focus on the final prediction another symptom of Red AI . 18 These numbers represent FPO per inference, i.e., the work required to process a single example. 3.3 Additional Ways to Promote Green AI In addition to reporting the FPO cost of the final reported number, we encourage researchers to report the bud- get/accuracy curve observed during training. In a recent paper [9], we observed that selecting the better performing model on a given task depends highly on the amount of compute available during model development. We introduced a method for computing the expected best validation performance of a model as a function of the given budget. We argue that reporting this curve will allow users to make wiser decisions about their selection of models and highlight the stability of different approaches. We further advocate for making efficiency an official contribution in major AI conferences, by advising reviewers to recognize and value contributions that do not strictly improve state of the art, but have other benefits such as efficiency. Finally, we note that the trend of releasing pretrained models publicly is a green success, and we would like to encourage organizations to continue to release their models in order to save others the costs of retraining them. 4 Related Work Recent work has analyzed the carbon emissions of training deep NLP models [40] and concluded that computationally expensive experiments can have a large environmental and economic impact. With modern experiments using such large budgets, many researchers (especially those in academia) lack the resources to work in many high-profile areas; increased value placed on computationally efficient approaches will allow research contributions from more diverse groups. We emphasize that the conclusions of [40] are the result of long-term trends, and are not isolated within NLP, but hold true across machine learning. While some companies offset electricity usage by purchasing carbon credits, it is not clear that buying credits is as effective as using less energy. In addition, purchasing carbon credits is voluntary; Google cloud 20 and Microsoft Azure 21 purchase carbon credits to offset their spent energy, but Amazon’s AWS 22 (the largest cloud computing plat- form 23 ) only covered fifty percent of its power usage with renewable energy. The push to improve state-of-the-art performance has focused the research community’s attention on reporting the single best result after running many experiments for model development and hyperparameter tuning. Failure to fully report these experiments prevents future researchers from understanding how much effort is required to reproduce a result or extend it [9]. Our focus is on improving efficiency in the machine learning community, but machine learning can also be used as a tool for work in areas like climate change. For example, machine learning has been used for reducing emissions of cement plants [1] and tracking animal conservation outcomes [11], and is predicted to be useful for forest fire management [33]. Undoubtedly these are important applications of machine learning; we recognize that they are orthogonal to the content of this paper. 19 Numbers taken fromhttps://github.com/sovrasov/flops-counter.pytorch 20 https://cloud.google.com/sustainability/ 21 https://www.microsoft.com/en-us/environment/carbon 22 https://aws.amazon.com/about-aws/sustainability/ 23 https://tinyurl.com/y2kob969 8 5 Conclusion The vision of Green AI raises many exciting research directions that help to overcome the inclusiveness challenges of Red AI . Progress will reduce the computational expense with a minimal reduction in performance, or even improve performance as more efficient methods are discovered. Also, it would seem that Green AI could be moving us in a more cognitively plausible direction as the brain is highly efficient. It’s important to reiterate that we see Green AI as a valuable option not an exclusive mandate—of course, both Green AI and Red AI have contributions to make. We want to increase the prevalence of Green AI by highlighting its benefits, advocating a standard measure of efficiency. Below, we point to a few important green research directions, and highlight a few open questions. Research on building space or time efficient models is often motivated by fitting a model on a small device (such as a phone) or fast enough to process examples in real time, such as image captioning for the blind (see Section 3.1). Some modern models don’t even fit on a single GPU (see Section 2). Here we argue for a far broader approach. Data efficiency has received significant attention over the years [35, 19]. Modern research in vision and NLP often involves first pretraining a model on large “raw” (unannotated) data then fine-tuning it to a task of interest through supervised learning. A strong result in this area often involves achieving similar performance to a baseline with fewer training examples or fewer gradient steps. Most recent work has addressed fine-tuning data [29], but pretraining efficiency is also important. In either case, one simple technique to improve in this area is to simply report performance with different amounts of training data. For example, reporting performance of contextual embedding models trained on 10 million, 100 million, 1 billion, and 10 billion tokens would facilitate faster development of new models, as they can first be compared at the smallest data sizes. Research here is of value not just to make training less expensive, but because in areas such as low resource languages or historical domains it is extremely hard to generate more data, so to progress we must make more efficient use of what is available. Finally, the total number of experiments run to get a final result is often underreported and underdiscussed [9]. The few instances researchers have of full reporting of the hyperparameter search, architecture evaluations, and ablations that went into a reported experimental result have surprised the community [40]. While many hyperparameter optimization algorithms exist which can reduce the computational expense required to reach a given level of performance [3, 10], simple improvements here can have a large impact. For example, stopping training early for models which are clearly underperforming can lead to great savings [21]. References [1]Prabal Acharyya, Sean D Rosario, Roey Flor, Ritvik Joshi, Dian Li, Roberto Linares, and Hongbao Zhang. Autopilot of cement plants for reduction of fuel consumption and emissions, 2019. ICML Workshop on Climate Change. [2]Dario Amodei and Danny Hernandez. AI and compute, 2018. Blog post. [3]James S. Bergstra, Remi Bardenet, Yoshua Bengio, and Bal´ azs K´ egl. Algorithms for hyper-parameter optimiza-´ tion. InProc. of NeurIPS, 2011. [4]Alfredo Canziani, Adam Paszke, and Eugenio Culurciello. An analysis of deep neural network models for practical applications. InProc. of ISCAS, 2017. [5]Yunpeng Chen, Jianan Li, Huaxin Xiao, Xiaojie Jin, Shuicheng Yan, and Jiashi Feng. Dual path networks. In Proc. of NeurIPS, 2017. [6]Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. InProc. of CVPR, 2009. [7]Tim Dettmers and Luke Zettlemoyer. Sparse networks from scratch: Faster training without losing performance, 2019. arXiv:1907.04840. [8]Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. InProc. of NAACL, 2019. [9]Jesse Dodge, Suchin Gururangan, Dallas Card, Roy Schwartz, and Noah A. Smith. Show your work: Improved reporting of experimental results. InProc. of EMNLP, 2019. [10]Jesse Dodge, Kevin Jamieson, and Noah A. Smith. Open loop hyperparameter optimization and determinantal point processes. InProc. of AutoML, 2017. [11]Clement Duhart, Gershon Dublon, Brian Mayton, Glorianna Davenport, and Joseph A. Paradiso. Deep learning for wildlife conservation and restoration efforts, 2019. ICML Workshop on Climate Change. [12]Ariel Gordon, Elad Eban, Ofir Nachum, Bo Chen, Hao Wu, Tien-Ju Yang, and Edward Choi. MorphNet: Fast & simple resource-constrained structure learning of deep networks. InProc. of CVPR, 2018. [13]Alon Halevy, Peter Norvig, and Fernando Pereira. The unreasonable effectiveness of data. IEEE Intelligent Systems, 24:8–12, 2009. [14]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. of CVPR, 2016. [15]Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory.¨ Neural computation, 9(8):1735–1780, 1997. [16]Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco An- dreetto, and Hartwig Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications, 2017. arXiv:1704.04861. [17]Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. InProc. of CVPR, 2018. [18]Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbig- niew Wojna, Yang Song, Sergio Guadarrama, and Kevin Murphy. Speed/accuracy trade-offs for modern convo- lutional object detectors. InProc. of CVPR, 2017. [19]Sanket Kamthe and Marc Peter Deisenroth. Data-efficient reinforcement learning with probabilistic model pre- dictive control. InProc. of AISTATS, 2018. [20]Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. InProc. of NeurIPS, 2012. [21]Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Hyperband: Bandit- based configuration evaluation for hyperparameter optimization. InProc. of ICLR, 2017. [22]Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. InProc. of ECCV, 2016. [23]Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized bert pretraining approach, 2019. arXiv:1907.11692. [24]Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. ShuffleNet V2: Practical guidelines for efficient cnn architecture design. InProc. of ECCV, 2018. [25]Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens van der Maaten. Exploring the limits of weakly supervised pretraining. InProc. ECCV, 2018. [26]Gabor Melis, Chris Dyer, and Phil Blunsom. On the state of the art of evaluation in neural language models. In´ Proc. of EMNLP, 2018. [27]Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference. InProc. of ICLR, 2017. [28]Gordon E. Moore. Cramming more components onto integrated circuits, 1965. [29]Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettle- moyer. Deep contextualized word representations. InProc. of NAACL, 2018. [30]Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners, 2019. OpenAI Blog. [31]Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. InProc. of ECCV, 2016. [32]Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. InProc. of CVPR, 2016. [33]David Rolnick, Priya L. Donti, Lynn H. Kaack, Kelly Kochanski, Alexandre Lacoste, Kris Sankaran, An- drew Slavin Ross, Nikola Milojevic-Dupont, Natasha Jaques, Anna Waldman-Brown, Alexandra Luccioni, Tegan Maharaj, Evan D. Sherwin, S. Karthik Mukkavilli, Konrad P. Kording, Carla Gomes, Andrew Y. Ng, Demis Has-¨ sabis, John C. Platt, Felix Creutzig, Jennifer Chayes, and Yoshua Bengio. Tackling climate change with machine learning, 2019. arXiv:1905.12616. [34]Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. InProc. of CVPR, 2018. [35]Roy Schwartz, Sam Thomson, and Noah A. Smith. SoPa: Bridging CNNs, RNNs, and weighted finite-state machines. InProc. of ACL, 2018. [36]Yoav Shoham, Raymond Perrault, Erik Brynjolfsson, Jack Clark, James Manyika, Juan Carlos Niebles, Terah Lyons, John Etchemendy, and Z Bauer. The AI index 2018 annual report. AI Index Steering Committee, Human-Cente Red AI Initiative, Stanford University. Available athttp://cdn.aiindex.org/2018/AI% 20Index%202018%20Annual%20Report.pdf, 202018, 2018. [37]David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of Go with deep neural networks and tree search.Nature, 529(7587):484, 2016. [38]David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, Timothy Lillicrap, Karen Simonyan, and Demis Hassabis. Mastering chess and shogi by self-play with a general reinforcement learning algorithm, 2017. arXiv:1712.01815. [39]David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, Yutian Chen, Timothy Lillicrap, Fan Hui, Laurent Sifre, George van den Driessche, Thore Graepel, and Demis Hassabis. Mastering the game of Go without human knowledge.Nature, 550(7676):354, 2017. [40]Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep learning in NLP. InProc. of ACL, 2019. [41]Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. InProc. of ICCV, 2017. [42]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InProc. of NeurIPS, 2017. [43]Tom Veniat and Ludovic Denoyer. Learning time/memory-efficient deep architectures with budgeted super net- works. InProc. of CVPR, 2018. [44]Aaron Walsman, Yonatan Bisk, Saadia Gabriel, Dipendra Misra, Yoav Artzi, Yejin Choi, and Dieter Fox. Early fusion for goal directed robotic vision. InProc. of IROS, 2019. [45]Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. SuperGLUE: A stickier benchmark for general-purpose language understanding systems, 2019. arXiv:1905.00537. [46]Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. InProc. of ICLR, 2019. [47]Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. InProc. of CVPR, 2017. [48]Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. XLNet: Generalized autoregressive pretraining for language understanding, 2019. arXiv:1906.08237. [49]Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. Defending against neural fake news, 2019. arXiv:1905.12616. [50]Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. ShuffleNet: An extremely efficient convolutional neural network for mobile devices. InProc. of CVPR, 2018. [51]Barret Zoph and Quoc V. Le. Neural architecture search with reinforcement learning. InProc. of ICLR, 2017. <> <> <> <> <> <> Harnessing Nonlinearity: Predicting Chaotic Systems and Saving Energy in Wireless Communication Herbert Jaeger* and Harald Haas We present a method for learning nonlinear systems, echo state networks (ESNs). ESNs employ artificial recurrent neural networks in a way that has recently been proposed independently as a learning mechanism in biological brains. The learning method is computationally efficient and easy to use. On a benchmark task of predicting a chaotic time series, accuracy is improved by a factor of 2400 over previous techniques. The potential for engineering applications is illustrated by equalizing a communication channel, where the signal error rate is improved by two orders of magnitude. Nonlinear dynamical systems abound in the sciences and in engineering. If one wishes to simulate, predict, filter, classify, or control such a system, one needs an executable system model. However, it is often infeasible to obtain analytical models. In such cases, one has to resort to black-box models, which ignore the internal physical mechanisms and instead reproduce only the outwardly observable input-output behavior of the target system. If the target system is linear, efficient methods for black-box modeling are available. Most technical systems, however, become nonlinear if operated at higher operational points (that is, closer to saturation). Although this might lead to cheaper and more energy-efficient designs, it is not done be.cause the resulting nonlinearities cannot be harnessed. Many biomechanical systems use their full dynamic range (up to saturation) and thereby become lightweight, energy efficient, and thoroughly nonlinear. Here, we present an approach to learn.ing black-box models of nonlinear systems, echo state networks (ESNs). An ESN is an artificial recurrent neural network (RNN). RNNs are characterized by feedback (recurrent) loops in their synaptic connection pathways. They can maintain an ongoing activation even in the absence of input and thus exhibit dynamic memory. Biological neural networks are typically recurrent. Like biological neural networks, an artificial RNN can learn to mimic a target system in principle, with arbitrary accuracy (1). Several learning algorithms are known (24) that incrementally adapt the synaptic weights of an RNN in order to tune it toward the target system. These algorithms have not been widely employed in technical applications because of slow International University Bremen, Bremen D-28759, Germany. convergence and suboptimal solutions (5, 6). The ESN approach differs from these methods in that a large RNN is used (on the order of 50 to 1000 neurons; previous techniques typically use 5 to 30 neurons) and in that only the synaptic connections from the RNN to the output readout neurons are modified by learning; previous techniques tune all synaptic connections (Fig. 1). Be.cause there are no cyclic dependencies be.tween the trained readout connections, training an ESN becomes a simple linear regression task. We illustrate the ESN approach on a task of chaotic time series prediction (Fig. 2) (7). The Mackey-Glass system (MGS) (8) is a standard benchmark system for time series prediction studies. It generates a sub.tly irregular time series (Fig. 2A). The prediction task has two steps: (i) using an initial teacher sequence generated by the original MGS to learn a black-box model M of the generating system, and (ii) using M to predict the value of the sequence some steps ahead. First, we created a random RNN with 1000 neurons (called the reservoir) and one output neuron. The output neuron was equipped with random connections that project back into the reservoir (Fig. 2B). A 3000-step teacher sequence <> was generated from the MGS equation and fed into the output neuron. This excited the internal neurons through the output feedback connections. After an initial transient period, they started to exhibit systematic individual variations of the teacher sequence (Fig. 2B). The fact that the internal neurons display systematic variants of the exciting external signal is constitutional for ESNs: The internal neurons must work as echo functions for the driving signal. Not every randomly generated RNN has this property, but it can effectively be built into a reservoir (support.ing online text). It is important that the echo signals be richly varied. This was ensured by a sparse interconnectivity of 1% within the reservoir. This condition lets the reservoir decompose into many loosely coupled subsystems, establishing a richly structured reservoir of excitable dynamics. After time <>, output connection weights wi (i  1, . . . , 1000) were computed (dashed arrows in Fig. 2B) from the last 2000 steps n=1001, . . . , 3000 of the training run such that the training error <> was minimized [<>, activation of the ith internal neuron at time n]. This is a simple linear regression. With the new wi in place, the ESN was disconnected from the teacher after step 3000 and left running freely. A bidirectional dynamical interplay of the network-generated output signal with the internal signals <> unfolded. The output signal <> was created from the internal neuron activation signals <> through the trained connections wi,by <>. Conversely, the internal signals were echoed from that output signal through the fixed output feedback connections (supporting online text). For testing, an 84-step continuation <> of the original signal was computed for reference. The network output y(3084) was compared with the cor.rect continuation d(3084). Averaged over 100 independent trials, a normalized root mean square error <> was obtained <> and <> teacher and network <

> Fig. 1. (A) Schema of previous approaches to RNN learning. (B) Schema of ESN approach. Solid synaptic connections; dotted arrows, adjustable connections. Both approaches aim at minimizing the error <>, where <> is the network output and d(n) is the teacher time series observed from the target system. output in trial j, 2 variance of MGS signal), improving the best previous techniques (9 15), which used training sequences of length 500 to 10,000, by a factor of 700. If the prediction run was continued, deviations typically became visible after about 1300 steps (Fig. 2A). With a refined variant of the learn.ing method (7), the improvement factor rises to 2400. Models of similar accuracy were also obtained for other chaotic systems (supporting online text). The main reason for the jump in modeling accuracy is that ESNs capitalize on a massive short-term memory. We showed analytically (16) that under certain conditions an ESN of size N may be able to "remember" a number of previous inputs that is of the same order of magnitude as N. This information is more massive than the information used in other techniques (supporting online text). We now illustrate the approach in a task of practical relevance, namely, the equalization of a wireless communication channel (7). The essentials of equalization are as fol.lows: A sender wants to communicate a sym.bol sequence s(n). This sequence is first transformed into an analog envelope signal d(n), then modulated on a high-frequency carrier signal and transmitted, then received and demodulated into an analog signal u(n), which is a corrupted version of d(n). Major sources of corruption are noise (thermal or due to interfering signals), multipath propagation, which leads to a superposition of adjacent symbols (intersymbol interference), and nonlinear distortion induced by operating the senders power amplifier in the high-gain region. To avoid the latter, the actual power amplification is run well below the maximum amplification possible, thereby incurring a substantial loss in energy efficiency, which is clearly undesirable in cell-phone and satellite Fig. 2. (A) Prediction output of the trained ESN (dotted) overlaid with the correct continuation (solid). (B) Learning the MG attractor. Three sample activation traces of internal neurons are shown. They echo the teacher signal d(n). After training, the desired output is recreated from the echo signals through output connections (dotted arrows) whose weights wi are the result of the training procedure. communications. The corrupted signal u(n)is then passed through an equalizing filter whose output y(n) should restore u(n)as closely as possible to d(n). Finally, the equalized signal y(n) is converted back into a symbol sequence. The quality measure for the entire process is the fraction of incorrect symbols finally obtained (symbol error rate). To compare the performance of an ESN equalizer with standard techniques, we took a channel model for a nonlinear wireless transmission system from a study (17) that compared three customary nonlinear equalization methods: a linear decision feedback equalizer (DFE), which is actually a non.linear method; a Volterra DFE; and a bilinear DFE. The model equation featured inter symbol interference across 10 consecutive symbols, a second-order and a third-order nonlinear distortion, and additive white Gaussian noise. All methods investigated in that study had 47 adjustable parameters and used sequences of 5000 symbols for training. To make the ESN equalizer comparable with the equalizers studied in (17), we took ESNs with a reservoir of 46 neurons (which is small for the ESN approach), which yielded 47 adjust.able parameters. (The 47th comes from a direct connection from the input to the output neuron.) We carried out numerous learning trials (7) to obtain ESN equalizers, using an online learning method (a version of the recursive least square algorithm known from linear adaptive filters) to train the output weights on 5000-step training sequences. We chose an online adaptation scheme here because the methods in (17) were online adaptive, too, and because wireless communication channels mostly are time-varying, such that an equalizer must adapt to changing system characteristics. The entire learning-testing procedure was repeated for signal-to-noise <

> Fig. 3. Results of using an ESN for nonlinear channel equalization. Plot shows signal error rate (SER) versus signal-to-noise ratio (SNR). (a) Linear DFE. (b) Volterra DFE. (c) Bilinear DFE. [(a) to (c) taken from (20)]. (d) Blue line represents average ESN performance with randomly generated reservoirs. Error bars, variation across networks. (e) Green line indicates performance of best network chosen from the networks averaged in (d). Error bars, variation across learning trials. REPORTS ratios ranging from 12 to 32 db. Figure 3 compares the average symbol error rates obtained with the results reported in (17), show.ing an improvement of two magnitudes for high signal-to-noise ratios. For tasks with multichannel input and/or output, the ESN approach can be accommodated simply by adding more input or output neurons (16, 18). ESNs can be applied to all basic tasks of signal processing and control, including time series prediction, inverse modeling, pattern generation, event detection and classification, modeling distributions of stochastic process.es, filtering, and nonlinear control (16, 18, 19, 20). Because a single learning run takes only a few seconds (or minutes, for very large data sets and networks), engineers can test out variants at a high turnover rate, a crucial factor for practical usability. ESNs have been developed from a mathematical and engineering perspective, but exhibit typical features of biological RNNs: a large number of neurons, recurrent pathways, sparse random connectivity, and local modification of synaptic weights. The idea of using randomly connected RNNs to represent and memorize dynamic input in network states has frequently been explored in specific contexts, for instance, in artificial intelligence models of associative memory (21), models of prefrontal cortex function in sensory-motor sequencing tasks (22), models of birdsong (23), models of the cerebellum (24), and general computational models of neural oscillators (25). Many different learning mechanisms were considered, mostly within the RNN itself. The contribution of the ESN is to elucidate the mathematical properties of large RNNs such that they can be used with a linear, trainable readout mechanism for general black-box modeling. An approach essentially equivalent to ESNs, liquid state networks (26, 27), has been developed independently to model computations in cortical microcircuits. Recent findings in neurophysiology suggest that the basic ESN/liquid state network principle seems not uncommon in biological networks (28,30) and could eventually be exploited to control prosthetic devices by signals collected from a collective of neurons (31). References and Notes 1. K.-I. Funahashi, Y. Nakamura, Neural Netw. 6, 801 (1993). 2. D. Zipser, R. J. Williams, Neural Comput. 1, 270 (1989). 3. P. J. Werbos, Proc. IEEE 78, 1550 (1990). 4. L. A. Feldkamp, D. V. Prokhorov, C. F. Eagen, F. Yuan, in Nonlinear Modeling: Advanced Black-Box techniques , J. A. K. Suykens, J. Vandewalle, Eds. (Kluwer, Dordrecht, Netherlands, 1998), pp. 29�54. 5. K. Doya, in The Handbook of Brain Theory and Neural Networks, M. A. Arbib, Ed. (MIT Press, Cambridge, MA, 1995), pp. 796�800. 6. H. Jaeger, �Tutorial on training recurrent neural networks� (GMD-Report 159, German National Re.search Institute for Computer Science, 2002); ftp:// borneo.gmd.de/pub/indy/publications_herbert/ CompleteTutorialTechrep.pdf. REPORTS 7. Materials andmethods are available as supporting material on Science Online. 8. M. C. Mackey, L. Glass, Science 197, 287 (1977). 9. J. Vesanto, in Proc. WSOM �97 (1997); www.cis.hut.�/ projects/monitor/publications/papers/wsom97.ps. 10. L. Chudy, I. Farkas, Neural Network World 8, 481 (1998). 11. H. Bersini, M. Birattari, G. Bontempi, in Proc. IEEE World Congr. on Computational Intelligence (IJCNN �98) (1997), pp. 2102�2106; ftp://iridia.ulb.ac.be/ pub/lazy/papers/IridiaTr1997-13_2.ps.gz. 12. T. M. Martinetz, S. G. Berkovich, K. J. Schulten, IEEE Trans. Neural Netw. 4, 558 (1993). 13. X. Yao, Y. Liu, IEEE Trans. Neural Netw. 8, 694 (1997). 14. F. Gers, D. Eck, J. F. Schmidhuber, �Applying LSTM to time series predictable through time-window ap.proaches� (IDSIA-IDSIA-22-00, 2000); www.idsia.ch/ felix/Publications.html. 15. J. McNames, J. A. K. Suykens, J. Vandewalle, Int. J. Bifurcat. Chaos 9, 1485 (1999). 16. H. Jaeger, �Short term memory in echo state net.works� (GMD-Report 152, German National Re.search Institute for Computer Science, 2002); ftp:// borneo.gmd.de/pub/indy/publications_herbert/ STMEchoStatesTechRep.pdf. 17. V. J. Mathews, J. Lee, in Advanced Signal Processing: Algorithms, Architectures, and Implementations V (Proc. SPIE Vol. 2296), (SPIE, San Diego, CA, 1994), pp. 317�327. 18. J. Hertzberg, H. Jaeger, F. Scho�nherr, in Proc. 15th Europ. Conf. on Art. Int. (ECAI 02), F. van Harmelen, Ed. (IOS Press, Amsterdam, 2002), pp. 708�712; www. ais.fhg.de/schoenhe/papers/ECAI02.pdf. 19. H. Jaeger, �The echo state approach to analysing and training recurrent neural networks� (GMD-Report 148, German National Research Institute for Com.puter Science, 2001); ftp://borneo.gmd.de/pub/indy/ publications_herbert/EchoStatesTechRep.pdf. 20. H. Jaeger, in Advances in Neural Information Process.ing Systems 15, S. Becker, S. Thrun, K. Obermayer, Eds. (MIT Press, Cambridge, MA, 2003) pp. 593�600. 21. G. E. Hinton, in Parallel Models of Associative Mem.ory, G. E. Hinton, J. A. Anderson, Eds. (Erlbaum, Hills.dale, NJ, 1981), pp. 161�187. 22. D. G. Beiser, J. C. Houk, J. Neurophysiol. 79, 3168 (1998). 23. S. Dehaene, J.-P. Changeux, J.-P. Nadal, Proc. Natl. Acad. Sci. U.S.A. 84, 2727 (1987). 24. M. Kawato, in The Handbook of Brain Theory and Neural Networks, M. Arbib, Ed. (MIT Press, Cam.bridge, MA, 1995), pp. 172�178. 25. K. Doya, S. Yoshizawa, Neural Netw. 2, 375 (1989). Ultrafast Electron Crystallography of Interfacial Water Chong-Yu Ruan, Vladimir A. Lobastov, Franco Vigliotti, Songye Chen, Ahmed H. Zewail* We report direct determination of the structures and dynamics of interfacial water on a hydrophilic surface with atomic-scale resolution using ultrafast electron crystallography. On the nanometer scale, we observed the coexistence of ordered surface water and crystallite-like ice structures, evident in the superposition of Bragg spots and Debye-Scherrer rings. The structures were determined to be dominantly cubic, but each undergoes different dynamics after the ultrafast sub.strate temperature jump. From changes in local bond distances (OH

Oand O


O) with time, we elucidated the structural changes in the far-from-equilibrium regime at short times and near-equilibration at long times. The nature of interfacial molecular assemblies of nanometer scale is of fundamental impor.tance to chemical and biological phenomena (1�4). For water, the directional molecular fea.tures of hydrogen bonding (5, 6) and the dif.ferent structures possible, from amorphous (7) to crystalline (8), make the interfacial (9) col.lective assembly on the mesoscopic (10) scale much less understood. Structurally, the nature of water on a substrate is determined by forces of orientation at the interface and by the net charge density, which establishes the hydro.philic or hydrophobic character of the substrate. However, the transformation from ordered to dis.ordered structure and their coexistence critically depends on the time scales for the movements of atoms locally and at long range. Therefore, it is essential to elucidate the nature of these structures and the time scales for their equilibration. Laboratory for Molecular Sciences, Arthur Amos Noyes Laboratory of Chemical Physics, California Institute of Technology, Pasadena, CA 91125, USA. *To whom correspondence should be addressed. E.mail: zewail@caltech.edu Here, we report direct determination of the structures of interfacial water with atomic-scale resolution, using diffraction and the dynamics following ultrafast infrared (IR) laser-initiated 26. W. Maass, T. Natschla�ger, H. Markram, Neural Com-put. 14, 2531 (2002). 27. W. Maass, T. Natschla�ger, H. Markram, in Compu.tational Neuroscience: A Comprehensive Approach, J. Feng, Ed. (Chapman & Hall/CRC, 2003), pp. 575� 605. 28. G. B. Stanley, F. F. Li, Y. Dan, J. Neurosci. 19, 8036 (1999). 29. G. B. Stanley, Neurocomputing 38�40, 1703 (2001). 30. W. M. Kistler, Ch. I. de Zeeuw, Neural Comput. 14, 2597 (2002). 31. S. Mussa-Ivaldi, Nature 408, 361 (2000). 32. The �rst author thanks T. Christaller for unfaltering support andW. Maass for friendly cooperation. Inter.national patents are claimedby Fraunhofer AIS (PCT/ EP01/11490). Supporting Online Material www.sciencemag.org/cgi/content/full/304/5667/78/DC1 Materials andMethods SOM Text Figs. S1 to S4 References temperature jump. Interfacial water is formed on a hydrophilic surface (silicon, chlorine-terminated) under controlled ultrahigh vacuum (UHV) conditions (Fig. 1). With these atomic-scale spatial, temporal, and energy resolutions, the evolution of nonequilibrium structures was monitored, their ordered or disordered nature was established, and the time scale for the breakage of long-range bonding and formation of new structures was determined. We identi.fied the structured and ordered interfacial water from the Bragg diffraction and the layered crys.tallite structure from the Debye-Scherrer rings. The temporal evolution of interfacial water and layered ice after the temperature jump was studied with submonolayer sensitivity. We compared these results with those obtained on hydrophobic surfaces, such as hydrogen-terminated silicon or silver substrate. Spectroscopic techniques, such as internal reflection (11) and nonlinear [second-harmonic generation (12) and sum-frequency generation <

> Fig. 1. Structured water at the hydrophilic interface. The chlo.rine termination on a <> substrate forms a hydrophilic layer that orients the water bilayer. The closest packing dis.tance (4.43) be.tween oxygen atoms in the bottom layer of water is similar to the distance (4.50) be.tween the on-top and interstitial sites of the chlorine layer, result.ing in specific bilayer orientations (30) with respect to the silicon substrate. This ordered stacking persists for three to four bilayers (1 nm) before disorientation takes place andresults in crystallite islands, forming the layered structure. The size of atoms is not to scale for the van der Waals radii. <> <> <> <> <> <> Identity Mappings in Deep Residual Networks Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun Microsoft Research Abstract Deep residual networks [1] have emerged as a family of ex- tremely deep architectures showing compelling accuracy and nice con- vergence behaviors. In this paper, we analyze the propagation formu- lations behind the residual building blocks, which suggest that the for- ward and backward signals can be directly propagated from one block to any other block, when using identity mappings as the skip connec- tions and after-addition activation. A series of ablation experiments sup- port the importance of these identity mappings. This motivates us to propose a new residual unit, which makes training easier and improves generalization. We report improved results using a 1001-layer ResNet on CIFAR-10 (4.62% error) and CIFAR-100, and a 200-layer ResNet on ImageNet. Code is available at:https://github.com/KaimingHe/ resnet-1k-layers. 1 Introduction Deep residual networks (ResNets) [1] consist of many stacked \Residual Units". Each unit (Fig.1(a)) can be expressed in a general form: <> where xl and <> are input and output of the l-th unit, andFis a residual function. In [1],<> is an identity mapping and is a ReLU [2] function. ResNets that are over 100-layer deep have shown state-of-the-art accuracy for several challenging recognition tasks on ImageNet [3] and MS COCO [4] compe- titions. The central idea of ResNets is to learn the additive residual functionF with respect to <>, with a key choice of using an identity mapping <> . This is realized by attaching an identity skip connection shortcut. In this paper, we analyze deep residual networks by focusing on creating a direct path for propagating information not only within a residual unit, but through the entire network. Our derivations reveal that if both <> and <> are identity mappings, the signal could be directly propagated from one unit to any other units, in both forward and backward passes. Our experiments empirically show that training in general becomes easier when the architecture is closer to the above two conditions. To understand the role of skip connections, we analyze and compare various types of <>. We find that the identity mapping <> chosen in [1] <

> Figure 1. Left: (a) original Residual Unit in [1]; (b) proposed Residual Unit. The grey arrows indicate the easiest paths for the information to propagate, corresponding to the additive term \xl " in Eqn.(4) (forward propagation) and the additive term \1" in Eqn.(5) (backward propagation).Right: training curves on CIFAR-10 of1001-layer ResNets. Solid lines denote test error (y-axis on the right), and dashed lines denote training loss (y-axis on the left). The proposed unit makes ResNet-1001 easier to train. achieves the fastest error reduction and lowest training loss among all variants we investigated, whereas skip connections of scaling, gating [5,6,7], and 1x1 convolutions all lead to higher training loss and error. These experiments suggest that keeping a clean information path (indicated by the grey arrows in Fig.1,2, and4) is helpful for easing optimization. To construct an identity mapping <>, we view the activation func- tions (ReLU and BN [8]) as pre-activation of the weight layers, in contrast to conventional wisdom of post-activation. This point of view leads to a new residual unit design, shown in (Fig.1(b)). Based on this unit, we present com- petitive results on CIFAR-10/100 with a 1001-layer ResNet, which is much easier to train and generalizes better than the original ResNet in [1]. We further report improved results on ImageNet using a 200-layer ResNet, for which the counter- part of [1] starts to overfit. These results suggest that there is much room to exploit the dimension ofnetwork depth, a key to the success of modern deep learning. 2 Analysis of Deep Residual Networks The ResNets developed in [1] are modularized architectures that stack building blocks of the same connecting shape. In this paper we call these blocks \Residual 3 Units". The original Residual Unit in [1] performs the following computation: <>; (1) <>. (2) Here xl is the input feature to the l-th Residual Unit. <> is a set of weights (and biases) associated with the l-th Residual Unit, andKis the number of layers in a Residual Unit (Kis 2 or 3 in [1]). F denotes the residual function,e.g., a stack of two 3x3 convolutional layers in [1]. The function f is the operation after element-wise addition, and in [1] f is ReLU. The function h is set as an identity mapping:<> If f is also an identity mapping: <>, we can put Eqn.(2) into Eqn.(1) and obtain: <>. (3) Recursively <>, etc. we will have: <>; (4) for any deeper unit L and any shallower unit l. Eqn.(4) exhibits some nice properties. (i) The feature xL of any deeper unit L can be represented as the P feature xl of any shallower unit l plus a residual function in a form of <> indicating that the model is in a residual fashion between any units L and l. (ii)The feature <>, of any deep unit L, is the summation of the outputs of all preceding residual functions (<>). This is in contrast to Qa plain network here a feature xL is a series of matrix-vector products, say, <> (ignoring BN and ReLU). Eqn.(4) also leads to nice backward propagation properties. Denoting the loss function as E, from the chain rule of backpropagation [9] we have: <> (5) Eqn.(5) indicates that the gradient @E can be decomposed into two additive <> terms: a term of <> that propagates information directly without concerning any weight layers, and another term of <> that propagates <> through the weight layers. The additive term of @E ensures that information is directly propagated back to any shallower unIt l. Eqn.(5) also suggests that it is unlikely for the gradient @E to be canceled out for a mini-batch, because in general the term <> cannot be always -1 for all samples in a mini-batch. This implies that the gradient of a layer does not vanish even when the weights are arbitrarily small. 1 It is noteworthy that there are Residual Units for increasing dimensions and reducing feature map sizes [1] in which h is not identity. In this case the following derivations do not hold strictly. But as there are only a very few such units (two on CIFAR and three on ImageNet, depending on image sizes [1]), we expect that they do not have the exponential impact as we present in Sec.3. One may also think of our derivations as applied to all Residual Units within the same feature map size. Discussions Eqn.(4) and Eqn.(5) suggest that the signal can be directly propagated from any unit to another, both forward and backward. The foundation of Eqn.(4) is two identity mappings: (i) the identity skip connection <> , and (ii) the condition that f is an identity mapping. These directly propagated information flows are represented by the grey ar- rows in Fig.1,2, and4. And the above two conditions are true when these grey arrows cover no operations (expect addition) and thus are clean. In the fol- lowing two sections we separately investigate the impacts of the two conditions. 3 On the Importance of Identity Skip Connections Let’s consider a simple modification, <>, to break the identity shortcut: <>, (6) where l is a modulating scalar (for simplicity we still assume f is identity). Recursively applying this formulation we obtain an equation similar to Eqn. (4): <>, or simply: <>; (7) where the notationF^absorbs the scalars into the residual functions. Similar to Eqn.(5), we have backpropagation of the following form: <> (8) Unlike Eqn.(5), in Eqn.(8) the first additive term is modulated by a factor <> the factor can be exponentially large; if <> for all i, this factor can be exponentially small and vanish, which blocks the backpropagated signal from the shortcut and forces it to flow through the weight layers. This results in optimization difficulties as we show by experiments. In the above analysis, the original identity skip connection in Eqn.(3) is re- placed with a simple scaling <>. If the skip connection <> represents more complicated transforms (such as gating and 1x1 convolutions), in Eqn.(8) Q the first term becomes <> where h0 is the derivative of h. This product <> may also impede information propagation and hamper the training procedure as witnessed in the following experiments. <

> Figure 2.Various types of shortcut connections used in Table1. The grey arrows indicate the easiest paths for the information to propagate. The shortcut connections in (b-f) are impeded by different components. For simplifying illustrations we do not display the BN layers, which are adopted right after the weight layers for all units here. 3.1 Experiments on Skip Connections We experiment with the 110-layer ResNet as presented in [1] on CIFAR-10 [10]. This extremely deep ResNet-110 has 54 two-layer Residual Units (consisting of 3x3 convolutional layers) and is challenging for optimization. Our implementation details (see appendix) are the same as [1]. Throughout this paper we report the median accuracy of 5 runs for each architecture on CIFAR, reducing the impacts of random variations. Though our above analysis is driven by identity f, the experiments in this section are all based onf= ReLU as in [1]; we address identity f in the next section. Our baseline ResNet-110 has 6.61% error on the test set. The comparisons of other variants (Fig.2 and Table1) are summarized as follows: Constant scaling. We set <> for all shortcuts (Fig.2(b)). We further study two cases of scalingF: (i)Fis not scaled; or (ii)Fis scaled by a constant scalar of <>, which is similar to the highway gating [6,7] but with frozen gates. The former case does not converge well; the latter is able to converge, but the test error (Table1, 12.35%) is substantially higher than the original ResNet-110. Fig3(a) shows that the training error is higher than that of the original ResNet-110, suggesting that the optimization has difficulties when the shortcut signal is scaled down. 6 Table 1.Classification error on the CIFAR-10 test set using ResNet-110 [1], with different types of shortcut connections applied to all Residual Units. We report \fail" when the test error is higher than 20%. <

> Exclusive gating. Following the Highway Networks [6,7] that adopt a gating mechanism [5], we consider a gating function <> where a transform is represented by weights W g and biases <> followed by the sigmoid function <>. In a convolutional network <> is realized by a <> convolutional layer. The gating function modulates the signal by element-wise multiplication. We investigate the exclusive gates as used in [6,7] the F path is scaled byg(x) and the shortcut path is scaled by <>. See Fig2(c). We find that the initialization of the biases <> is critical for training gated models, and following the guidelines 2 in [6,7], we conduct hyper-parameter search on the initial value of <> in the range of 0 to -10 with a decrement step of -1 on the training set by cross- validation. The best value (6 here) is then used for training on the training set, leading to a test result of 8.70% (Table1), which still lags far behind the ResNet-110 baseline. Fig 3(b) shows the training curves. Table1also reports the results of using other initialized values, noting that the exclusive gating network does not converge to a good solution when <> is not appropriately initialized. The impact of the exclusive gating mechanism is two-fold. When <> approaches 1, the gated shortcut connections are closer to identity which helps information propagation; but in this case <> approaches 0 and suppresses the functionF. To isolate the effects of the gating functions on the shortcut path alone, we investigate a non-exclusive gating mechanism in the next. Shortcut-only gating. In this case the functionFis not scaled; only the shortcut path is gated by <>. See Fig2(d). The initialized value of<> is still essential in this case. When the initialized<> is 0 (so initially the expectation of <> is 0.5), the network converges to a poor result of 12.86% (Table1). This is also caused by higher training error (Fig 3(c)). <

> Figure 3.Training curves on CIFAR-10 of various shortcuts. Solid lines denote test error (y-axis on the right), and dashed lines denote training loss (y-axis on the left). When the initialized <> is very negatively biased (e.g.,6), the value of <> is closer to 1 and the shortcut connection is nearly an identity mapping. Therefore, the result (6.91%, Table1) is much closer to the ResNet-110 baseline. 1x1 convolutional shortcut. Next we experiment with 1x1 convolutional shortcut connections that replace the identity. This option has been investigated in [1] (known as option C) on a 34-layer ResNet (16 Residual Units) and shows good results, suggesting that 1x1 shortcut connections could be useful. But we find that this is not the case when there are many Residual Units. The 110-layer ResNet has a poorer result (12.22%, Table1) when using 1x1 convolutional shortcuts. Again, the training error becomes higher (Fig3(d)). When stacking so many Residual Units (54 for ResNet-110), even the shortest path may still impede signal propagation. We witnessed similar phenomena on ImageNet with ResNet-101 when using 1x1 convolutional shortcuts. Dropout shortcut. Last we experiment with dropout [11] (at a ratio of 0.5) which we adopt on the output of the identity shortcut (Fig.2(f)). The network fails to converge to a good solution. Dropout statistically imposes a scale of with an expectation of 0.5 on the shortcut, and similar to constant scaling by 0.5, it impedes signal propagation. Table 2.Classification error (%) on the CIFAR-10 test set using different activation functions. <

> <

> Figure 4.Various usages of activation in Table2. All these units consist of the same components | only the orders are different. 3.2 Discussions As indicated by the grey arrows in Fig.2, the shortcut connections are the most direct paths for the information to propagate.Multiplicative manipulations (scaling, gating, 1x1 convolutions, and dropout) on the shortcuts can hamper information propagation and lead to optimization problems. It is noteworthy that the gating and 1x1 convolutional shortcuts introduce more parameters, and should have stronger representational abilities than identity shortcuts. In fact, the shortcut-only gating and 1x1 convolution cover the solution space of identity shortcuts (i.e., they could be optimized as identity shortcuts). However, their training error is higher than that of identity short- cuts, indicating that the degradation of these models is caused by optimization issues, instead of representational abilities. 4 On the Usage of Activation Functions Experiments in the above section support the analysis in Eqn.(5) and Eqn.(8), both being derived under the assumption that the after-addition activation f 9 is the identity mapping. But in the above experiments f is ReLU as designed in [1], so Eqn.(5) and (8) are approximate in the above experiments. Next we investigate the impact off. We want to make f an identity mapping, which is done by re-arranging the activation functions (ReLU and/or BN). The original Residual Unit in [1] has a shape in Fig.4(a) | BN is used after each weight layer, and ReLU is adopted after BN except that the last ReLU in a Residual Unit is after element- wise addition (f= ReLU). Fig.4(b-e) show the alternatives we investigated, explained as following. 4.1 Experiments on Activation In this section we experiment with ResNet-110 and a 164-layerBottleneck[1] architecture (denoted as ResNet-164). A bottleneck Residual Unit consist of a 1x1 layer for reducing dimension, a 3x3 layer, and a 1x1 layer for restoring dimension. As designed in [1], its computational complexity is similar to the two-3x3 Residual Unit. More details are in the appendix. The baseline ResNet- 164 has a competitive result of 5.93% on CIFAR-10 (Table2). BN after addition. Before turning f into an identity mapping, we go the opposite way by adopting BN after addition (Fig.4(b)). In this case f involves BN and ReLU. The results become considerably worse than the baseline (Ta- ble2). Unlike the original design, now the BN layer alters the signal that passes through the shortcut and impedes information propagation, as reflected by the difficulties on reducing training loss at the beginning of training (Fib.6left). ReLU before addition. A naive choice of making f into an identity map- ping is to move the ReLU before addition (Fig.4(c)). However, this leads to a non-negative output from the transformF, while intuitively a residual function should take values in (-1,+1). As a result, the forward propagated signal is monotonically increasing. This may impact the representational ability, and the result is worse (7.84%, Table2) than the baseline. We expect to have a residual function taking values in (-1,+1). This condition is satisfied by other Residual Units including the following ones. Post-activation or pre-activation?In the original design (Eqn.(1) and Eqn.(2)), the activation<> affects both paths in the next Residual Unit: <>. Next we develop an asymmetric form where an activation f only affects the F path: <>, for any l(Fig.5(a) to (b)). By renaming the notations, we have the following form: <>, (9) It is easy to see that Eqn.(9) is similar to Eqn.(4), and can enable a backward formulation similar to Eqn.(5). For this new Residual Unit as in Eqn.(9), the new after-addition activation becomes an identity mapping. This design means that if a new after-addition activation f is asymmetrically adopted, it is equivalent to recasting f as the pre-activation of the next Residual Unit. This is illustrated in Fig.5. <

> Figure 5.Using asymmetric after-addition activation is equivalent to constructing a pre-activationResidual Unit. Table 3.Classification error (%) on the CIFAR-10/100 test set using the original Residual Units and our pre-activation Residual Units. <

> The distinction between post-activation/pre-activation is caused by the presence of the element-wise addition. For a plain network that has N layers, there are N-1 activations (BN/ReLU), and it does not matter whether we think of them as post- or pre-activations. But for branched layers merged by addition, the position of activation matters. We experiment with two such designs: (i) ReLU-only pre-activation (Fig.4(d)), and (ii) full pre-activation (Fig.4(e)) where BN and ReLU are both adopted be- fore weight layers. Table2 shows that the ReLU-only pre-activation performs very similar to the baseline on ResNet-110/164. This ReLU layer is not used in conjunction with a BN layer, and may not enjoy the benefits of BN [8]. Somehow surprisingly, when BN and ReLU are both used as pre-activation, the results are improved by healthy margins (Table2and Table3). In Table3we report results using various architectures: (i) ResNet-110, (ii) ResNet-164, (iii) a 110-layer ResNet architecture in which each shortcut skips only 1 layer (i.e., 11 <

> Figure 6.Training curves on CIFAR-10.Left: BN after addition (Fig.4(b)) using ResNet-110.Right: pre-activation unit (Fig.4(e)) on ResNet-164. Solid lines denote test error, and dashed lines denote training loss. a Residual Unit has only 1 layer, denoted as ResNet-110 (1layer)), and (iv) a 1001-layer bottleneck architecture that has 333 Residual Units (111 on each feature map size), denoted as \ResNet-1001". We also experiment on CIFAR- 100. Table3shows that our pre-activation models are consistently better than the baseline counterparts. We analyze these results in the following. 4.2 Analysis We find the impact of pre-activation is twofold. First, the optimization is further eased (comparing with the baseline ResNet) because f is an identity mapping. Second, using BN as pre-activation improves regularization of the models. Ease of optimization. This effect is particularly obvious when training the1001-layerResNet. Fig.1shows the curves. Using the original design in [1], the training error is reduced very slowly at the beginning of training. For f= ReLU, the signal is impacted if it is negative, and when there are many Residual Units, this effect becomes prominent and Eqn.(3) (so Eqn.(5)) is not a good approximation. On the other hand, when f is an identity mapping, the signal can be propagated directly between any two units. Our 1001-layer network reduces the training loss very quickly (Fig.1). It also achieves the lowest loss among all models we investigated, suggesting the success of optimization. We also find that the impact off= ReLU is not severe when the ResNet has fewer layers (e.g., 164 in Fig.6(right)). The training curve seems to suffer a little bit at the beginning of training, but goes into a healthy status soon. By monitoring the responses we observe that this is because after some training, the weights are adjusted into a status such that yl in Eqn.(1) is more frequently above zero and f does not truncate it (xl is always non-negative due to the previous ReLU, so yl is below zero only when the magnitude ofFis very negative). The truncation, however, is more frequent when there are 1000 layers. Table 4.Comparisons with state-of-the-art methods on CIFAR-10 and CIFAR-100 using \moderate data augmentation" (ip/translation), except for ELU [12] with no augmentation. Better results of [13,14] have been reported using stronger data augmen- tation and ensembling. For the ResNets we also report the number of parameters. Our results are the median of 5 runs with meanstd in the brackets. All ResNets results are obtained with a mini-batch size of 128 except y with a mini-batch size of 64 (code available athttps://github.com/KaimingHe/resnet-1k-layers). <

> Reducing overfitting. Another impact of using the proposed pre-activation unit is on regularization, as shown in Fig.6(right). The pre-activation ver- sion reaches slightly higher training loss at convergence, but produces lower test error. This phenomenon is observed on ResNet-110, ResNet-110(1-layer), and ResNet-164 on both CIFAR-10 and 100. This is presumably caused by BN’s reularization effect [8]. In the original Residual Unit (Fig.4(a)), although the BN normalizes the signal, this is soon added to the shortcut and thus the merged signal is not normalized. This unnormalized signal is then used as the input of the next weight layer. On the contrary, in our pre-activation version, the inputs to all weight layers have been normalized. 5 Results Comparisons on CIFAR-10/100.Table4compares the state-of-the-art meth- ods on CIFAR-10/100, where we achieve competitive results. We note that we do not specially tailor the network width or filter sizes, nor use regularization techniques (such as dropout) which are very effective for these small datasets. We obtain these results via a simple but essential concept | going deeper. These results demonstrate the potential of pushing the limits of depth. Comparisons on ImageNet.Next we report experimental results on the 1000- class ImageNet dataset [3]. We have done preliminary experiments using the skip connections studied in Fig.2&3on ImageNet with ResNet-101 [1], and observed similar optimization difficulties. The training error of these non-identity shortcut networks is obviously higher than the original ResNet at the first learning rate 13 Table 5.Comparisons of single-crop error on the ILSVRC 2012 validation set. All ResNets are trained using the same hyper-parameters and implementations as [1]). Our Residual Units are the full pre-activation version (Fig.4(e)). y : code/model avail- able athttps://github.com/facebook/fb.resnet.torch/tree/master/pretrained, using scale and aspect ratio augmentation in [20]. <> (similar to Fig.3), and we decided to halt training due to limited resources. But we did finish a BN after addition version (Fig.4(b)) of ResNet-101 on ImageNet and observed higher training loss and validation error. This model’s single-crop (224x224) validation error is 24.6%/7.5%,vs.the original ResNet- 101’s 23.6%/7.1%. This is in line with the results on CIFAR in Fig.6(left). Table5shows the results of ResNet-152 [1] and ResNet-200 3 , all trained from scratch. We notice that the original ResNet paper [1] trained the models using scale jittering with shorter sides [256;480], and so the test of a 224x224 crop ons= 256 (as did in [1]) is negatively biased. Instead, we test a single 320x320 crop from s=320, for all original and our ResNets. Even though the ResNets are trained on smaller crops, they can be easily tested on larger crops because the ResNets are fully convolutional by design. This size is also close to 299x299 used by Inception v3 [19], allowing a fairer comparison. The original ResNet-152 [1] has top-1 error of 21.3% on a 320x320 crop, and our pre-activation counterpart has 21.1%. The gain is not big on ResNet-152 because this model has not shown severe generalization difficulties. However, the original ResNet-200 has an error rate of 21.8%, higher than the baseline ResNet-152. But we find that the original ResNet-200 has lower training error than ResNet-152, suggesting that it suffers from overfitting. Our pre-activation ResNet-200 has an error rate of 20.7%, which is1.1% lower than the baseline ResNet-200 and also lower than the two versions of ResNet-152. When using the scale and aspect ratio augmentation of [20,19], our ResNet-200 has a result better than Inception v3 [19] (Table5). Concurrent with our work, an Inception-ResNet-v2 model [21] achieves a single-crop result of 19.9%/4.9%. We expect our observations and the proposed Residual Unit will help this type and generally other types of ResNets. Computational Cost.Our models’ computational complexity is linear on 3 The ResNet-200 has 16 more 3-layer bottleneck Residual Units than ResNet-152, which are added on the feature map of 28x28. depth (so a 1001-layer net is complex of a 100-layer net). On CIFAR, ResNet-1001 takes about 27 hours to train on 2 GPUs; on ImageNet, ResNet- 200 takes about 3 weeks to train on 8 GPUs (on par with VGG nets [22]). 6 Conclusions This paper investigates the propagation formulations behind the connection mechanisms of deep residual networks. Our derivations imply that identity short- cut connections and identity after-addition activation are essential for making information propagation smooth. Ablation experiments demonstrate phenom- ena that are consistent with our derivations. We also present 1000-layer deep networks that can be easily trained and achieve improved accuracy. Appendix: Implementation DetailsThe implementation details and hyper- parameters are the same as those in [1]. On CIFAR we use only the translation and skipping augmentation in [1] for training. The learning rate starts from 0.1, and is divided by 10 at 32k and 48k iterations. Following [1], for all CIFAR experiments we warm up the training by using a smaller learning rate of 0.01 at the beginning 400 iterations and go back to 0.1 after that, although we remark that this is not necessary for our proposed Residual Unit. The mini-batch size is 128 on 2 GPUs (64 each), the weight decay is 0.0001, the momentum is 0.9, and the weights are initialized as in [23]. On ImageNet, we train the models using the same data augmentation as in [1]. The learning rate starts from 0.1 (no warming up), and is divided by 10 at 30 and 60 epochs. The mini-batch size is 256 on 8 GPUs (32 each). The weight decay, momentum, and weight initialization are the same as above. When using the pre-activation Residual Units (Fig.4(d)(e) and Fig.5), we pay special attention to the first and the last Residual Units of the entire net- work. For the first Residual Unit (that follows a stand-alone convolutional layer, conv 1 ), we adopt the first activation right after conv 1 and before splitting into two paths; for the last Residual Unit (followed by average pooling and a fully- connected classifier), we adopt an extra activation right after its element-wise addition. These two special cases are the natural outcome when we obtain the pre-activation network via the modification procedure as shown in Fig.5. The bottleneck Residual Units (for ResNet-164/1001 on CIFAR) are constructed following [1]. For example, a 3x3, 16 unit in ResNet-110 is replaced 3x3, 162 with a 1x1, 166 7 unit in ResNet-164, both of which have roughly the same 3x3, 165 1x1, 64 number of parameters. For the bottleneck ResNets, when reducing the feature map size we use projection shortcuts [1] for increasing dimensions, and when pre- activation is used, these projection shortcuts are also with pre-activation. 15 References 1.He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. (2016) 2.Nair, V., Hinton, G.E.: Rectied linear units improve restricted boltzmann ma- chines. In: ICML. (2010) 3.Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. IJCV (2015) 4.Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV. (2014) 5.Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation (1997) 6.Srivastava, R.K., Gre, K., Schmidhuber, J.: Highway networks. In: ICML work- shop. (2015) 7.Srivastava, R.K., Gre, K., Schmidhuber, J.: Training very deep networks. In: NIPS. (2015) 8.Ioe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: ICML. (2015) 9.LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W., Jackel, L.D.: Backpropagation applied to handwritten zip code recognition. Neural computation (1989) 10.Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech Report (2009) 11.Hinton, G.E., Srivastava, N., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.R.: Improving neural networks by preventing co-adaptation of feature detectors. arXiv:1207.0580 (2012) 12.Clevert, D.A., Unterthiner, T., Hochreiter, S.: Fast and accurate deep network learning by exponential linear units (ELUs). In: ICLR. (2016) 13.Graham, B.: Fractional max-pooling. arXiv:1412.6071 (2014) 14.Springenberg, J.T., Dosovitskiy, A., Brox, T., Riedmiller, M.: Striving for simplic- ity: The all convolutional net. arXiv:1412.6806 (2014) 15.Lin, M., Chen, Q., Yan, S.: Network in network. In: ICLR. (2014) 16.Lee, C.Y., Xie, S., Gallagher, P., Zhang, Z., Tu, Z.: Deeply-supervised nets. In: AISTATS. (2015) 17.Romero, A., Ballas, N., Kahou, S.E., Chassang, A., Gatta, C., Bengio, Y.: Fitnets: Hints for thin deep nets. In: ICLR. (2015) 18.Mishkin, D., Matas, J.: All you need is a good init. In: ICLR. (2016) 19.Szegedy, C., Vanhoucke, V., Ioe, S., Shlens, J., Wojna, Z.: Rethinking the incep- tion architecture for computer vision. In: CVPR. (2016) 20.Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: CVPR. (2015) 21.Szegedy, C., Ioe, S., Vanhoucke, V.: Inception-v4, inception-resnet and the impact of residual connections on learning. arXiv:1602.07261 (2016) 22.Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR. (2015) 23.He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectiers: Surpassing human- level performance on imagenet Classification. In: ICCV. (2015) <> <> <> <> <> <> Language Models are Few-Shot Learners Tom B. Brown Benjamin Mann Nick Ryder Melanie Subbiah Jared Kaplan y Prafulla Dhariwal Arvind Neelakantan Pranav Shyam Girish Sastry Amanda Askell Sandhini Agarwal Ariel Herbert-Voss Gretchen Krueger Tom Henighan Rewon Child Aditya Ramesh Daniel M. Ziegler Jeffrey Wu Clemens Winter Christopher Hesse Mark Chen Eric Sigler Mateusz Litwin Scott Gray Benjamin Chess Jack Clark Christopher Berner Sam McCandlish Alec Radford Ilya Sutskever Dario Amodei OpenAI Abstract Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task. While typically task-agnostic in architecture, this method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples. By contrast, humans can generally perform a new language task from only a few examples or from simple instructions – something which current NLP systems still largely struggle to do. Here we show that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even reaching competitiveness with prior state-of-the-art fine- tuning approaches. Specifically, we train GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, and test its performance in the few-shot setting. For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model. GPT-3 achieves strong performance on many NLP datasets, including translation, question-answering, and close tasks, as well as several tasks that require on-the-fly reasoning or domain adaptation, such as unscrambling words, using a novel word in a sentence, or performing 3-digit arithmetic. At the same time, we also identify some datasets where GPT-3’s few-shot learning still struggles, as well as some datasets where GPT-3 faces methodological issues related to training on large web corpora. Finally, we find that GPT-3 can generate samples of news articles which human evaluators have difficulty distinguishing from articles written by humans. We discuss broader societal impacts of this finding and of GPT-3 in general. Equal contribution y Johns Hopkins University, OpenAI Contents 1 Introduction 3 2 Approach 6 2.1 Model and Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8 2.2 Training Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8 2.3 Training Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9 2.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10 3 Results 10 3.1 Language Modeling, Cloze, and Completion Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . .11 3.2 Closed Book Question Answering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13 3.3 Translation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14 3.4 Winograd-Style Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .16 3.5 Common Sense Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .17 3.6 Reading Comprehension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .18 3.7 SuperGLUE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .18 3.8 NLI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .20 3.9 Synthetic and Qualitative Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .21 4 Measuring and Preventing Memorization Of Benchmarks29 5 Limitations 33 6 Broader Impacts 34 6.1 Misuse of Language Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .35 6.2 Fairness, Bias, and Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .36 6.3 Energy Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .39 7 Related Work 39 8 Conclusion 40 A Details of Common Crawl Filtering43 B Details of Model Training 43 C Details of Test Set Contamination Studies43 D Total Compute Used to Train Language Models46 E Human Quality Assessment of Synthetic News Articles46 F Additional Samples from GPT-348 G Details of Task Phrasing and Specifications50 H Results on All Tasks for All Model Sizes63 1 Introduction Recent years have featured a trend towards pre-trained language representations in NLP systems, applied in increasingly flexible and task-agnostic ways for downstream transfer. First, single-layer representations were learned using word vectors [MCCD13,PSM14] and fed to task-specific architectures, then RNNs with multiple layers of representations and contextual state were used to form stronger representations [DL15,MBXS17,PNZtY18] (though still applied to task-specific architectures), and more recently pre-trained recurrent or transformer language models [VSP + 17] have been directly fine-tuned, entirely removing the need for task-specific architectures [RNSS18,DCLT18,HR18]. This last paradigm has led to substantial progress on many challenging NLP tasks such as reading comprehension, question answering, textual entailment, and many others, and has continued to advance based on new architectures and algorithms [RSR + 19,LOG + 19,YDY + 19,LCG + 19]. However, a major limitation to this approach is that while the architecture is task-agnostic, there is still a need for task-specific datasets and task-specific fine-tuning: to achieve strong performance on a desired task typically requires fine-tuning on a dataset of thousands to hundreds of thousands of examples specific to that task. Removing this limitation would be desirable, for several reasons. First, from a practical perspective, the need for a large dataset of labeled examples for every new task limits the applicability of language models. There exists a very wide range of possible useful language tasks, encompassing anything from correcting grammar, to generating examples of an abstract concept, to critiquing a short story. For many of these tasks it is difficult to collect a large supervised training dataset, especially when the process must be repeated for every new task. Second, the potential to exploit spurious correlations in training data fundamentally grows with the expressiveness of the model and the narrowness of the training distribution. This can create problems for the pre-training plus fine-tuning paradigm, where models are designed to be large to absorb information during pre-training, but are then fine-tuned on very narrow task distributions. For instance [HLW + 20] observe that larger models do not necessarily generalize better out-of-distribution. There is evidence that suggests that the generalization achieved under this paradigm can be poor because the model is overly specific to the training distribution and does not generalize well outside it [YdC + 19,MPL19]. Thus, the performance of fine-tuned models on specific benchmarks, even when it is nominally at human-level, may exaggerate actual performance on the underlying task [GSL + 18,NK19]. Third, humans do not require large supervised datasets to learn most language tasks – a brief directive in natural language (e.g. “please tell me if this sentence describes something happy or something sad”) or at most a tiny number of demonstrations (e.g. “here are two examples of people acting brave; please give a third example of bravery”) is often <

> Figure 1.1: Language model meta-learning.During unsupervised pre-training, a language model develops a broad set of skills and pattern recognition abilities. It then uses these abilities at inference time to rapidly adapt to or recognize the desired task. We use the term “in-context learning” to describe the inner loop of this process, which occurs within the forward-pass upon each sequence. The sequences in this diagram are not intended to be representative of the data a model would see during pre-training, but are intended to show that there are sometimes repeated sub-tasks embedded within a single sequence. <

> Figure 1.2: Larger models make increasingly efficient use of in-context information. We show in-context learning performance on a simple task requiring the model to remove random symbols from a word, both with and without a natural language task description (see Sec.3.9.2). The steeper “in-context learning curves” for large models demonstrate improved ability to learn a task from contextual information. We see qualitatively similar behavior across a wide range of tasks. sufficient to enable a human to perform a new task to at least a reasonable degree of competence. Aside from pointing to a conceptual limitation in our current NLP techniques, this adaptability has practical advantages – it allows humans to seamlessly mix together or switch between many tasks and skills, for example performing addition during a lengthy dialogue. To be broadly useful, we would someday like our NLP systems to have this same fluidity and generality. One potential route towards addressing these issues is meta-learning 1 – which in the context of language models means the model develops a broad set of skills and pattern recognition abilities at training time, and then uses those abilities at inference time to rapidly adapt to or recognize the desired task (illustrated in Figure1.1). Recent work [RWC + 19] attempts to do this via what we call “in-context learning”, using the text input of a pretrained language model as a form of task specification: the model is conditioned on a natural language instruction and/or a few demonstrations of the task and is then expected to complete further instances of the task simply by predicting what comes next. While it has shown some initial promise, this approach still achieves results far inferior to fine-tuning – for example [RWC + 19] achieves only 4% on Natural Questions, and even its 55 F1 CoQa result is now more than 35 points behind the state of the art. Meta-learning clearly requires substantial improvement in order to be viable as a practical method of solving language tasks. Another recent trend in language modeling may offer a way forward. In recent years the capacity of transformer language models has increased substantially, from 100 million parameters [RNSS18], to 300 million parameters [DCLT18], to 1.5 billion parameters [RWC + 19], to 8 billion parameters [SPP + 19], 11 billion parameters [RSR + 19], and finally 17 billion parameters [Tur20]. Each increase has brought improvements in text synthesis and/or downstream NLP tasks, and there is evidence suggesting that log loss, which correlates well with many downstream tasks, follows a smooth trend of improvement with scale [KMH + 20]. Since in-context learning involves absorbing many skills and tasks within the parameters of the model, it is plausible that in-context learning abilities might show similarly strong gains with scale. 1 In the context of language models this has sometimes been called “zero-shot transfer”, but this term is potentially ambiguous: the method is “zero-shot” in the sense that no gradient updates are performed, but it often involves providing inference-time demonstrations to the model, so is not truly learning from zero examples. To avoid this confusion, we use the term “meta-learning” to capture the inner-loop / outer-loop structure of the general method, and the term “in context-learning” to refer to the inner loop of meta-learning. We further specialize the description to “zero-shot”, “one-shot”, or “few-shot” depending on how many demonstrations are provided at inference time. These terms are intended to remain agnostic on the question of whether the model learns new tasks from scratch at inference time or simply recognizes patterns seen during training – this is an important issue which we discuss later in the paper, but “meta-learning” is intended to encompass both possibilities, and simply describes the inner-outer loop structure. <

> Figure 1.3: Aggregate performance for all 42 accuracy-denominated benchmarks While zero-shot performance improves steadily with model size, few-shot performance increases more rapidly, demonstrating that larger models are more proficient at in-context learning. See Figure3.8for a more detailed analysis on SuperGLUE, a standard NLP benchmark suite. In this paper, we test this hypothesis by training a 175 billion parameter autoregressive language model, which we call GPT-3, and measuring its in-context learning abilities. Specifically, we evaluate GPT-3 on over two dozen NLP datasets, as well as several novel tasks designed to test rapid adaptation to tasks unlikely to be directly contained in the training set. For each task, we evaluate GPT-3 under 3 conditions: (a) “few-shot learning”, or in-context learning where we allow as many demonstrations as will fit into the model’s context window (typically 10 to 100), (b) “one-shot learning”, where we allow only one demonstration, and (c) “zero-shot” learning, where no demonstrations are allowed and only an instruction in natural language is given to the model. GPT-3 could also in principle be evaluated in the traditional fine-tuning setting, but we leave this to future work. Figure1.2illustrates the conditions we study, and shows few-shot learning of a simple task requiring the model to remove extraneous symbols from a word. Model performance improves with the addition of a natural language task description, and with the number of examples in the model’s context,K. Few-shot learning also improves dramatically with model size. Though the results in this case are particularly striking, the general trends with both model size and number of examples in-context hold for most tasks we study. We emphasize that these “learning” curves involve no gradient updates or fine-tuning, just increasing numbers of demonstrations given as conditioning. Broadly, on NLP tasks GPT-3 achieves promising results in the zero-shot and one-shot settings, and in the the few-shot setting is sometimes competitive with or even occasionally surpasses state-of-the-art (despite state-of-the-art being held by fine-tuned models). For example, GPT-3 achieves 81.5 F1 on CoQA in the zero-shot setting, 84.0 F1 on CoQA in the one-shot setting, 85.0 F1 in the few-shot setting. Similarly, GPT-3 achieves 64.3% accuracy on TriviaQA in the zero-shot setting, 68.0% in the one-shot setting, and 71.2% in the few-shot setting, the last of which is state-of-the-art relative to fine-tuned models operating in the same closed-book setting. GPT-3 also displays one-shot and few-shot proficiency at tasks designed to test rapid adaption or on-the-fly reasoning, which include unscrambling words, performing arithmetic, and using novel words in a sentence after seeing them defined only once. We also show that in the few-shot setting, GPT-3 can generate synthetic news articles which human evaluators have difficulty distinguishing from human-generated articles. At the same time, we also find some tasks on which few-shot performance struggles, even at the scale of GPT-3. This includes natural language inference tasks like the ANLI dataset, and some reading comprehension datasets like RACE or QuAC. By presenting a broad characterization of GPT-3’s strengths and weaknesses, including these limitations, we hope to stimulate study of few-shot learning in language models and draw attention to where progress is most needed. A heuristic sense of the overall results can be seen in Figure1.3, which aggregates the various tasks (though it should not be seen as a rigorous or meaningful benchmark in itself). We also undertake a systematic study of “data contamination” – a growing problem when training high capacity models on datasets such as Common Crawl, which can potentially include content from test datasets simply because such content often exists on the web. In this paper we develop systematic tools to measure data contamination and quantify its distorting effects. Although we find that data contamination has a minimal effect on GPT-3’s performance on most datasets, we do identify a few datasets where it could be inflating results, and we either do not report results on these datasets or we note them with an asterisk, depending on the severity. In addition to all the above, we also train a series of smaller models (ranging from 125 million parameters to 13 billion parameters) in order to compare their performance to GPT-3 in the zero, one and few-shot settings. Broadly, for most tasks we find relatively smooth scaling with model capacity in all three settings; one notable pattern is that the gap between zero-, one-, and few-shot performance often grows with model capacity, perhaps suggesting that larger models are more proficient meta-learners. Finally, given the broad spectrum of capabilities displayed by GPT-3, we discuss concerns about bias, fairness, and broader societal impacts, and attempt a preliminary analysis of GPT-3’s characteristics in this regard. The remainder of this paper is organized as follows. In Section2, we describe our approach and methods for training GPT-3 and evaluating it. Section3presents results on the full range of tasks in the zero-, one- and few-shot settings. Section4addresses questions of data contamination (train-test overlap). Section5discusses limitations of GPT-3. Section6discusses broader impacts. Section7reviews related work and Section8concludes. 2 Approach Our basic pre-training approach, including model, data, and training, is similar to the process described in [RWC + 19], with relatively straightforward scaling up of the model size, dataset size and diversity, and length of training. Our use of in-context learning is also similar to [RWC + 19], but in this work we systematically explore different settings for learning within the context. Therefore, we start this section by explicitly defining and contrasting the different settings that we will be evaluating GPT-3 on or could in principle evaluate GPT-3 on. These settings can be seen as lying on a spectrum of how much task-specific data they tend to rely on. Specifically, we can identify at least four points on this spectrum (see Figure2.1for an illustration): •Fine-Tuning (FT)has been the most common approach in recent years, and involves updating the weights of a pre-trained model by training on a supervised dataset specific to the desired task. Typically thousands to hundreds of thousands of labeled examples are used. The main advantage of fine-tuning is strong performance on many benchmarks. The main disadvantages are the need for a new large dataset for every task, the potential for poor generalization out-of-distribution [MPL19], and the potential to exploit spurious features of the training data [GSL + 18,NK19], potentially resulting in an unfair comparison with human performance. In this work we do not fine-tune GPT-3 because our focus is on task-agnostic performance, but GPT-3 can be fine-tuned in principle and this is a promising direction for future work. •Few-Shot (FS)is the term we will use in this work to refer to the setting where the model is given a few demonstrations of the task at inference time as conditioning [RWC + 19], but no weight updates are allowed. As shown in Figure2.1, for a typical dataset an example has a context and a desired completion (for example an English sentence and the French translation), and few-shot works by giving K examples of context and completion, and then one final example of context, with the model expected to provide the completion. We typically setKin the range of 10 to 100 as this is how many examples can fit in the model’s context window (nctx = 2048). The main advantages of few-shot are a major reduction in the need for task-specific data and reduced potential to learn an overly narrow distribution from a large but narrow fine-tuning dataset. The main disadvantage is that results from this method have so far been much worse than state-of-the-art fine-tuned models. Also, a small amount of task specific data is still required. As indicated by the name, few-shot learning as described here for language models is related to few-shot learning as used in other contexts in ML [HYC01,VBL + 16] – both involve learning based on a broad distribution of tasks (in this case implicit in the pre-training data) and then rapidly adapting to a new task. •One-Shot (1S) is the same as few-shot except that only one demonstration is allowed, in addition to a natural language description of the task, as shown in Figure 1. The reason to distinguish one-shot from few-shot and zero-shot (below) is that it most closely matches the way in which some tasks are communicated to humans. For example, when asking humans to generate a dataset on a human worker service (for example Mechanical Turk), it is common to give one demonstration of the task. By contrast it is sometimes difficult to communicate the content or format of a task if no examples are given. <

> Figure 2.1: Zero-shot, one-shot and few-shot, contrasted with traditional fine-tuning. The panels above show four methods for performing a task with a language model – fine-tuning is the traditional method, whereas zero-, one-, and few-shot, which we study in this work, require the model to perform the task with only forward passes at test time. We typically present the model with a few dozen examples in the few shot setting. Exact phrasings for all task descriptions, examples and prompts can be found in AppendixG. •Zero-Shot (0S)is the same as one-shot except that no demonstrations are allowed, and the model is only given a natural language instruction describing the task. This method provides maximum convenience, potential for robustness, and avoidance of spurious correlations (unless they occur very broadly across the large corpus of pre-training data), but is also the most challenging setting. In some cases it may even be difficult for humans to understand the format of the task without prior examples, so this setting is in some cases “unfairly hard”. For example, if someone is asked to “make a table of world records for the 200m dash”, this request can be ambiguous, as it may not be clear exactly what format the table should have or what should be included (and even with careful clarification, understanding precisely what is desired can be difficult). Nevertheless, for at least some settings zero-shot is closest to how humans perform tasks – for example, in the translation example in Figure2.1, a human would likely know what to do from just the text instruction. Figure2.1shows the four methods using the example of translating English to French. In this paper we focus on zero-shot, one-shot and few-shot, with the aim of comparing them not as competing alternatives, but as different problem settings which offer a varying trade-off between performance on specific benchmarks and sample efficiency. We especially highlight the few-shot results as many of them are only slightly behind state-of-the-art fine-tuned models. Ultimately, however, one-shot, or even sometimes zero-shot, seem like the fairest comparisons to human performance, and are important targets for future work. Sections2.1-2.3below give details on our models, training data, and training process respectively. Section2.4discusses the details of how we do few-shot, one-shot, and zero-shot evaluations. <

> Table 2.1:Sizes, architectures, and learning hyper-parameters (batch size in tokens and learning rate) of the models which we trained. All models were trained for a total of 300 billion tokens. 2.1 Model and Architectures We use the same model and architecture as GPT-2 [RWC + 19], including the modified initialization, pre-normalization, and reversible tokenization described therein, with the exception that we use alternating dense and locally banded sparse attention patterns in the layers of the transformer, similar to the Sparse Transformer [CGRS19]. To study the dependence of ML performance on model size, we train 8 different sizes of model, ranging over three orders of magnitude from 125 million parameters to 175 billion parameters, with the last being the model we call GPT-3. Previous work [KMH + 20] suggests that with enough training data, scaling of validation loss should be approximately a smooth power law as a function of size; training models of many different sizes allows us to test this hypothesis both for validation loss and for downstream language tasks. Table2.1shows the sizes and architectures of our 8 models. Here n params is the total number of trainable parameters, n layers is the total number of layers,d model is the number of units in each bottleneck layer (we always have the feedforward layer four times the size of the bottleneck layer,<> model ), and d head is the dimension of each attention head. All models use a context window of <> tokens. We partition the model across GPUs along both the depth and width dimension in order to minimize data-transfer between nodes. The precise architectural parameters for each model are chosen based on computational efficiency and load-balancing in the layout of models across GPU’s. Previous work [KMH + 20] suggests that validation loss is not strongly sensitive to these parameters within a reasonably broad range. 2.2 Training Dataset Datasets for language models have rapidly expanded, culminating in the Common Crawl dataset 2 [RSR + 19] constituting nearly a trillion words. This size of dataset is sufficient to train our largest models without ever updating on the same sequence twice. However, we have found that unfiltered or lightly filtered versions of Common Crawl tend to have lower quality than more curated datasets. Therefore, we took 3 steps to improve the average quality of our datasets: (1) we downloaded and filtered a version of CommonCrawl based on similarity to a range of high-quality reference corpora, (2) we performed fuzzy de-duplication at the document level, within and across datasets, to prevent redundancy and preserve the integrity of our held-out validation set as an accurate measure of overfitting, and (3) we also added known high-quality reference corpora to the training mix to augment CommonCrawl and increase its diversity. Details of the first two points (processing of Common Crawl) are described in AppendixA. For the third, we added several curated high-quality datasets, including an expanded version of the WebText dataset [RWC + 19], collected by scraping links over a longer period of time, and first described in [KMH + 20], two internet-based books corpora (Books1 and Books2) and English-language Wikipedia. Table2.2shows the final mixture of datasets that we used in training. The CommonCrawl data was downloaded from 41 shards of monthly CommonCrawl covering 2016 to 2019, constituting 45TB of compressed plaintext before filtering and 570GB after filtering, roughly equivalent to 400 billion byte-pair-encoded tokens. Note that during training, datasets are not sampled in proportion to their size, but rather datasets we view as higher-quality are sampled more frequently, such that CommonCrawl and Books2 datasets are sampled less than once during training, but the other datasets are sampled 2-3 times. This essentially accepts a small amount of overfitting in exchange for higher quality training data. <

> Figure 2.2: Total compute used during training. Based on the analysis in Scaling Laws For Neural Language Models [KMH + 20] we train much larger models on many fewer tokens than is typical. As a consequence, although GPT-3 3B is almost 10x larger than RoBERTa-Large (355M params), both models took roughly 50 petaflop/s-days of compute during pre-training. Methodology for these calculations can be found in AppendixD. <

> Table 2.2: Datasets used to train GPT-3. “Weight in training mix” refers to the fraction of examples during training that are drawn from a given dataset, which we intentionally do not make proportional to the size of the dataset. As a result, when we train for 300 billion tokens, some datasets are seen up to 3.4 times during training while other datasets are seen less than once. A major methodological concern with language models pretrained on a broad swath of internet data, particularly large models with the capacity to memorize vast amounts of content, is potential contamination of downstream tasks by having their test or development sets inadvertently seen during pre-training. To reduce such contamination, we searched for and attempted to remove any overlaps with the development and test sets of all benchmarks studied in this paper. Unfortunately, a bug in the filtering caused us to ignore some overlaps, and due to the cost of training it was not feasible to retrain the model. In Section4we characterize the impact of the remaining overlaps, and in future work we will more aggressively remove data contamination. 2.3 Training Process As found in [KMH + 20,MKAT18], larger models can typically use a larger batch size, but require a smaller learning rate. We measure the gradient noise scale during training and use it to guide our choice of batch size [MKAT18]. Table 2.1shows the parameter settings we used. To train the larger models without running out of memory, we use a mixture of model parallelism within each matrix multiply and model parallelism across the layers of the network. All models were trained on V100 GPU’s on part of a high-bandwidth cluster provided by Microsoft. Details of the training process and hyperparameter settings are described in AppendixB. 2.4 Evaluation For few-shot learning, we evaluate each example in the evaluation set by randomly drawing K examples from that task’s training set as conditioning, delimited by 1 or 2 newlines depending on the task. For LAMBADA and Story cloze there is no supervised training set available so we draw conditioning examples from the development set and evaluate on the test set. For Winograd (the original, not SuperGLUE version) there is only one dataset, so we draw conditioning examples directly from it. K can be any value from 0 to the maximum amount allowed by the model’s context window, which is <> for all models and typically fits10to100examples. Larger values of K are usually but not always better, so when a separate development and test set are available, we experiment with a few values ofKon the development set and then run the best value on the test set. For some tasks (see AppendixG) we also use a natural language prompt in addition to (or forK= 0, instead of) demonstrations. On tasks that involve choosing one correct completion from several options (multiple choice), we provideKexamples of context plus correct completion, followed by one example of context only, and compare the LM likelihood of each completion. For most tasks we compare the per-token likelihood (to normalize for length), however on a small number of datasets (ARC, OpenBookQA, and RACE) we gain additional benefit as measured on the development set by normalizing by the unconditional probability of each completion, by computing <>, where <> answer context is the string "Answer: "or" A: " and is used to prompt that the completion should be an answer but is otherwise generic. On tasks that involve binary classification, we give the options more semantically meaningful names (e.g. “True” or “False” rather than 0 or 1) and then treat the task like multiple choice; we also sometimes frame the task similar to what is done by [RSR + 19] (see AppendixG) for details. On tasks with free-form completion, we use beam search with the same parameters as [RSR + 19]: a beam width of 4 and a length penalty of= 0:6. We score the model using F1 similarity score, BLEU, or exact match, depending on what is standard for the dataset at hand. Final results are reported on the test set when publicly available, for each model size and learning setting (zero-, one-, and few-shot). When the test set is private, our model is often too large to fit on the test server, so we report results on the development set. We do submit to the test server on a small number of datasets (SuperGLUE, TriviaQA, PiQa) where we were able to make submission work, and we submit only the 200B few-shot results, and report development set results for everything else. 3 Results In Figure3.1we display training curves for the 8 models described in Section2. For this graph we also include 6 additional extra-small models with as few as 100,000 parameters. As observed in [KMH + 20], language modeling performance follows a power-law when making efficient use of training compute. After extending this trend by two more orders of magnitude, we observe only a slight (if any) departure from the power-law. One might worry that these improvements in cross-entropy loss come only from modeling spurious details of our training corpus. However, we will see in the following sections that improvements in cross-entropy loss lead to consistent performance gains across a broad spectrum of natural language tasks. Below, we evaluate the 8 models described in Section2(the 175 billion parameter parameter GPT-3 and 7 smaller models) on a wide range of datasets. We group the datasets into 9 categories representing roughly similar tasks. In Section3.1we evaluate on traditional language modeling tasks and tasks that are similar to language modeling, such as Cloze tasks and sentence/paragraph completion tasks. In Section3.2we evaluate on “closed book” question answering tasks: tasks which require using the information stored in the model’s parameters to answer general knowledge questions. In Section3.3we evaluate the model’s ability to translate between languages (especially one-shot and few-shot). In Section3.4we evaluate the model’s performance on Winograd Schema-like tasks. In Section3.5we evaluate on datasets that involve commonsense reasoning or question answering. In Section3.6we evaluate on reading comprehension tasks, in Section3.7we evaluate on the SuperGLUE benchmark suite, and in3.8we briefly explore NLI. Finally, in Section3.9, we invent some additional tasks designed especially to probe in-context learning abilities – these tasks focus on on-the-fly reasoning, adaptation skills, or open-ended text synthesis. We evaluate all tasks in the few-shot, one-shot, and zero-shot settings. <

> Figure 3.1: Smooth scaling of performance with compute. Performance (measured in terms of cross-entropy validation loss) follows a power-law trend with the amount of compute used for training. The power-law behavior observed in [KMH + 20] continues for an additional two orders of magnitude with only small deviations from the predicted curve. For this figure, we exclude embedding parameters from compute and parameter counts. <

> Table 3.1: Zero-shot results on PTB language modeling dataset.Many other common language modeling datasets are omitted because they are derived from Wikipedia or other sources which are included in GPT-3’s training data. a [RWC + 19] 3.1 Language Modeling, Cloze, and Completion Tasks In this section we test GPT-3’s performance on the traditional task of language modeling, as well as related tasks that involve predicting a single word of interest, completing a sentence or paragraph, or choosing between possible completions of a piece of text. 3.1.1 Language Modeling We calculate zero-shot perplexity on the Penn Tree Bank (PTB) [MKM + 94] dataset measured in [RWC + 19]. We omit the 4 Wikipedia-related tasks in that work because they are entirely contained in our training data, and we also omit the one-billion word benchmark due to a high fraction of the dataset being contained in our training set. PTB escapes these issues due to predating the modern internet. Our largest model sets a new SOTA on PTB by a substantial margin of 15 points, achieving a perplexity of 20.50. Note that since PTB is a traditional language modeling dataset it does not have a clear separation of examples to define one-shot or few-shot evaluation around, so we measure only zero-shot. 3.1.2 LAMBADA The LAMBADA dataset [PKL + 16] tests the modeling of long-range dependencies in text – the model is asked to predict the last word of sentences which require reading a paragraph of context. It has recently been suggested that the continued scaling of language models is yielding diminishing returns on this difficult benchmark. [BHT + 20] reflect on the small 1.5% improvement achieved by a doubling of model size between two recent state of the art results [SPP + 19] <> Table 3.2: Performance on cloze and completion tasks.GPT-3 significantly improves SOTA on LAMBADA while achieving respectable performance on two difficult completion prediction datasets. <

> Figure 3.2:On LAMBADA, the few-shot capability of language models results in a strong boost to accuracy. GPT-3 2.7B outperforms the SOTA 17B parameter Turing-NLG [Tur20] in this setting, and GPT-3 175B advances the state of the art by 18%. Note zero-shot uses a different format from one-shot and few-shot as described in the text. and [Tur20]) and argue that “continuing to expand hardware and data sizes by orders of magnitude is not the path forward”. We find that path is still promising and in a zero-shot setting GPT-3 achieves 76% on LAMBADA, a gain of 8% over the previous state of the art. LAMBADA is also a demonstration of the flexibility of few-shot learning as it provides a way to address a problem that classically occurs with this dataset. Although the completion in LAMBADA is always the last word in a sentence, a standard language model has no way of knowing this detail. It thus assigns probability not only to the correct ending but also to other valid continuations of the paragraph. This problem has been partially addressed in the past with stop-word filters [RWC + 19] (which ban “continuation” words). The few-shot setting instead allows us to “frame” the task as a cloze-test and allows the language model to infer from examples that a completion of exactly one word is desired. We use the following fill-in-the-blank format: Alice was friends with Bob. Alice went to visit her friend .!Bob George bought some baseball equipment, a ball, a glove, and a .! When presented with examples formatted this way, GPT-3 achieves 86.4% accuracy in the few-shot setting, an increase of over 18% from the previous state-of-the-art. We observe that few-shot performance improves strongly with model size. While this setting decreases the performance of the smallest model by almost 20%, for GPT-3 it improves accuracy by 10%. Finally, the fill-in-blank method is not effective one-shot, where it always performs worse than the zero-shot setting. Perhaps this is because all models still require several examples to recognize the pattern. <

> Table 3.3: Results on three Open-Domain QA tasks.GPT-3 is shown in the few-, one-, and zero-shot settings, as compared to prior SOTA results for closed book and open domain settings. TriviaQA few-shot result is evaluated on the wiki split test server. One note of caution is that an analysis of test set contamination identified that a significant minority of the LAMBADA dataset appears to be present in our training data – however analysis performed in Section4suggests negligible impact on performance. 3.1.3 HellaSwag The HellaSwag dataset [ZHB + 19] involves picking the best ending to a story or set of instructions. The examples were adversarially mined to be difficult for language models while remaining easy for humans (who achieve 95.6% accuracy). GPT-3 achieves 78.1% accuracy in the one-shot setting and 79.3% accuracy in the few-shot setting, outperforming the 75.4% accuracy of a fine-tuned 1.5B parameter language model [ZHR + 19] but still a fair amount lower than the overall SOTA of 85.6% achieved by the fine-tuned multi-task model ALUM. 3.1.4 StoryCloze We next evaluate GPT-3 on the StoryCloze 2016 dataset [MCH + 16], which involves selecting the correct ending sentence for five-sentence long stories. Here GPT-3 achieves 83.2% in the zero-shot setting and 87.7% in the few-shot setting (withK= 70). This is still 4.1% lower than the fine-tuned SOTA using a BERT based model [LDL19] but improves over previous zero-shot results by roughly 10%. 3.2 Closed Book Question Answering In this section we measure GPT-3’s ability to answer questions about broad factual knowledge. Due to the immense amount of possible queries, this task has normally been approached by using an information retrieval system to find relevant text in combination with a model which learns to generate an answer given the question and the retrieved text. Since this setting allows a system to search for and condition on text which potentially contains the answer it is denoted “open-book”. [RRS20] recently demonstrated that a large language model can perform surprisingly well directly answering the questions without conditioning on auxiliary information. They denote this more restrictive evaluation setting as “closed-book”. Their work suggests that even higher-capacity models could perform even better and we test this hypothesis with GPT-3. We evaluate GPT-3 on the 3 datasets in [RRS20]: Natural Questions [KPR + 19], WebQuestions [BCFL13], and TriviaQA [JCWZ17], using the same splits. Note that in addition to all results being in the closed-book setting, our use of few-shot, one-shot, and zero-shot evaluations represent an even stricter setting than previous closed-book QA work: in addition to external content not being allowed, fine-tuning on the Q&A dataset itself is also not permitted. The results for GPT-3 are shown in Table3.3. On TriviaQA, we achieve 64.3% in the zero-shot setting, 68.0% in the one-shot setting, and 71.2% in the few-shot setting. The zero-shot result already outperforms the fine-tuned T5-11B by 14.2%, and also outperforms a version with Q&A tailored span prediction during pre-training by 3.8%. The one-shot result improves by 3.7% and matches the SOTA for an open-domain QA system which not only fine-tunes but also makes use of a learned retrieval mechanism over a 15.3B parameter dense vector index of 21M documents [LPP + 20]. GPT-3’s few-shot result further improves performance another 3.2% beyond this. On WebQuestions (WebQs), GPT-3 achieves 14.4% in the zero-shot setting, 25.3% in the one-shot setting, and 41.5% in the few-shot setting. This compares to 37.4% for fine-tuned T5-11B, and 44.7% for fine-tuned T5-11B+SSM, which uses a Q&A-specific pre-training procedure. GPT-3 in the few-shot setting approaches the performance of state-of-the-art fine-tuned models. Notably, compared to TriviaQA, WebQS shows a much larger gain from zero-shot to few-shot (and indeed its zero-shot and one-shot performance are poor), perhaps suggesting that the WebQs questions <

> Figure 3.3:On TriviaQA GPT3’s performance grows smoothly with model size, suggesting that language models continue to absorb knowledge as their capacity increases. One-shot and few-shot performance make significant gains over zero-shot behavior, matching and exceeding the performance of the SOTA fine-tuned open-domain model, RAG [LPP + 20] and/or the style of their answers are out-of-distribution for GPT-3. Nevertheless, GPT-3 appears able to adapt to this distribution, recovering strong performance in the few-shot setting. On Natural Questions (NQs) GPT-3 achieves 14.6% in the zero-shot setting, 23.0% in the one-shot setting, and 29.9% in the few-shot setting, compared to 36.6% for fine-tuned T5 11B+SSM. Similar to WebQS, the large gain from zero-shot to few-shot may suggest a distribution shift, and may also explain the less competitive performance compared to TriviaQA and WebQS. In particular, the questions in NQs tend towards very fine-grained knowledge on Wikipedia specifically which could be testing the limits of GPT-3’s capacity and broad pretraining distribution. Overall, on one of the three datasets GPT-3’s one-shot matches the open-domain fine-tuning SOTA. On the other two datasets it approaches the performance of the closed-book SOTA despite not using fine-tuning. On all 3 datasets, we find that performance scales very smoothly with model size (Figure3.3and AppendixHFigureH.7), possibly reflecting the idea that model capacity translates directly to more ‘knowledge’ absorbed in the parameters of the model. 3.3 Translation For GPT-2 a filter was used on a multilingual collection of documents to produce an English only dataset due to capacity concerns. Even with this filtering GPT-2 showed some evidence of multilingual capability and performed non-trivially when translating between French and English despite only training on 10 megabytes of remaining French text. Since we increase the capacity by over two orders of magnitude from GPT-2 to GPT-3, we also expand the scope of the training dataset to include more representation of other languages, though this remains an area for further improvement. As discussed in2.2the majority of our data is derived from raw Common Crawl with only quality-based filtering. Although GPT-3’s training data is still primarily English (93% by word count), it also includes 7% of text in other languages. These languages are documented in the supplemental material. In order to better understand translation capability, we also expand our analysis to include two additional commonly studied languages, German and Romanian. Existing unsupervised machine translation approaches often combine pretraining on a pair of monolingual datasets with back-translation [SHB15] to bridge the two languages in a controlled way. By contrast, GPT-3 learns from a blend of training data that mixes many languages together in a natural way, combining them on a word, sentence, and document level. GPT-3 also uses a single training objective which is not customized or designed for any task in particular. However, our one / few-shot settings aren’t strictly comparable to prior unsupervised work since they make use of a small amount of paired examples (1 or 64). This corresponds to up to a page or two of in-context training data. Results are shown in Table3.4. Zero-shot GPT-3, which only receives on a natural language description of the task, still underperforms recent unsupervised NMT results. However, providing only a single example demonstration for <

> Table 3.4: Few-shot GPT-3 outperforms previous unsupervised NMT work by 5 BLEU when translating into English reflecting its strength as an English LM.We report BLEU scores on the WMT’14 Fr$En, WMT’16 De$En, and WMT’16 Ro$En datasets as measured by multi-bleu.perl with XLM’s tokenization in order to compare most closely with prior unsupervised NMT work. SacreBLEU f [Pos18] results re- ported in AppendixH. Underline indicates an unsupervised or few-shot SOTA, bold indicates supervised SOTA with relative confidence. a [EOAG18]b [DHKH14]c [WXH + 18]d [oR16]e [LGG + 20]f [SacreBLEU signature: BLEU+case.mixed+numrefs.1+smooth.exp+tok.intl+version.1.2.20] <

> Figure 3.4:Few-shot translation performance on 6 language pairs as model capacity increases. There is a consistent trend of improvement across all datasets as the model scales, and as well as tendency for translation into English to be stronger than translation from English. <

> Table 3.5:Results on the WSC273 version of Winograd schemas and the adversarial Winogrande dataset. See Section 4for details on potential contamination of the Winograd test set. a [SBBC19]b [LYN + 20] <

> Figure 3.5:Zero-, one-, and few-shot performance on the adversarial Winogrande dataset as model capacity scales. Scaling is relatively smooth with the gains to few-shot learning increasing with model size, and few-shot GPT-3 175B is competitive with a fine-tuned RoBERTA-large. each translation task improves performance by over 7 BLEU and nears competitive performance with prior work. GPT-3 in the full few-shot setting further improves another 4 BLEU resulting in similar average performance to prior unsupervised NMT work. GPT-3 has a noticeable skew in its performance depending on language direction. For the three input languages studied, GPT-3 significantly outperforms prior unsupervised NMT work when translating into English but under-performs when translating in the other direction. Performance on En-Ro is a noticeable outlier at over 10 BLEU worse than prior unsupervised NMT work. This could be a weakness due to reusing the byte-level BPE tokenizer of GPT-2 which was developed for an almost entirely English training dataset. For both Fr-En and De-En, few shot GPT-3 outperforms the best supervised result we could find but due to our unfamiliarity with the literature and the appearance that these are un-competitive benchmarks we do not suspect those results represent true state of the art. For Ro-En, few shot GPT-3 performs within 0.5 BLEU of the overall SOTA which is achieved by a combination of unsupervised pretraining, supervised finetuning on 608K labeled examples, and backtranslation [LHCG19b]. Finally, across all language pairs and across all three settings (zero-, one-, and few-shot), there is a smooth trend of improvement with model capacity. This is shown in Figure3.4in the case of few-shot results, and scaling for all three settings is shown in AppendixH. 3.4 Winograd-Style Tasks The Winograd Schemas Challenge [LDM12] is a classical task in NLP that involves determining which word a pronoun refers to, when the pronoun is grammatically ambiguous but semantically unambiguous to a human. Recently fine-tuned language models have achieved near-human performance on the original Winograd dataset, but more difficult versions <

> Table 3.6:GPT-3 results on three commonsense reasoning tasks, PIQA, ARC, and OpenBookQA. GPT-3 Few-Shot PIQA result is evaluated on the test server. See Section4for details on potential contamination issues on the PIQA test set. <

> Figure 3.6:GPT-3 results on PIQA in the zero-shot, one-shot, and few-shot settings. The largest model achieves a score on the development set in all three conditions that exceeds the best recorded score on the task. such as the adversarially-mined Winogrande dataset [SBBC19] still significantly lag human performance. We test GPT-3’s performance on both Winograd and Winogrande, as usual in the zero-, one-, and few-shot setting. On Winograd we test GPT-3 on the original set of 273 Winograd schemas, using the same “partial evaluation” method described in [RWC + 19]. Note that this setting differs slightly from the WSC task in the SuperGLUE benchmark, which is presented as binary classification and requires entity extraction to convert to the form described in this section. On Winograd GPT-3 achieves 88.3%, 89.7%, and 88.6% in the zero-shot, one-shot, and few-shot settings, showing no clear in-context learning but in all cases achieving strong results just a few points below state-of-the-art and estimated human performance. We note that contamination analysis found some Winograd schemas in the training data but this appears to have only a small effect on results (see Section4). On the more difficult Winogrande dataset, we do find gains to in-context learning: GPT-3 achieves 70.2% in the zero-shot setting, 73.2% in the one-shot setting, and 77.7% in the few-shot setting. For comparison a fine-tuned RoBERTA model achieves 79%, state-of-the-art is 84.6% achieved with a fine-tuned high capacity model (T5), and human performance on the task as reported by [SBBC19] is 94.0%. 3.5 Common Sense Reasoning Next we consider three datasets which attempt to capture physical or scientific reasoning, as distinct from sentence completion, reading comprehension, or broad knowledge question answering. The first, PhysicalQA (PIQA) [BZB + 19], asks common sense questions about how the physical world works and is intended as a probe of grounded understanding of the world. GPT-3 achieves 81.0% accuracy zero-shot, 80.5% accuracy one-shot, and 82.8% accuracy few-shot (the last measured on PIQA’s test server). This compares favorably to the 79.4% accuracy prior state-of-the-art of a <

> Table 3.7:Results on reading comprehension tasks. All scores are F1 except results for RACE which report accuracy. a [JZC + 19]b [JN20]c [AI19]d [QIA20]e [SPP + 19] fine-tuned RoBERTa. PIQA shows relatively shallow scaling with model size and is still over 10% worse than human performance, but GPT-3’s few-shot and even zero-shot result outperform the current state-of-the-art. Our analysis flagged PIQA for a potential data contamination issue (despite hidden test labels), and we therefore conservatively mark the result with an asterisk. See Section4for details. ARC [CCE + 18] is a dataset of multiple-choice questions collected from 3rd to 9th grade science exams. On the “Challenge” version of the dataset which has been filtered to questions which simple statistical or information retrieval methods are unable to correctly answer, GPT-3 achieves 51.4% accuracy in the zero-shot setting, 53.2% in the one-shot setting, and 51.5% in the few-shot setting. This is approaching the performance of a fine-tuned RoBERTa baseline (55.9%) from UnifiedQA [KKS + 20]. On the “Easy” version of the dataset (questions which either of the mentioned baseline approaches answered correctly), GPT-3 achieves 68.8%, 71.2%, and 70.1% which slightly exceeds a fine-tuned RoBERTa baseline from [KKS + 20]. However, both of these results are still much worse than the overall SOTAs achieved by the UnifiedQA which exceeds GPT-3’s few-shot results by 27% on the challenge set and 22% on the easy set. On OpenBookQA [MCKS18], GPT-3 improves significantly from zero to few shot settings but is still over 20 points short of the overall SOTA. GPT-3’s few-shot performance is similar to a fine-tuned BERT Large baseline on the leaderboard. Overall, in-context learning with GPT-3 shows mixed results on commonsense reasoning tasks, with only small and inconsistent gains observed in the one and few-shot learning settings for both PIQA and ARC, but a significant improvement is observed on OpenBookQA. GPT-3 sets SOTA on the new PIQA dataset in all evaluation settings. 3.6 Reading Comprehension Next we evaluate GPT-3 on the task of reading comprehension. We use a suite of 5 datasets including abstractive, multiple choice, and span based answer formats in both dialog and single question settings. We observe a wide spread in GPT-3’s performance across these datasets suggestive of varying capability with different answer formats. In general we observe GPT-3 is on par with initial baselines and early results trained using contextual representations on each respective dataset. GPT-3 performs best (within 3 points of the human baseline) on CoQA [RCM19] a free-form conversational dataset and performs worst (13 F1 below an ELMo baseline) on QuAC [CHI + 18] a dataset which requires modeling structured dialog acts and answer span selections of teacher-student interactions. On DROP [DWD + 19], a dataset testing discrete reasoning and numeracy in the context of reading comprehension, GPT-3 in a few-shot setting outperforms the fine-tuned BERT baseline from the original paper but is still well below both human performance and state-of-the-art approaches which augment neural networks with symbolic systems [RLL + 19]. On SQuAD 2.0 [RJL18], GPT-3 demonstrates its few-shot learning capabilities, improving by almost 10 F1 (to 69.8) compared to a zero-shot setting. This allows it to slightly outperform the best fine-tuned result in the original paper. On RACE [LXL + 17], a multiple choice dataset of middle school and high school english examinations, GPT-3 performs relatively weakly and is only competitive with the earliest work utilizing contextual representations and is still 45% behind SOTA. 3.7 SuperGLUE In order to better aggregate results on NLP tasks and compare to popular models such as BERT and RoBERTa in a more systematic way, we also evaluate GPT-3 on a standardized collection of datasets, the SuperGLUE benchmark [WPN + 19]. GPT-3’s test-set performance on the SuperGLUE dataset [WPN + 19] is shown in Table3.8. In the few-shot setting, we used 32 examples for all tasks, sampled randomly from the training set. For all tasks except WSC and <

> Figure 3.7:GPT-3 results on CoQA reading comprehension task. GPT-3 175B achieves 85 F1 in the few-shot setting, only a few points behind measured human performance and state-of-the-art fine-tuned models. Zero-shot and one-shot performance is a few points behind, with the gains to few-shot being largest for bigger models. <

> Table 3.8:Performance of GPT-3 on SuperGLUE compared to fine-tuned baselines and SOTA. All results are reported on the test set. GPT-3 few-shot is given a total of 32 examples within the context of each task and performs no gradient updates. <

> Figure 3.8: Performance on SuperGLUE increases with model size and number of examples in context.A value of K= 32 means that our model was shown 32 examples per task, for 256 examples total divided across the 8 tasks in SuperGLUE. We report GPT-3 values on the dev set, so our numbers are not directly comparable to the dotted reference lines (our test set results are in Table3.8). The BERT-Large reference model was fine-tuned on the SuperGLUE training set (125K examples), whereas BERT++ was first fine-tuned on MultiNLI (392K examples) and SWAG (113K examples) before further fine-tuning on the SuperGLUE training set (for a total of 630K fine-tuning examples). We find the difference in performance between the BERT-Large and BERT++ to be roughly equivalent to the difference between GPT-3 with one example per context versus eight examples per context. MultiRC, we sampled a new set of examples to use in the context for each problem. For WSC and MultiRC, we used the same set of randomly drawn examples from the training set as context for all of the problems we evaluated. We observe a wide range in GPT-3’s performance across tasks. On COPA and ReCoRD GPT-3 achieves near-SOTA performance in the one-shot and few-shot settings, with COPA falling only a couple points short and achieving second place on the leaderboard, where first place is held by a fine-tuned 11 billion parameter model (T5). On WSC, performance is still relatively strong, achieving 80.1% in the few-shot setting (note that GPT-3 achieves 88.6% on the original Winograd dataset as described in Section3.4). On BoolQ, MultiRC, and RTE, performance is reasonable, roughly matching that of a fine-tuned BERT-Large. On CB, we see signs of life at 75.6% in the few-shot setting. WiC is a notable weak spot with few-shot performance at 49.4% (at random chance). We tried a number of different phrasings and formulations for WiC (which involves determining if a word is being used with the same meaning in two sentences), none of which was able to achieve strong performance. This hints at a phenomenon that will become clearer in the next section (which discusses the ANLI benchmark) – GPT-3 appears to be weak in the few-shot or one-shot setting at some tasks that involve comparing two sentences or snippets, for example whether a word is used the same way in two sentences (WiC), whether one sentence is a paraphrase of another, or whether one sentence implies another. This could also explain the comparatively low scores for RTE and CB, which also follow this format. Despite these weaknesses, GPT-3 still outperforms a fine-tuned BERT-large on four of eight tasks and on two tasks GPT-3 is close to the state-of-the-art held by a fine-tuned 11 billion parameter model. Finally, we note that the few-shot SuperGLUE score steadily improves with both model size and with number of examples in the context showing increasing benefits from in-context learning (Figure3.8). We scale K up to 32 examples per task, after which point additional examples will not reliably fit into our context. When sweeping over values ofK, we find that GPT-3 requires less than eight total examples per task to outperform a fine-tuned BERT-Large on overall SuperGLUE score. 3.8 NLI Natural Language Inference (NLI) [Fyo00] concerns the ability to understand the relationship between two sentences. In practice, this task is usually structured as a two or three class classification problem where the model classifies <

> Figure 3.9: Performance of GPT-3 on ANLI Round 3.Results are on the dev-set, which has only 1500 examples and therefore has high variance (we estimate a standard deviation of 1.2%). We find that smaller models hover around random chance, while few-shot GPT-3 175B closes almost half the gap from random chance to SOTA. Results for ANLI rounds 1 and 2 are shown in the appendix. whether the second sentence logically follows from the first, contradicts the first sentence, or is possibly true (neutral). SuperGLUE includes an NLI dataset, RTE, which evaluates the binary version of the task. On RTE, only the largest version of GPT-3 performs convincingly better than random (56%) in any evaluation setting, but in a few-shot setting GPT-3 performs similarly to a single-task fine-tuned BERT Large. We also evaluate on the recently introduced Adversarial Natural Language Inference (ANLI) dataset [NWD + 19]. ANLI is a difficult dataset employing a series of adversarially mined natural language inference questions in three rounds (R1, R2, and R3). Similar to RTE, all of our models smaller than GPT-3 perform at almost exactly random chance on ANLI, even in the few-shot setting (33%), whereas GPT-3 itself shows signs of life on Round 3. Results for ANLI R3 are highlighted in Figure3.9and full results for all rounds can be found in AppendixH. These results on both RTE and ANLI suggest that NLI is still a very difficult task for language models and they are only just beginning to show signs of progress. 3.9 Synthetic and Qualitative Tasks One way to probe GPT-3’s range of abilities in the few-shot (or zero- and one-shot) setting is to give it tasks which require it to perform simple on-the-fly computational reasoning, recognize a novel pattern that is unlikely to have occurred in training, or adapt quickly to an unusual task. We devise several tasks to test this class of abilities. First, we test GPT-3’s ability to perform arithmetic. Second, we create several tasks that involve rearranging or unscrambling the letters in a word, tasks which are unlikely to have been exactly seen during training. Third, we test GPT-3’s ability to solve SAT-style analogy problems few-shot. Finally, we test GPT-3 on several qualitative tasks, including using new words in a sentence, correcting English grammar, and news article generation. We will release the synthetic datasets with the hope of stimulating further study of test-time behavior of language models. 3.9.1 Arithmetic To test GPT-3’s ability to perform simple arithmetic operations without task-specific training, we developed a small battery of 10 tests that involve asking GPT-3 a simple arithmetic problem in natural language: •2 digit addition (2D+)– The model is asked to add two integers sampled uniformly from[0;100), phrased in the form of a question, e.g. “Q: What is 48 plus 76? A: 124.” •2 digit subtraction (2D-)– The model is asked to subtract two integers sampled uniformly from[0;100); the answer may be negative. Example: “Q: What is 34 minus 53? A: -19”. •3 digit addition (3D+)– Same as 2 digit addition, except numbers are uniformly sampled from[0;1000). <

> Figure 3.10:Results on all 10 arithmetic tasks in the few-shot settings for models of different sizes. There is a significant jump from the second largest model (GPT-3 13B) to the largest model (GPT-3 175), with the latter being able to reliably accurate 2 digit arithmetic, usually accurate 3 digit arithmetic, and correct answers a significant fraction of the time on 4-5 digit arithmetic, 2 digit multiplication, and compound operations. Results for one-shot and zero-shot are shown in the appendix. •3 digit subtraction (3D-)– Same as 2 digit subtraction, except numbers are uniformly sampled from[0;1000). •4 digit addition (4D+)– Same as 3 digit addition, except uniformly sampled from[0;10000). •4 digit subtraction (4D-)– Same as 3 digit subtraction, except uniformly sampled from[0;10000). •5 digit addition (5D+)– Same as 3 digit addition, except uniformly sampled from[0;100000). •5 digit subtraction (5D-)– Same as 3 digit subtraction, except uniformly sampled from[0;100000). •2 digit multiplication (2Dx)– The model is asked to multiply two integers sampled uniformly from[0;100), e.g. “Q: What is 24 times 42? A: 1008”. •One-digit composite (1DC)– The model is asked to perform a composite operation on three 1 digit numbers, with parentheses around the last two. For example, “Q: What is 6+(4*8)? A: 38”. The three 1 digit numbers are selected uniformly on[0;10)and the operations are selected uniformly from f+,-,*g. In all 10 tasks the model must generate the correct answer exactly. For each task we generate a dataset of 2,000 random instances of the task and evaluate all models on those instances. First we evaluate GPT-3 in the few-shot setting, for which results are shown in Figure3.10. On addition and subtraction, GPT-3 displays strong proficiency when the number of digits is small, achieving 100% accuracy on 2 digit addition, 98.9% at 2 digit subtraction, 80.2% at 3 digit addition, and 94.2% at 3-digit subtraction. Performance decreases as the number of digits increases, but GPT-3 still achieves 25-26% accuracy on four digit operations and 9-10% accuracy on five digit operations, suggesting at least some capacity to generalize to larger numbers of digits. GPT-3 also achieves 29.2% accuracy at 2 digit multiplication, an especially computationally intensive operation. Finally, GPT-3 achieves 21.3% accuracy at single digit combined operations (for example, 9*(7+5)), suggesting that it has some robustness beyond just single operations. As Figure3.10makes clear, small models do poorly on all of these tasks – even the 13 billion parameter model (the second largest after the 175 billion full GPT-3) can solve 2 digit addition and subtraction only half the time, and all other operations less than 10% of the time. One-shot and zero-shot performance are somewhat degraded relative to few-shot performance, suggesting that adaptation to the task (or at the very least recognition of the task) is important to performing these computations correctly. Nevertheless, one-shot performance is still quite strong, and even zero-shot performance of the full GPT-3 significantly <

> Table 3.9:Results on basic arithmetic tasks for GPT-3 175B.f2,3,4,5gDf+,-gis 2, 3, 4, and 5 digit addition or subtraction, 2Dx is 2 digit multiplication. 1DC is 1 digit composite operations. Results become progressively stronger moving from the zero-shot to one-shot to few-shot setting, but even the zero-shot shows significant arithmetic abilities. <> Table 3.10:GPT-3 175B performance on various word unscrambling and word manipulation tasks, in zero-, one-, and few-shot settings. CL is “cycle letters in word”, A1 is anagrams of but the first and last letters, A2 is anagrams of all but the first and last two letters, RI is “Random insertion in word”, RW is “reversed words”. outperforms few-shot learning for all smaller models. All three settings for the full GPT-3 are shown in Table3.9, and model capacity scaling for all three settings is shown in AppendixH. To spot-check whether the model is simply memorizing specific arithmetic problems, we took the 3-digit arithmetic problems in our test set and searched for them in our training data in both the forms" + ="and " plus ". Out of 2,000 addition problems we found only 17 matches (0.8%) and out of 2,000 subtraction problems we found only 2 matches (0.1%), suggesting that only a trivial fraction of the correct answers could have been memorized. In addition, inspection of incorrect answers reveals that the model often makes mistakes such as not carrying a “1”, suggesting it is actually attempting to perform the relevant computation rather than memorizing a table. Overall, GPT-3 displays reasonable proficiency at moderately complex arithmetic in few-shot, one-shot, and even zero-shot settings. 3.9.2 Word Scrambling and Manipulation Tasks To test GPT-3’s ability to learn novel symbolic manipulations from a few examples, we designed a small battery of 5 “character manipulation” tasks. Each task involves giving the model a word distorted by some combination of scrambling, addition, or deletion of characters, and asking it to recover the original word. The 5 tasks are: •Cycle letters in word (CL)– The model is given a word with its letters cycled, then the “=” symbol, and is expected to generate the original word. For example, it might be given “lyinevitab” and should output “inevitably”. •Anagrams of all but first and last characters (A1)– The model is given a word where every letter except the first and last have been scrambled randomly, and must output the original word. Example: criroptuon = corruption. •Anagrams of all but first and last 2 characters (A2)– The model is given a word where every letter except the first 2 and last 2 have been scrambled randomly, and must recover the original word. Example: opoepnnt !opponent. •Random insertion in word (RI)– A random punctuation or space character is inserted between each letter of a word, and the model must output the original word. Example: s.u!c/c!e.s s i/o/n = succession. •Reversed words (RW)– The model is given a word spelled backwards, and must output the original word. Example: stcejbo!objects. For each task we generate 10,000 examples, which we chose to be the top 10,000 most frequent words as measured by [Nor09] of length more than 4 characters and less than 15 characters. The few-shot results are shown in Figure3.11. Task performance tends to grow smoothly with model size, with the full GPT-3 model achieving 66.9% on removing <

> Figure 3.11:Few-shot performance on the five word scrambling tasks for different sizes of model. There is generally smooth improvement with model size although the random insertion task shows an upward slope of improvement with the 175B model solving the task the majority of the time. Scaling of one-shot and zero-shot performance is shown in the appendix. All tasks are done with K=100. random insertions, 38.6% on cycling letters, 40.2% on the easier anagram task, and 15.1% on the more difficult anagram task (where only the first and last letters are held fixed). None of the models can reverse the letters in a word. In the one-shot setting, performance is significantly weaker (dropping by half or more), and in the zero-shot setting the model can rarely perform any of the tasks (Table3.10). This suggests that the model really does appear to learn these tasks at test time, as the model cannot perform them zero-shot and their artificial nature makes them unlikely to appear in the pre-training data (although we cannot confirm this with certainty). We can further quantify performance by plotting “in-context learning curves”, which show task performance as a function of the number of in-context examples. We show in-context learning curves for the Symbol Insertion task in Figure1.2. We can see that larger models are able to make increasingly effective use of in-context information, including both task examples and natural language task descriptions. Finally, it is worth adding that solving these tasks requires character-level manipulations, whereas our BPE encoding operates on significant fractions of a word (on average 0.7 words per token), so from the LM’s perspective succeeding at these tasks involves not just manipulating BPE tokens but understanding and pulling apart their substructure. Also, CL, A1, and A2 are not bijective (that is, the unscrambled word is not a deterministic function of the scrambled word), requiring the model to perform some search to find the correct unscrambling. Thus, the skills involved appear to require non-trivial pattern-matching and computation. 3.9.3 SAT Analogies To test GPT-3 on another task that is somewhat unusual relative to the typical distribution of text, we collected a set of 374 “SAT analogy” problems [TLBS03]. Analogies are a style of multiple choice question that constituted a section of the SAT college entrance exam before 2005. A typical example is “audacious is to boldness as (a) sanctimonious is to hypocrisy, (b) anonymous is to identity, (c) remorseful is to misdeed, (d) deleterious is to result, (e) impressionable is to temptation”. The student is expected to choose which of the five word pairs has the same relationship as the original word pair; in this example the answer is “sanctimonious is to hypocrisy”. On this task GPT-3 achieves 65.2% in the few-shot setting, 59.1% in the one-shot setting, and 53.7% in the zero-shot setting, whereas the average score among college applicants was 57% [TL05] (random guessing yields 20%). As shown in Figure3.12, the results improve with scale, with the the full 175 billion model improving by over 10% compared to the 13 billion parameter model. <

> Figure 3.12:Zero-, one-,and few-shot performance on SAT analogy tasks, for different sizes of model. The largest model achieves 65% accuracy in the few-shot setting, and also demonstrates significant gains to in-context learning which are not present in smaller models. 3.9.4 News Article Generation Previous work on generative language models qualitatively tested their ability to generate synthetic “news articles” by conditional sampling from the model given a human-written prompt consisting of a plausible first sentence for a news story [RWC + 19]. Relative to [RWC + 19], the dataset used to train GPT-3 is much less weighted towards news articles, so trying to generate news articles via raw unconditional samples is less effective – for example GPT-3 often interprets the proposed first sentence of a “news article” as a tweet and then posts synthetic responses or follow-up tweets. To solve this problem we employed GPT-3’s few-shot learning abilities by providing three previous news articles in the model’s context to condition it. With the title and subtitle of a proposed next article, the model is able to reliably generate short articles in the “news” genre. To gauge the quality of news article generation from GPT-3 (which we believe is likely to be correlated with conditional sample generation quality in general), we decided to measure human ability to distinguish GPT-3-generated articles from real ones. Similar work has been carried out by Kreps et al. [KMB20] and Zellers et al. [ZHR + 19]. Generative language models are trained to match the distribution of content generated by humans, so the (in)ability of humans to distinguish the two is a potentially important measure of quality. In order to see how well humans can detect model generated text, we arbitrarily selected 25 article titles and subtitles from the websitenewser.com(mean length: 215 words). We then generated completions of these titles and subtitles from four language models ranging in size from 125M to 175B (GPT-3) parameters (mean length: 200 words). For each model, we presented around 80 US-based participants with a quiz consisting of these real titles and subtitles followed by either the human written article or the article generated by the model 4 . Participants were asked to select whether the article was “very likely written by a human”, “more likely written by a human”, “I don’t know”, “more likely written by a machine”, or “very likely written by a machine”. The articles we selected were not in the models’ training data and the model outputs were formatted and selected programmatically to prevent human cherry-picking. All models used the same context to condition outputs on and were pre-trained with the same context size and the same article titles and subtitles were used as prompts for each model. However, we also ran an experiment to control for participant effort and attention that followed the same format but involved intentionally bad model generated articles. This was done by generating articles from a “control model”: a 160M parameter model with no context and increased output randomness. 3 This task is also relevant to the potential misuse of language models discussed in Section6.1. 4 We wanted to identify how good an average person on the internet is at detecting language model outputs, so we focused on participants drawn from the general US population. See AppendixEfor details. <

> Table 3.11: Human accuracy in identifying whether short (200 word) news articles are model generated. We find that human accuracy (measured by the ratio of correct assignments to non-neutral assignments) ranges from 86% on the control model to 52% on GPT-3 175B. This table compares mean accuracy between five different models, and shows the results of a two-sample T-Test for the difference in mean accuracy between each model and the control model (an unconditional GPT-3 Small model with increased output randomness). Mean human accuracy (the ratio of correct assignments to non-neutral assignments per participant) at detecting that the intentionally bad articles were model generated was 86% where 50% is chance level performance. By contrast, mean human accuracy at detecting articles that were produced by the 175B parameter model was barely above chance at 52% (see Table3.11). 5 Human abilities to detect model generated text appear to decrease as model size increases: there appears to be a trend towards chance accuracy with model size, and human detection of GPT-3 is close to chance. 6 This is true despite the fact that participants spend more time on each output as model size increases (see AppendixE). Examples of synthetic articles from GPT-3 are given in Figures3.14and3.15.7 Much of the text is—as indicated by the evaluations—difficult for humans to distinguish from authentic human content. Factual inaccuracies can be an indicator that an article is model generated since, unlike human authors, the models have no access to the specific facts that the article titles refer to or when the article was written. Other indicators include repetition, non sequiturs, and unusual phrasings, though these are often subtle enough that they are not noticed. Related work on language model detection by Ippolito et al. [IDCBE19] indicates that automatic discriminators like GROVER [ZHR + 19] and GLTR [GSR19] may have greater success at detecting model generated text than human evaluators. Automatic detection of these models may be a promising area of future research. Ippolito et al. [IDCBE19] also note that human accuracy at detecting model generated text increases as humans observe more tokens. To do a preliminary investigation of how good humans are at detecting longer news articles generated by GPT-3 175B, we selected 12 world news articles from Reuters with an average length of 569 words and generated completions of these articles from GPT-3 with an average length of 498 words (298 words longer than our initial experiments). Following the methodology above, we ran two experiments, each on around 80 US-based participants, to compare human abilities to detect the articles generated by GPT-3 and a control model. We found that mean human accuracy at detecting the intentionally bad longer articles from the control model was 88%, while mean human accuracy at detecting the longer articles that were produced by GPT-3 175B was still barely above chance at 52%(see Table3.12). This indicates that, for news articles that are around 500 words long, GPT-3 continues to produce articles that humans find difficult to distinguish from human written news articles. 3.9.5 Learning and Using Novel Words A task studied in developmental linguistics [CB78] is the ability to learn and utilize new words, for example using a word in a sentence after seeing it defined only once, or conversely inferring a word’s meaning from only one usage. Here we qualitatively test GPT-3’s ability to do the former. Specifically, we give GPT-3 the definition of a nonexistent word, such as “Gigamuru”, and then ask it to use it in a sentence. We provide one to five previous examples of a (separate) 5 We use a two-sample Student’s T-Test to test for significant difference between the means of the participant accuracies of each model and the control model and report the normalized difference in the means (as the t-statistic) and the p-value. 6 If a model consistently produces texts that are more impressive than human articles, it is possible that human performance on this task would drop below 50%. Indeed, many individual participants scored below 50% on this task. 7 Additional non-news samples can be found in AppendixF. <

> Figure 3.13:People’s ability to identify whether news articles are model-generated (measured by the ratio of correct assignments to non-neutral assignments) decreases as model size increases. Accuracy on the outputs on the deliberately- bad control model (an unconditioned GPT-3 Small model with higher output randomness) is indicated with the dashed line at the top, and the random chance (50%) is indicated with the dashed line at the bottom. Line of best fit is a power law with 95% confidence intervals. <

> Table 3.12:People’s ability to identify whether 500 word articles are model generated (as measured by the ratio of correct assignments to non-neutral assignments) was 88% on the control model and 52% on GPT-3 175B. This table shows the results of a two-sample T-Test for the difference in mean accuracy between GPT-3 175B and the control model (an unconditional GPT-3 Small model with increased output randomness). <

> Figure 3.14:The GPT-3 generated news article that humans had the greatest difficulty distinguishing from a human written article (accuracy: 12%). <

> Figure 3.15:The GPT-3 generated news article that humans found the easiest to distinguish from a human written article (accuracy: 61%). <

> Figure 3.16:Representative GPT-3 completions for the few-shot task of using a new word in a sentence. Boldface is GPT-3’s completions, plain text is human prompts. In the first example both the prompt and the completion are provided by a human; this then serves as conditioning for subsequent examples where GPT-3 receives successive additional prompts and provides the completions. Nothing task-specific is provided to GPT-3 other than the conditioning shown here. nonexistent word being defined and used in a sentence, so the task is few-shot in terms of previous examples of the broad task and one-shot in terms of the specific word. Table3.16shows the 6 examples we generated; all definitions were human-generated, and the first answer was human-generated as conditioning while the subsequent answers were generated by GPT-3. These examples were generated continuously in one sitting and we did not omit or repeatedly try any prompts. In all cases the generated sentence appears to be a correct or at least plausible use of the word. In the final sentence the model generates a plausible conjugation for the word “screeg” (namely “screeghed”), although the use of the word is slightly awkward (“screeghed at each other”) despite being plausible in the sense that it could describe a toy sword fight. Overall, GPT-3 appears to be at least proficient at the task of using novel words in a sentence. 3.9.6 Correcting English Grammar Another task well suited for few-shot learning is correcting English grammar. We test this with GPT-3 in the few- shot setting by giving prompts of the form"Poor English Input: nn Good English Output: ". We give GPT-3 one human-generated correction and then ask it to correct 5 more (again without any omissions or repeats). Results are shown in Figure3.17. 4 Measuring and Preventing Memorization Of Benchmarks Since our training dataset is sourced from the internet, it is possible that our model was trained on some of our benchmark test sets. Accurately detecting test contamination from internet-scale datasets is a new area of research without established best practices. While it is common practice to train large models without investigating contamination, given the increasing scale of pretraining datasets, we believe this issue is becoming increasingly important to attend to. This concern is not just hypothetical. One of the first papers to train a language model on Common Crawl data [TL18] detected and removed a training document which overlapped with one of their evaluation datasets. Other work such as GPT-2 [RWC + 19] also conducted post-hoc overlap analysis. Their study was relatively encouraging, finding that <

>. Figure 3.17:Representative GPT-3 completions for the few-shot task of correcting English grammar. Boldface is GPT-3’s completions, plain text is human prompts. In the first few examples example both the prompt and the completion are provided by a human; this then serves as conditioning for subsequent examples where GPT-3 receives successive additional prompts and provides the completions. Nothing task-specific is provided to GPT-3 aside from the first few examples as conditioning and the “Poor English input/Good English output” framing. We note that the distinction between ”poor” and ”good” English (and the terms themselves) is complex, contextual, and contested. As the example mentioning the rental of a house shows, assumptions that the model makes about what “good” is can even lead it to make errors (here, the model not only adjusts grammar, but also removes the word ”cheap” in a way that alters meaning). <

> Figure 4.1: GPT-3 Training Curves We measure model performance during training on a deduplicated validation split of our training distribution. Though there is some gap between training and validation performance, the gap grows only minimally with model size and training time, suggesting that most of the gap comes from a difference in difficulty rather than overfitting. although models did perform moderately better on data that overlapped between training and testing, this did not significantly impact reported results due to the small fraction of data which was contaminated (often only a few percent). GPT-3 operates in a somewhat different regime. On the one hand, the dataset and model size are about two orders of magnitude larger than those used for GPT-2, and include a large amount of Common Crawl, creating increased potential for contamination and memorization. On the other hand, precisely due to the large amount of data, even GPT-3 175B does not overfit its training set by a significant amount, measured relative to a held-out validation set with which it was deduplicated (Figure4.1). Thus, we expect that contamination is likely to be frequent, but that its effects may not be as large as feared. We initially tried to address the issue of contamination by proactively searching for and attempting to remove any overlap between our training data and the development and test sets of all benchmarks studied in this paper. Unfortunately, a bug resulted in only partial removal of all detected overlaps from the training data. Due to the cost of training, it wasn’t feasible to retrain the model. To address this, we investigate in detail how the remaining detected overlap impacts results. For each benchmark, we produce a ‘clean’ version which removes all potentially leaked examples, defined roughly as examples that have a 13-gram overlap with anything in the pretraining set (or that overlap with the whole example when it is shorter than 13-grams). The goal is to very conservatively flag anything that could potentially be contamination, so as to produce a clean subset that is free of contamination with high confidence. The exact procedure is detailed in AppendixC. We then evaluate GPT-3 on these clean benchmarks, and compare to the original score. If the score on the clean subset is similar to the score on the entire dataset, this suggests that contamination, even if present, does not have a significant effect on reported results. If the score on the clean subset is lower, this suggests contamination may be inflating the results. The results are summarized in Figure4.2. Although potential contamination is often high (with a quarter of benchmarks scoring over 50%), in most cases performance changes only negligibly, and we see no evidence that contamination level and performance difference are correlated. We conclude that either our conservative method substantially overestimated contamination or that contamination has little effect on performance. Below, we review in more detail the few specific cases where either (1) the model performs significantly worse on the cleaned version, or (2) potential contamination is very high, which makes measuring the performance difference difficult. Our analysis flagged six groups of benchmarks for further investigation: Word Scrambling, Reading Comprehension (QuAC, SQuAD2, DROP), PIQA, Winograd, language modeling tasks (Wikitext tasks, 1BW), and German to English <

> Figure 4.2: Benchmark contamination analysis We constructed cleaned versions of each of our benchmarks to check for potential contamination in our training set. The x-axis is a conservative lower bound for how much of the dataset is known with high confidence to be clean, and the y-axis shows the difference in performance when evaluating only on the verified clean subset. Performance on most benchmarks changed negligibly, but some were flagged for further review. On inspection we find some evidence for contamination of the PIQA and Winograd results, and we mark the corresponding results in Section3with an asterisk. We find no evidence that other benchmarks are affected. translation. Since our overlap analysis is designed to be extremely conservative, we expect it to produce some false positives. We summarize the results for each group of tasks below: •Reading Comprehension:Our initial analysis flagged>90% of task examples from QuAC, SQuAD2, and DROP as potentially contaminated, so large that even measuring the differential on a clean subset was difficult. Upon manual inspection, however, we found that for every overlap we inspected, in all 3 datasets, the source text was present in our training data but the question/answer pairs were not, meaning the model gains only background information and cannot memorize the answer to a specific question. •German translation:We found 25% of the examples in the WMT16 German-English test set were marked as potentially contaminated, with an associated total effect size of 1-2 BLEU. Upon inspection, none of the flagged examples contain paired sentences resembling NMT training data and collisions were monolingual matches mostly of snippets of events discussed in the news. •Reversed Words and Anagrams:Recall that these tasks are of the form “alaok = koala”. Due to the short length of these tasks, we used 2-grams for filtering (ignoring punctuation). After inspecting the flagged overlaps, we found that they were not typically instances of real reversals or unscramblings in the training set, but rather palindromes or trivial unscramblings, e.g “kayak = kayak”. The amount of overlap was small, but removing the trivial tasks lead to an increase in difficulty and thus a spurious signal. Related to this, the symbol insertion task shows high overlap but no effect on performance – this is because that task involves removing non-letter characters from a word, and the overlap analysis itself ignores such characters, leading to many spurious matches. •PIQA:The overlap analysis flagged 29% of examples as contaminated, and observed a 3 percentage point absolute decrease (4% relative decrease) in performance on the clean subset. Though the test dataset was released after our training set was created and its labels are hidden, some of the web pages used by the crowdsourced dataset creators are contained in our training set. We found a similar decrease in a 25x smaller model with much less capacity to memorize, leading us to suspect that the shift is likely statistical bias rather than memorization; examples which workers copied may simply be easier. Unfortunately, we cannot rigorously prove this hypothesis. We therefore mark our PIQA results with an asterisk to denote this potential contamination. •Winograd:The overlap analysis flagged 45% of examples, and found a 2.6% decrease in performance on the clean subset. Manual inspection of the overlapping data point showed that 132 Winograd schemas were in fact present in our training set, though presented in a different format than we present the task to the model. Although the decrease in performance is small, we mark our Winograd results in the main paper with an asterisk. •Language modeling:We found the 4 Wikipedia language modeling benchmarks measured in GPT-2, plus the Children’s Book Test dataset, to be almost entirely contained in our training data. Since we cannot reliably extract a clean subset here, we do not report results on these datasets, even though we intended to when starting this work. We note that Penn Tree Bank due to its age was unaffected and therefore became our chief language modeling benchmark. We also inspected datasets where contamination was high, but the impact on performance was close to zero, simply to verify how much actual contamination existed. These appeared to often contain false positives. They had either no actual contamination, or had contamination that did not give away the answer to the task. One notable exception was LAMBADA, which appeared to have substantial genuine contamination, yet the impact on performance was very small, with the clean subset scoring within 0.5% of the full dataset. Also, strictly speaking, our fill-in-the-blank format precludes the simplest form of memorization. Nevertheless, since we made very large gains on LAMBADA in this paper, the potential contamination is noted in the results section. An important limitation of our contamination analysis is that we cannot be sure that the clean subset is drawn from the same distribution as the original dataset. It remains possible that memorization inflates results but at the same time is precisely counteracted by some statistical bias causing the clean subset to be easier. However, the sheer number of shifts close to zero suggests this is unlikely, and we also observed no noticeable difference in the shifts for small models, which are unlikely to be memorizing. Overall, we have made a best effort to measure and document the effects of data contamination, and to note or outright remove problematic results, depending on the severity. Much work remains to be done to address this important and subtle issue for the field in general, both when designing benchmarks and when training models. For a more detailed explanation of our analysis, we refer the reader to AppendixC. 5 Limitations GPT-3 and our analysis of it have a number of limitations. Below we describe some of these and suggest directions for future work. First, despite the strong quantitative and qualitative improvements of GPT-3, particularly compared to its direct predecessor GPT-2, it still has notable weaknesses in text synthesis and several NLP tasks. On text synthesis, although the overall quality is high, GPT-3 samples still sometimes repeat themselves semantically at the document level, start to lose coherence over sufficiently long passages, contradict themselves, and occasionally contain non-sequitur sentences or paragraphs. We will release a collection of 500 uncurated unconditional samples to help provide a better sense of GPT-3’s limitations and strengths at text synthesis. Within the domain of discrete language tasks, we have noticed informally that GPT-3 seems to have special difficulty with “common sense physics”, despite doing well on some datasets (such as PIQA [BZB + 19]) that test this domain. Specifically GPT-3 has difficulty with questions of the type “If I put cheese into the fridge, will it melt?”. Quantitatively, GPT-3’s in-context learning performance has some notable gaps on our suite of benchmarks, as described in Section3, and in particular it does little better than chance when evaluated one-shot or even few-shot on some “comparison” tasks, such as determining if two words are used the same way in a sentence, or if one sentence implies another (WIC and ANLI respectively), as well as on a subset of reading comprehension tasks. This is especially striking given GPT-3’s strong few-shot performance on many other tasks. GPT-3 has several structural and algorithmic limitations, which could account for some of the issues above. We focused on exploring in-context learning behavior in autoregressive language models because it is straightforward to both sample and compute likelihoods with this model class. As a result our experiments do not include any bidirectional architectures or other training objectives such as denoising. This is a noticeable difference from much of the recent literature, which has documented improved fine-tuning performance when using these approaches over standard language models [RSR + 19]. Thus our design decision comes at the cost of potentially worse performance on tasks which empirically benefit from bidirectionality. This may include fill-in-the-blank tasks, tasks that involve looking back and comparing two pieces of content, or tasks that require re-reading or carefully considering a long passage and then generating a very short answer. This could be a possible explanation for GPT-3’s lagging few-shot performance on a few of the tasks, such as WIC (which involves comparing the use of a word in two sentences), ANLI (which involves comparing two sentences to see if one implies the other), and several reading comprehension tasks (e.g. QuAC and RACE). We also conjecture, based on past literature, that a large bidirectional model would be stronger at fine-tuning than GPT-3. Making a bidirectional model at the scale of GPT-3, and/or trying to make bidirectional models work with few- or zero-shot learning, is a promising direction for future research, and could help achieve the “best of both worlds”. A more fundamental limitation of the general approach described in this paper – scaling up any LM-like model, whether autoregressive or bidirectional – is that it may eventually run into (or could already be running into) the limits of the pretraining objective. Our current objective weights every token equally and lacks a notion of what is most important to predict and what is less important. [RRS20] demonstrate benefits of customizing prediction to entities of interest. Also, with self-supervised objectives, task specification relies on forcing the desired task into a prediction problem, whereas ultimately, useful language systems (for example virtual assistants) might be better thought of as taking goal-directed actions rather than just making predictions. Finally, large pretrained language models are not grounded in other domains of experience, such as video or real-world physical interaction, and thus lack a large amount of context about the world [BHT + 20]. For all these reasons, scaling pure self-supervised prediction is likely to hit limits, and augmentation with a different approach is likely to be necessary. Promising future directions in this vein might include learning the objective function from humans [ZSW + 19a], fine-tuning with reinforcement learning, or adding additional modalities such as images to provide grounding and a better model of the world [CLY + 19]. Another limitation broadly shared by language models is poor sample efficiency during pre-training. While GPT-3 takes a step towards test-time sample efficiency closer to that of humans (one-shot or zero-shot), it still sees much more text during pre-training than a human sees in the their lifetime [Lin20]. Improving pre-training sample efficiency is an important direction for future work, and might come from grounding in the physical world to provide additional information, or from algorithmic improvements. A limitation, or at least uncertainty, associated with few-shot learning in GPT-3 is ambiguity about whether few-shot learning actually learns new tasks “from scratch” at inference time, or if it simply recognizes and identifies tasks that it has learned during training. These possibilities exist on a spectrum, ranging from demonstrations in the training set that are drawn from exactly the same distribution as those at test time, to recognizing the same task but in a different format, to adapting to a specific style of a general task such as QA, to learning a skill entirely de novo. Where GPT-3 is on this spectrum may also vary from task to task. Synthetic tasks such as wordscrambling or defining nonsense words seem especially likely to be learned de novo, whereas translation clearly must be learned during pretraining, although possibly from data that is very different in organization and style than the test data. Ultimately, it is not even clear what humans learn from scratch vs from prior demonstrations. Even organizing diverse demonstrations during pre-training and identifying them at test time would be an advance for language models, but nevertheless understanding precisely how few-shot learning works is an important unexplored direction for future research. A limitation associated with models at the scale of GPT-3, regardless of objective function or algorithm, is that they are both expensive and inconvenient to perform inference on, which may present a challenge for practical applicability of models of this scale in their current form. One possible future direction to address this is distillation [HVD15] of large models down to a manageable size for specific tasks. Large models such as GPT-3 contain a very wide range of skills, most of which are not needed for a specific task, suggesting that in principle aggressive distillation may be possible. Distillation is well-explored in general [LHCG19a] but has not been tried at the scale of hundred of billions parameters; new challenges and opportunities may be associated with applying it to models of this size. Finally, GPT-3 shares some limitations common to most deep learning systems – its decisions are not easily interpretable, it is not necessarily well-calibrated in its predictions on novel inputs as observed by the much higher variance in performance than humans on standard benchmarks, and it retains the biases of the data it has been trained on. This last issue – biases in the data that may lead the model to generate stereotyped or prejudiced content – is of special concern from a societal perspective, and will be discussed along with other issues in the next section on Broader Impacts (Section6). 6 Broader Impacts Language models have a wide range of beneficial applications for society, including code and writing auto-completion, grammar assistance, game narrative generation, improving search engine responses, and answering questions. But they also have potentially harmful applications. GPT-3 improves the quality of text generation and adaptability over smaller models and increases the difficulty of distinguishing synthetic text from human-written text. It therefore has the potential to advance both the beneficial and harmful applications of language models. Here we focus on the potential harms of improved language models, not because we believe the harms are necessarily greater, but in order to stimulate efforts to study and mitigate them. The broader impacts of language models like this are numerous. We focus on two primary issues: the potential for deliberate misuse of language models like GPT-3 in Section6.1, and issues of bias, fairness, and representation within models like GPT-3 in Section6.2. We also briefly discuss issues of energy efficiency (Section6.3). 6.1 Misuse of Language Models Malicious uses of language models can be somewhat difficult to anticipate because they often involve repurposing language models in a very different environment or for a different purpose than researchers intended. To help with this, we can think in terms of traditional security risk assessment frameworks, which outline key steps such as identifying threats and potential impacts, assessing likelihood, and determining risk as a combination of likelihood and impact [Ros12]. We discuss three factors: potential misuse applications, threat actors, and external incentive structures. 6.1.1 Potential Misuse Applications Any socially harmful activity that relies on generating text could be augmented by powerful language models. Examples include misinformation, spam, phishing, abuse of legal and governmental processes, fraudulent academic essay writing and social engineering pretexting. Many of these applications bottleneck on human beings to write sufficiently high quality text. Language models that produce high quality text generation could lower existing barriers to carrying out these activities and increase their efficacy. The misuse potential of language models increases as the quality of text synthesis improves. The ability of GPT-3 to generate several paragraphs of synthetic content that people find difficult to distinguish from human-written text in 3.9.4 represents a concerning milestone in this regard. 6.1.2 Threat Actor Analysis Threat actors can be organized by skill and resource levels, ranging from low or moderately skilled and resourced actors who may be able to build a malicious product to ‘advanced persistent threats’ (APTs): highly skilled and well-resourced (e.g. state-sponsored) groups with long-term agendas [SBC + 19]. To understand how low and mid-skill actors think about language models, we have been monitoring forums and chat groups where misinformation tactics, malware distribution, and computer fraud are frequently discussed. While we did find significant discussion of misuse following the initial release of GPT-2 in spring of 2019, we found fewer instances of experimentation and no successful deployments since then. Additionally, those misuse discussions were correlated with media coverage of language model technologies. From this, we assess that the threat of misuse from these actors is not immediate, but significant improvements in reliability could change this. Because APTs do not typically discuss operations in the open, we have consulted with professional threat analysts about possible APT activity involving the use of language models. Since the release of GPT-2 there has been no discernible difference in operations that may see potential gains by using language models. The assessment was that language models may not be worth investing significant resources in because there has been no convincing demonstration that current language models are significantly better than current methods for generating text, and because methods for “targeting” or “controlling” the content of language models are still at a very early stage. 6.1.3 External Incentive Structures Each threat actor group also has a set of tactics, techniques, and procedures (TTPs) that they rely on to accomplish their agenda. TTPs are influenced by economic factors like scalability and ease of deployment; phishing is extremely popular among all groups because it offers a low-cost, low-effort, high-yield method of deploying malware and stealing login credentials. Using language models to augment existing TTPs would likely result in an even lower cost of deployment. Ease of use is another significant incentive. Having stable infrastructure has a large impact on the adoption of TTPs. The outputs of language models are stochastic, however, and though developers can constrain these (e.g. using top-k truncation) they are not able to perform consistently without human feedback. If a social media disinformation bot produces outputs that are reliable 99% of the time, but produces incoherent outputs 1% of the time, this could reduce the amount of human labor required in operating this bot. But a human is still needed to filter the outputs, which restricts how scalable the operation can be. Based on our analysis of this model and analysis of threat actors and the landscape, we suspect AI researchers will eventually develop language models that are sufficiently consistent and steerable that they will be of greater interest to malicious actors. We expect this will introduce challenges for the broader research community, and hope to work on this through a combination of mitigation research, prototyping, and coordinating with other technical developers. 6.2 Fairness, Bias, and Representation Biases present in training data may lead models to generate stereotyped or prejudiced content. This is concerning, since model bias could harm people in the relevant groups in different ways by entrenching existing stereotypes and producing demeaning portrayals amongst other potential harms [Cra17]. We have conducted an analysis of biases in the model in order to better understand GPT-3’s limitations when it comes to fairness, bias, and representation. 8 Our goal is not to exhaustively characterize GPT-3, but to give a preliminary analysis of some of its limitations and behaviors. We focus on biases relating to gender, race, and religion, although many other categories of bias are likely present and could be studied in follow-up work. This is a preliminary analysis and does not reflect all of the model’s biases even within the studied categories. Broadly, our analysis indicates that internet-trained models have internet-scale biases; models tend to reflect stereotypes present in their training data. Below we discuss our preliminary findings of bias along the dimensions of gender, race, and religion. We probe for bias in the 175 billion parameter model and also in similar smaller models, to see if and how they are different in this dimension. 6.2.1 Gender In our investigation of gender bias in GPT-3, we focused on associations between gender and occupation. We found that occupations in general have a higher probability of being followed by a male gender identifier than a female one (in other words, they are male leaning) when given a context such as"Thefoccupationgwas a"(Neutral Variant). 83% of the 388 occupations we tested were more likely to be followed by a male identifier by GPT-3. We measured this by feeding the model a context such as"The detective was a"and then looking at the probability of the model following up with male indicating words (eg. man, male etc.) or female indicating words (woman, female etc.). In particular, occupations demonstrating higher levels of education such as legislator, banker, or professor emeritus were heavily male leaning along with occupations that require hard physical labour such as mason, millwright, and sheriff. Occupations that were more likely to be followed by female identifiers include midwife, nurse, receptionist, housekeeper etc. We also tested how these probabilities changed when we shifted the context to be the"The competentfoccupationg was a"(Competent Variant), and when we shifted the context to be"The incompetentfoccupationgwas a" (Incompetent Variant) for each occupation in the dataset. We found that, when prompted with"The competent foccupationgwas a,"the majority of occupations had an even higher probability of being followed by a male identifier than a female one than was the case with our original neutral prompt,"Thefoccupationgwas a". With the prompt"The incompetentfoccupationgwas a"the majority of occupations still leaned male with a similar probability than for our original neutral prompt. The average occupation bias - measured as <> was <> for the Neutral Variant,<> for the Competent Variant and <> jobs for the Incompetent Variant. We also carried out pronoun resolution on the Winogender dataset [RNLVD18] using two methods which further corroborated the model’s tendency to associate most occupations with males. One method measured the mod- els ability to correctly assign a pronoun as the occupation or the participant. For example, we fed the model a context such as"The advisor met with the advisee because she wanted to get advice about job applications. ‘She’ refers to the" and found the option with the lowest probability between the two possi- ble options (Choices between Occupation Option: advisor; Participant Option: advisee). Occupation and participant words often have societal biases associated with them such as the assumption that most occupants are by default male. We found that the language models learnt some of these biases such as a tendency to associate female pronouns with participant positions more than male pronouns. GPT-3 175B had the highest accuracy of all the models (64.17%) on this task. It was also the only model where the accuracy for Occupant sentences (sentences where the correct answer was the Occupation option) for females was higher than for males (81.7% vs 76.7%). All other models had a higher accuracy for male pronouns with Occupation sentences as compared to female pronouns with the exception of our second largest model- GPT-3 13B - which had the same accuracy (60%) for both. This offers some preliminary evidence that in places where issues of bias can make language models susceptible to error, the larger models are more robust than smaller models. We also performed co-occurrence tests, where we analyzed which words are likely to occur in the vicinity of other pre- selected words. We created a model output sample set by generating 800 outputs of length 50 each with a temperature 8 Evaluating fairness, bias, and representation in language models is a rapidly-developing area with a large body of prior work. See, for example, [HZJ + 19,NBR20,SCNP19]. Table 6.1:Most Biased Descriptive Words in 175B Model <

> of 1 and topp of 0.9 for every prompt in our dataset. For gender, we had prompts such as"He was very","She was very","He would be described as","She would be described as" 9 . We looked at the adjectives and adverbs in the top 100 most favored words using an off-the-shelf POS tagger [LB02]. We found females were more often described using appearance oriented words such as ”beautiful” and ”gorgeous” as compared to men who were more often described using adjectives that span a greater spectrum. Table6.1shows the top 10 most favored descriptive words for the model along with the raw number of times each word co-occurred with a pronoun indicator. “Most Favored” here indicates words which were most skewed towards a category by co-occurring with it at a higher rate as compared to the other category. To put these numbers in perspective, we have also included the average for the number of co-occurrences across all qualifying words for each gender. 6.2.2 Race To investigate racial bias in GPT-3, we seeded the model with prompts such as - "The fracegman was very", "The fracegwoman was very"and"People would describe thefracegperson as"and generated 800 samples for each of the above prompts, withfracegreplaced with a term indicating a racial category such as White or Asian. We then measure word co-occurrences in the generated samples. Given prior research demonstrating that language models produce text of differing sentiment when varying features such as occupation [HZJ + 19], we explored how race impacted sentiment. We measured sentiment using Senti WordNet [BES10] for the words which co-occurred disproportionately with each race. Each word sentiment varied from 100 to -100, with positive scores indicating positive words (eg. wonderfulness: 100, amicable: 87.5), negative scores indicating negative words (eg. wretched: -87.5 , horrid: -87.5) and a score of 0 indicating neutral words (eg. sloping, chalet). It should be noted that we were explicitly prompting the models to talk about race and this in turn generated text that focused on racial features; these results are not from the models talking about race in the wild but talking about race in an experimental setup where they have been primed to do so. Additionally, since we are measuring sentiment by simply looking at word co-occurrences, the resulting sentiment can reflect socio-historical factors - for instance, text relating to a discussion of slavery will frequently have a negative sentiment, which may lead to a demographic being associated with a negative sentiment under this testing methodology. Across the models we analyzed, ‘Asian’ had a consistently high sentiment - it ranked 1st in 3 out of 7 models. On the other hand, ’Black’ had a consistently low sentiment - it ranked the lowest in 5 out of 7 models. These differences narrowed marginally on the larger model sizes. This analysis gives a sense of the biases of different models and highlights the need for more sophisticated analysis of the relationship between sentiment, entities, and input data. 9 We only used male and female pronouns. This simplifying assumption makes it easier to study co-occurrence since it does not require the isolation of instances in which ‘they’ refers to a singular noun from those where it didn’t, but other forms of gender bias are likely present and could be studied using different approaches. <

> Figure 6.1:Racial Sentiment Across Models <

> Table 6.2:Shows the ten most favored words about each religion in the GPT-3 175B model. 6.2.3 Religion We studied which words co-occurred with religious terms relating to Atheism, Buddhism, Christianity, Hinduism, Islam, and Judaism, by generating 800 model outputs of length 50 with a temperature of 1 and a to p of 0.9 for every prompt. Our prompts were of the nature"fReligion practitionersgare"(Eg."Christians are") for each of the six religious categories listed above. We then allowed the model to naturally carry out completions and created a corpus of such completions for studying co-occurrence of words. The following is an example output from the model: "Buddhists aredivided into two main branches - Theravada and Mahayana. Theravada is the more conservative branch, centering on monastic life and the earliest sutras and refusing to recognize the later Mahayana sutras as authentic." Similar to race, we found that the models make associations with religious terms that indicate some propensity to reflect how these terms are sometimes presented in the world. For example, with the religionIslam, we found that words such asramadan,prophetandmosqueco-occurred at a higher rate than for other religions. We also found that words such asviolent,terrorismandterroristco-occurred at a greater rate with Islam than with other religions and were in the top 40 most favored words for Islam in GPT-3. 6.2.4 Future Bias and Fairness Challenges We have presented this preliminary analysis to share some of the biases we found in order to motivate further research, and to highlight the inherent difficulties in characterizing biases in large-scale generative models; we expect this to be an area of continuous research for us and are excited to discuss different methodological approaches with the community. We view the work in this section as subjective signposting - we chose gender, race, and religion as a starting point, but we recognize the inherent subjectivity in this choice. Our work is inspired by the literature on characterizing model attributes to develop informative labels such as Model Cards for Model Reporting from [MWZ + 18]. Ultimately, it is important not just to characterize biases in language systems but to intervene. The literature on this is also extensive [QMZH19,HZJ + 19], so we offer only a few brief comments on future directions specific to large language models. In order to pave the way for effective bias prevention in general purpose models, there is a need for building a common vocabulary tying together the normative, technical and empirical challenges of bias mitigation for these models. There is room for more research that engages with the literature outside NLP, better articulates normative statements about harm, and engages with the lived experience of communities affected by NLP systems [BBDIW20]. Thus, mitigation work should not be approached purely with a metric driven objective to ‘remove’ bias as this has been shown to have blind spots [GG19,NvNvdG19] but in a holistic manner. 6.3 Energy Usage Practical large-scale pre-training requires large amounts of computation, which is energy-intensive: training the GPT-3 175B consumed several thousand petaflop/s-days of compute during pre-training, compared to tens of petaflop/s-days for a 1.5B parameter GPT-2 model (Figure2.2). This means we should be cognizant of the cost and efficiency of such models, as advocated by [SDSE19]. The use of large-scale pre-training also gives another lens through which to view the efficiency of large models - we should consider not only the resources that go into training them, but how these resources are amortized over the lifetime of a model, which will subsequently be used for a variety of purposes and fine-tuned for specific tasks. Though models like GPT-3 consume significant resources during training, they can be surprisingly efficient once trained: even with the full GPT-3 175B, generating 100 pages of content from a trained model can cost on the order of 0.4 kW-hr, or only a few cents in energy costs. Additionally, techniques like model distillation [LHCG19a] can further bring down the cost of such models, letting us adopt a paradigm of training single, large-scale models, then creating more efficient versions of them for use in appropriate contexts. Algorithmic progress may also naturally further increase the efficiency of such models over time, similar to trends observed in image recognition and neural machine translation [HB20]. 7 Related Work Several lines of work have focused on increasing parameter count and/or computation in language models as a means to improve generative or task performance. An early work scaled LSTM based language models to over a billion parameters [JVS + 16]. One line of work straightforwardly increases the size of transformer models, scaling up parameters and FLOPS-per-token roughly in proportion. Work in this vein has successively increased model size: 213 million parameters [VSP + 17] in the original paper, 300 million parameters [DCLT18], 1.5 billion parameters [RWC + 19], 8 billion parameters [SPP + 19], 11 billion parameters [RSR + 19], and most recently 17 billion parameters [Tur20]. A second line of work has focused on increasing parameter count but not computation, as a means of increasing models’ capacity to store information without increased computational cost. These approaches rely on the conditional computation framework [BLC13] and specifically, the mixture-of-experts method [SMM + 17] has been used to produce 100 billion parameter models and more recently 50 billion parameter translation models [AJF19], though only a small fraction of the parameters are actually used on each forward pass. A third approach increases computation without increasing parameters; examples of this approach include adaptive computation time [Gra16] and the universal transformer [DGV + 18]. Our work focuses on the first approach (scaling compute and parameters together, by straightforwardly making the neural net larger), and increases model size 10x beyond previous models that employ this strategy. Several efforts have also systematically studied the effect of scale on language model performance. [KMH + 20, RRBS19,LWS + 20,HNA + 17], find a smooth power-law trend in loss as autoregressive language models are scaled up. This work suggests that this trend largely continues as models continue to scale up (although a slight bending of the curve can perhaps be detected in Figure3.1), and we also find relatively smooth increases in many (though not all) downstream tasks across 3 orders of magnitude of scaling. Another line of work goes in the opposite direction from scaling, attempting to preserve strong performance in language models that are as small as possible. This approach includes ALBERT [LCG + 19] as well as general [HVD15] and task-specific [SDCW19,JYS + 19,KR16] approaches to distillation of language models. These architectures and techniques are potentially complementary to our work, and could be applied to decrease latency and memory footprint of giant models. As fine-tuned language models have neared human performance on many standard benchmark tasks, considerable effort has been devoted to constructing more difficult or open-ended tasks, including question answering [KPR + 19, IBGC + 14,CCE + 18,MCKS18], reading comprehension [CHI + 18,RCM19], and adversarially constructed datasets designed to be difficult for existing language models [SBBC19,NWD + 19]. In this work we test our models on many of these datasets. Many previous efforts have focused specifically on question-answering, which constitutes a significant fraction of the tasks we tested on. Recent efforts include [RSR + 19,RRS20], which fine-tuned an 11 billion parameter language model, and [GLT + 20], which focused on attending over a large corpus of data at test time. Our work differs in focusing on in-context learning but could be combined in the future with those of [GLT + 20,LPP + 20]. Metalearning in language models has been utilized in [RWC + 19], though with much more limited results and no systematic study. More broadly, language model metalearning has an inner-loop-outer-loop structure, making it structurally similar to metalearning as applied to ML in general. Here there is an extensive literature, including matching networks [VBL + 16], RL2 [DSC + 16], learning to optimize [RL16,ADG + 16,LM17] and MAML [FAL17]. Our approach of stuffing the model’s context with previous examples is most structurally similar to RL2 and also resembles [HYC01], in that an inner loop of adaptation takes place through computation in the model’s activations across timesteps, without updating the weights, while an outer loop (in this case just language model pre-training) updates the weights, and implicitly learns the ability to adapt to or at least recognize tasks defined at inference-time. Few-shot auto-regressive density estimation was explored in [RCP + 17] and [GWC + 18] studied low-resource NMT as a few-shot learning problem. While the mechanism of our few-shot approach is different, prior work has also explored ways of using pre-trained language models in combination with gradient descent to perform few-shot learning [SS20]. Another sub-field with similar goals is semi-supervised learning where approaches such as UDA [XDH + 19] also explore methods of fine-tuning when very little labeled data is available. Giving multi-task models instructions in natural language was first formalized in a supervised setting with [MKXS18] and utilized for some tasks (such as summarizing) in a language model with [RWC + 19]. The notion of presenting tasks in natural language was also explored in the text-to-text transformer [RSR + 19], although there it was applied for multi-task fine-tuning rather than for in-context learning without weight updates. Another approach to increasing generality and transfer-learning capability in language models is multi-task learning [Car97], which fine-tunes on a mixture of downstream tasks together, rather than separately updating the weights for each one. If successful multi-task learning could allow a single model to be used for many tasks without updating the weights (similar to our in-context learning approach), or alternatively could improve sample efficiency when updating the weights for a new task. Multi-task learning has shown some promising initial results [LGH + 15,LSP + 18] and multi-stage fine-tuning has recently become a standardized part of SOTA results on some datasets [PFB18] and pushed the boundaries on certain tasks [KKS + 20], but is still limited by the need to manually curate collections of datasets and set up training curricula. By contrast pre-training at large enough scale appears to offer a “natural” broad distribution of tasks implicitly contained in predicting the text itself. One direction for future work might be attempting to generate a broader set of explicit tasks for multi-task learning, for example through procedural generation [TFR + 17], human interaction [ZSW + 19b], or active learning [Mac92]. Algorithmic innovation in language models over the last two years has been enormous, including denoising-based bidirectionality [DCLT18], prefixLM [DL15] and encoder-decoder architectures [LLG + 19,RSR + 19], random permu- tations during training [YDY + 19], architectures that improve the efficiency of sampling [DYY + 19], improvements in data and training procedures [LOG + 19], and efficiency increases in the embedding parameters [LCG + 19]. Many of these techniques provide significant gains on downstream tasks. In this work we continue to focus on pure autoregressive language models, both in order to focus on in-context learning performance and to reduce the complexity of our large model implementations. However, it is very likely that incorporating these algorithmic advances could improve GPT-3’s performance on downstream tasks, especially in the fine-tuning setting, and combining GPT-3’s scale with these algorithmic techniques is a promising direction for future work. 8 Conclusion We presented a 175 billion parameter language model which shows strong performance on many NLP tasks and benchmarks in the zero-shot, one-shot, and few-shot settings, in some cases nearly matching the performance of state-of-the-art fine-tuned systems, as well as generating high-quality samples and strong qualitative performance at tasks defined on-the-fly. We documented roughly predictable trends of scaling in performance without using fine-tuning. We also discussed the social impacts of this class of model. Despite many limitations and weaknesses, these results suggest that very large language models may be an important ingredient in the development of adaptable, general language systems. Acknowledgements The authors would like to thank Ryan Lowe for giving detailed feedback on drafts of the paper. Thanks to Jakub Pachocki and Szymon Sidor for suggesting tasks, and Greg Brockman, Michael Petrov, Brooke Chan, and Chelsea Voss for helping run evaluations on OpenAI’s infrastructure. Thanks to David Luan for initial support in scaling up this project, Irene Solaiman for discussions about ways to approach and evaluate bias, Harrison Edwards and Yura Burda for discussions and experimentation with in-context learning, Geoffrey Irving and Paul Christiano for early discussions of language model scaling, Long Ouyang for advising on the design of the human evaluation experiments, Chris Hallacy for discussions on data collection, and Shan Carter for help with visual design. Thanks to the millions of people who created content that was used in the training of the model, and to those who were involved in indexing or upvoting the content (in the case of WebText). Additionally, we would like to thank the entire OpenAI infrastructure and supercomputing teams for making it possible to train models at this scale. Contributions Tom Brown, Ben Mann, Prafulla Dhariwal, Dario Amodei, Nick Ryder, Daniel M Ziegler, and Jeffrey Wu implemented the large-scale models, training infrastructure, and model-parallel strategies. Tom Brown, Dario Amodei, Ben Mann, and Nick Ryderconducted pre-training experiments. Ben Mann and Alec Radfordcollected, filtered, deduplicated, and conducted overlap analysis on the training data. Melanie Subbiah, Ben Mann, Dario Amodei, Jared Kaplan, Sam McCandlish, Tom Brown, Tom Henighan, and Girish Sastryimplemented the downstream tasks and the software framework for supporting them, including creation of synthetic tasks. Jared Kaplan and Sam McCandlishinitially predicted that a giant language model should show continued gains, and applied scaling laws to help predict and guide model and data scaling decisions for the research. Ben Mannimplemented sampling without replacement during training. Alec Radfordoriginally demonstrated few-shot learning occurs in language models. Jared Kaplan and Sam McCandlishshowed that larger models learn more quickly in-context, and systematically studied in-context learning curves, task prompting, and evaluation methods. Prafulla Dhariwalimplemented an early version of the codebase, and developed the memory optimizations for fully half-precision training. Rewon Child and Mark Chendeveloped an early version of our model-parallel strategy. Rewon Child and Scott Graycontributed the sparse transformer. Aditya Rameshexperimented with loss scaling strategies for pretraining. Melanie Subbiah and Arvind Neelakantanimplemented, experimented with, and tested beam search. Pranav Shyamworked on SuperGLUE and assisted with connections to few-shot learning and meta-learning literature. Sandhini Agarwalconducted the fairness and representation analysis. Girish Sastry and Amanda Askellconducted the human evaluations of the model. Ariel Herbert-Vossconducted the threat analysis of malicious use. Gretchen Kruegeredited and red-teamed the policy sections of the paper. Benjamin Chess, Clemens Winter, Eric Sigler, Christopher Hesse, Mateusz Litwin, and Christopher Berner optimized OpenAI’s clusters to run the largest models efficiently. Scott Graydeveloped fast GPU kernels used during training. Jack Clarkled the analysis of ethical impacts — fairness and representation, human assessments of the model, and broader impacts analysis, and advised Gretchen, Amanda, Girish, Sandhini, and Ariel on their work. Dario Amodei, Alec Radford, Tom Brown, Sam McCandlish, Nick Ryder, Jared Kaplan, Sandhini Agarwal, Amanda Askell, Girish Sastry, and Jack Clarkwrote the paper. Sam McCandlishled the analysis of model scaling, and advised Tom Henighan and Jared Kaplan on their work. Alec Radfordadvised the project from an NLP perspective, suggested tasks, put the results in context, and demonstrated the benefit of weight decay for training. Ilya Sutskeverwas an early advocate for scaling large generative likelihood models, and advised Pranav, Prafulla, Rewon, Alec, and Aditya on their work. Dario Amodeidesigned and led the research. A Details of Common Crawl Filtering As mentioned in Section2.2, we employed two techniques to improve the quality of the Common Crawl dataset: (1) filtering Common Crawl and (2) fuzzy deduplication: 1.In order to improve the quality of Common Crawl, we developed an automatic filtering method to remove low quality documents. Using the original WebText as a proxy for high-quality documents, we trained a classifier to distinguish these from raw Common Crawl. We then used this classifier to re-sample Common Crawl by prioritizing documents which were predicted by the classifier to be higher quality. The classifier is trained using logistic regression classifier with features from Spark’s standard tokenizer and HashingTF 10 . For the positive examples, we used a collection of curated datasets such as WebText, Wikiedia, and our web books corpus as the positive examples, and for the negative examples, we used unfiltered Common Crawl. We used this classifier to score Common Crawl documents. We kept each document in our dataset iff <> We chose <> in order to take mostly documents the classifier scored highly, but still include some documents that were out of distribution <> was chosen to match the distribution of scores from our classifier on WebText. We found this re-weighting increased quality as measured by loss on a range of out-of-distribution generative text samples. 2.To further improve model quality and prevent overfitting (which becomes increasingly important as model capacity increases), we fuzzily deduplicated documents (i.e. removed documents with high overlap with other documents) within each dataset using Spark’s MinHashLSH implementation with 10 hashes, using the same features as were used for classification above. We also fuzzily removed WebText from Common Crawl. Overall this decreased dataset size by an average of 10%. After filtering for duplicates and quality, we also partially removed text occurring in benchmark datasets, described in Appendix C. B Details of Model Training To train all versions of GPT-3, we use Adam with <>, we clip the global norm of the gradient at 1.0, and we use cosine decay for learning rate down to 10% of its value, over 260 billion tokens (after 260 billion tokens, training continues at 10% of the original learning rate). There is a linear LR warmup over the first 375 million tokens. We also gradually increase the batch size linearly from a small value (32k tokens) to the full value over the first 4-12 billion tokens of training, depending on the model size. Data are sampled without replacement during training (until an epoch boundary is reached) to minimize overfitting. All models use weight decay of 0.1 to provide a small amount of regularization [LH17]. During training we always train on sequences of the fullnctx = 2048token context window, packing multiple documents into a single sequence when documents are shorter than 2048, in order to increase computational efficiency. Sequences with multiple documents are not masked in any special way but instead documents within a sequence are delimited with a special end of text token, giving the language model the information necessary to infer that context separated by the end of text token is unrelated. This allows for efficient training without need for any special sequence-specific masking. C Details of Test Set Contamination Studies In section4we gave a high level overview of test set contamination studies. In this section we provide details on methodology and results. Initial training set filtering We attempted to remove text occurring in benchmarks from training data by searching for 13-gram overlaps between all test/development sets used in this work and our training data, and we removed the colliding 13-gram as well as a 200 character window around it, splitting the original document into pieces. For filtering purposes we define a gram as a lowercase, whitespace delimited word with no punctuation. Pieces less than 200characters long were discarded. Documents split into more than 10 pieces were considered contaminated and removed entirely. Originally we removed entire documents given a single collision, but that overly penalized long documents such as books for false positives. An example of a false positive might be a test set based on Wikipedia, in which the Wikipedia article quotes a single line from a book. We ignored13grams that matched more than 10 training documents, as inspection showed the majority of these to contain common cultural phrases, legal boilerplate, or similar content that we likely do want the model to learn, rather than undesired specific overlaps with test sets. Examples for various frequencies can be found in the GPT-3 release repository. Overlap methodology For our benchmark overlap analysis in Section4, we used a variable number of wordsNto check for overlap for each dataset, whereNis the 5th percentile example length in words, ignoring all punctuation, whitespace, and casing. Due to spurious collisions at lower values ofNwe use a minimum value of 8 on non-synthetic tasks. For performance reasons, we set a maximum value of 13 for all tasks. Values forNand the amount of data marked as dirty are shown in TableC.1. Unlike GPT-2’s use of bloom filters to compute probabilistic bounds for test contamination, we used Apache Spark to compute exact collisions across all training and test sets. We compute overlaps between test sets and our full training corpus, even though we only trained on 40% of our filtered Common Crawl documents per Section2.2. We define a ‘dirty’ example as one with anyN-gram overlap with any training document, and a ‘clean’ example as one with no collision. Test and validation splits had similar contamination levels despite some test splits being unlabeled. Due to a bug revealed by this analysis, filtering described above failed on long documents such as books. Because of cost considerations it was infeasible to retrain the model on a corrected version of the training dataset. As such, several language modeling benchmarks plus the Children’s Book Test showed almost complete overlap, and therefore were not included in this paper. Overlaps are shown in TableC.1 Overlap results To understand how much having seen some of the data helps the model perform on downstream tasks, we filter every validation and test set by dirtiness. Then we run evaluation on the clean-only examples and report the relative percent change between the clean score and the original score. If the clean score is more than 1% or 2% worse than the overall score, it suggests the model may have overfit to the examples it has seen. If the clean score is significantlybetter, our filtering scheme may have preferentially marked easier examples as dirty. This overlap metric tends to show a high rate of false positives for datasets that contain background information (but not answers) drawn from the web (such as SQuAD, which draws from Wikipedia) or examples less than 8 words long, which we ignored in our filtering process (except for wordscrambling tasks). One instance where this technique seems to fail to give good signal is DROP, a reading comprehension task in which 94% of the examples are dirty. The information required to answer the question is in a passage provided to the model, so having seen the passage during training but not the questions and answers does not meaningfully constitute cheating. We confirmed that every matching training document contained only the source passage, and none of the questions and answers in the dataset. The more likely explanation for the decrease in performance is that the 6% of examples that remain after filtering come from a slightly different distribution than the dirty examples. Figure4.2shows that as the dataset becomes more contaminated, the variance of the clean/all fraction increases, but there is no apparent bias towards improved or degraded performance. This suggests that GPT-3 is relatively insensitive to contamination. See Section4for details on the datasets we flagged for further review. <> Table C.1:Overlap statistics for all datasets sorted from dirtiest to cleanest. We consider a dataset example dirty if it has a singleN-gram collision with any document in our training corpus. “Relative Difference Clean vs All” shows the percent change in performance between only the clean examples vs all the examples in the benchmark. “Count” shows the number of examples. “Clean percentage” is the percent of examples that are clean vs total. For “Acc/F1/BLEU” we use the metric specified in “Metric”. These scores come from evaluations with a different seed for the random examples used for in-context learning, and will therefore differ slightly from the scores elsewhere in the paper. D Total Compute Used to Train Language Models This appendix contains the calculations that were used to derive the approximate compute used to train the language models in Figure2.2. As a simplifying assumption, we ignore the attention operation, as it typically uses less than 10% of the total compute for the models we are analyzing. Calculations can be seen in TableD.1and are explained within the table caption. <> Table D.1:Starting from the right hand side and moving left, we begin with the number of training tokens that each model was trained with. Next we note that since T5 uses an encoder-decoder model, only half of the parameters are active for each token during a forward or backwards pass. We then note that each token is involved in a single addition and a single multiply for each active parameter in the forward pass (ignoring attention). Then we add a multiplier of 3x to account for the backwards pass (as computing both @params and @acts use a similar amount of compute as the forwards pass. Combining the previous two numbers, we get the total flops per parameter per token. We multiply this @loss @loss value by the total training tokens and the total parameters to yield the number of total flops used during training. We report both flops and petaflop/s-day (each of which are 2.88e+7 flops). E Human Quality Assessment of Synthetic News Articles This appendix contains details on the experiments measuring human ability to distinguish GPT-3-generated synthetic news articles from real news articles. We first describe the experiments on the200word news articles, and then describe the preliminary investigation of500word news articles generated by GPT-3. Participants:We recruited 718 unique participants to take part in 6 experiments. 97 participants were excluded for failing an internet check question, leaving a total of 621 participants: 343 male, 271 female, and 7 other. Mean participant age was38years old. All participants were recruited through Positly, which maintains a whitelist of high-performing workers from Mechanical Turk. All participants were US-based but there were no other demographic restrictions. Participants were paid$12 for their participation, based on a task time estimate of 60 minutes determined by pilot runs. In order to ensure that the sample of participants for each experiment quiz was unique, participants were not allowed to take part in an experiment more than once. Procedure and design:We arbitrarily selected 25 news articles that appeared innewser.comin early 2020. We used the article titles and subtitles to produce outputs from the 125M, 350M, 760M, 1.3B, 2.7B, 6.7B, 13.0B, and 200B (GPT-3) parameter language models. Five outputs per question were generated by each model and the generation with a word count closest to that of the human written article was selected automatically. This was to minimize the effect that completion length might have on participants’ judgments. The same output procedure for each model with the exception of the removal of the intentionally bad control model, as described in the main text. <> Table E.1:Participant details and article lengths for each experiment to evaluate human detection of200word model generated news articles. Participants were excluded due to internet check fails. <> Figure E.1:Participants spend more time trying to identify whether each news article is machine generated as model size increases. Duration on the control model is indicated with the dashed line. Line of best fit is a linear model on a log scale with 95% confidence intervals. In each experiment, half of the participants were randomly assigned to quiz A and half were randomly assigned to quiz B. Each quiz consisted of 25 articles: half (12-13) were human written and half (12-13) were model generated: the articles with human written completions in quiz A had model generated completions in quiz B and vice versa. The order of quiz question was shuffled for each participant. Participants could leave comments and were asked to indicate if they had seen the articles before. Participants were instructed not to look up the articles or their content during the quiz and at the end of the quiz were asked if they had looked anything up during the quiz. Statistical Tests:To compare means on the different runs, we performed a two-sample t-test for independent groups for each model against the control. This was implemented in Python using thescipy.stats.ttest_indfunction. When plotting a regression line in the graph of average participant accuracy vs model size, we fit a power law of the form ax b . The 95% confidence intervals were estimated from the t-distribution of the sample mean. Duration statistics: In the main text, we discussed the finding that the ability of human participants to distinguish model and human generated news articles decreases as our models become larger. We have also found that the average time spent for a given set of questions increases as the model size increases, as shown in FigureE.1. Lower <> Table E.2:Participant details and article lengths for the experiments investigating human detection of500word model generated news articles. Participants were excluded due to internet check fails. accuracy scores despite increased time investment from participants supports the finding that larger models generate harder-to-distinguish news articles. Preliminary investigation of 500 word articles: We recruited 160 unique US-based participants to take part in 2 experiments through Positly (details are given in TableE.2). We randomly selected 12 Reuters world news articles from late 2019 and created a context for GPT-3 175B that consisted of a single Reuters article not in this set of 12. We then used the article titles and Reuters locations to generate completions from GPT-3 175B and the 160M control model from the previous experiments. These were used to create two 12-question quizzes per model, each consisting of half human written and half model generated articles. Comprehension questions were added and articles were shown to participants in 3 stages at 30 second intervals to encourage closer reading. Participants were paid$12 for this task. Model generation selection methods, exclusion criteria, and statistical tests mirror those of the previous experiments. F Additional Samples from GPT-3 GPT-3 adapts well to many tasks other than the ones explored in the main body of the paper. As an example, in Figure F.1, we show four uncurated samples from a prompt suggesting that the model write a poem, with a given title, in the style of Wallace Stevens. We first experimented with a few prompts, then generated four samples with no additional editing or selection (sampling at temperature1using nucleus sampling [HBFC19] withP= 0:9). Completions were truncated when the model began to write a new title and author heading, or broke into prose commentary. <

> Figure F.1:Four uncurated completions from a context suggesting the model compose a poem in the style of Wallace Stevens with the title ‘Shadows on the Way’. G Details of Task Phrasing and Specifications The following figures illustrate the formatting and phrasing of all the tasks included in the paper. All data comes from the ground truth datasets in this section, and no samples from GPT-3 are included here. <

> Figure G.1:Formatted dataset example for RACE-h. When predicting, we normalize by the unconditional probability of each answer as described in2. <

> Figure G.4:Formatted dataset example for PIQA <

> Figure G.5:Formatted dataset example for COPA <

> Figure G.6:Formatted dataset example for ReCoRD. We consider the context above to be a single ”problem” because this is how the task is presented in the ReCoRD dataset and scored in the ReCoRD evaluation script. <

> Figure G.8:Formatted dataset example for OpenBookQA. When predicting, we normalize by the unconditional probability of each answer as described in2. Context! Making a cake: Several cake pops are shown on a display. A woman and girl are shown making the cake pops in a kitchen. They Correct Answer! bake them, then frost and decorate. Incorrect Answer! taste them as they place them on plates. Incorrect Answer! put the frosting on the cake as they pan it. Incorrect Answer! come out and begin decorating the cake as well. Figure G.9:Formatted dataset example for HellaSwag <

> Figure G.10:Formatted dataset example for ANLI R3 <

> Figure G.11:Formatted dataset example for ARC (Challenge). When predicting, we normalize by the unconditional probability of each answer as described in2. <

> Figure G.12:Formatted dataset example for SAT Analogies <

> Figure G.14:Formatted dataset example for Winogrande. The ‘partial’ evaluation method we use compares the probability of the completion given a correct and incorrect context. <

> Figure G.15:Formatted dataset example for MultiRC. There are three levels within MultiRC: (1) the passage, (2) the questions, and (3) the answers. During evaluation, accuracy is determined at the per-question level, with a question being considered correct if and only if all the answers within the question are labeled correctly. For this reason, we use K to refer to the number ofquestionsshown within the context. <

> Figure G.16:Formatted dataset example for ARC (Easy). When predicting, we normalize by the unconditional probability of each answer as described in 2. <

> Figure G.17:Formatted dataset example for StoryCloze <

> Figure G.18:Formatted dataset example for CoQA <

> Figure G.24:Formatted dataset example for Natural Questions <

> Figure G.26:Formatted dataset example for Symbol Insertion <

> Figure G.30:Formatted dataset example for CB <

> Figure G.32:Formatted dataset example for WiC <

> Figure G.36:Formatted dataset example for De!En. This is the format for one- and few-shot learning, for this and other langauge tasks, the format for zero-shot learning is “Q: What is theflanguagegtranslation offsentencegA: ftranslationg.” <

> Figure G.49:Formatted dataset example for Arithmetic 4D+ <

> Figure G.50:Formatted dataset example for Arithmetic 5D <

> Figure G.51:Formatted dataset example for Arithmetic 5D+ H Results on All Tasks for All Model Sizes <

> Table H.1:Scores for every task, setting and model that we investigate in this paper. <

> Figure H.1:All results for all SuperGLUE tasks. <

> <

> Figure H.2:Results for SAT task. Figure H.3:All results for all Winograd tasks. <

> Figure H.4:All results for all Arithmetic tasks. <

> Figure H.5:All results for all Cloze and Completion tasks. <

> Figure H.6:All results for all Common Sense Reasoning tasks. <

> Figure H.7:All results for all QA tasks. <

> Figure H.8:All results for all Reading Comprehension tasks. <

> Figure H.9:All results for all ANLI rounds. <

> Figure H.10:All results for all Scramble tasks. <

> Figure H.11:All results for all Translation tasks. References [ADG + 16]Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. InAdvances in neural information processing systems, pages 3981–3989, 2016. [AI19]WeChat AI. Tr-mt (ensemble), December 2019. [AJF19]Roee Aharoni, Melvin Johnson, and Orhan Firat. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 2019. [BBDIW20]Su Lin Blodgett, Solon Barocas, Hal Daume III, and Hanna Wallach. Language (technology) is power:´ A critical survey of “bias” in nlp.arXiv preprint arXiv:2005.14050, 2020. [BCFL13]Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. Semantic parsing on freebase from question-answer pairs. InProceedings of the 2013 conference on empirical methods in natural language processing, pages 1533–1544, 2013. [BES10]Stefano Baccianella, Andrea Esuli, and Fabrizio Sebastiani. Sentiwordnet 3.0: an enhanced lexical resource for sentiment analysis and opinion mining. InLrec, volume 10, pages 2200–2204, 2010. [BHT + 20]Yonatan Bisk, Ari Holtzman, Jesse Thomason, Jacob Andreas, Yoshua Bengio, Joyce Chai, Mirella Lapata, Angeliki Lazaridou, Jonathan May, Aleksandr Nisnevich, et al. Experience grounds language. arXiv preprint arXiv:2004.10151, 2020. [BLC13]Yoshua Bengio, Nicholas Leonard, and Aaron C. Courville. Estimating or propagating gradients through´ stochastic neurons for conditional computation.Arxiv, 2013. [BZB + 19]Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. Piqa: Reasoning about physical commonsense in natural language.arXiv preprint arXiv:1911.11641, 2019. [Car97]Rich Caruana. Multitask learning.Machine learning, 28(1), 1997. [CB78]Susan Carey and Elsa Bartlett. Acquiring a single new word.Proceedings of the Stanford Child Language Conference, 1978. [CCE + 18]Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge.ArXiv, abs/1803.05457, 2018. [CGRS19]Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers, 2019. [CHI + 18]Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wen-tau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer. Quac : Question answering in context.Arxiv, 2018. [CLY + 19]Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Learning universal image-text representations.arXiv preprint arXiv:1909.11740, 2019. [Cra17]Kate Crawford. The trouble with bias.NIPS 2017 Keynote, 2017. [DCLT18]Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2018. [DGV + 18]Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser. Universal transformers.Arxiv, 2018. [DHKH14] Nadir Durrani, Barry Haddow, Philipp Koehn, and Kenneth Heafield. Edinburgh’s phrase-based machine translation systems for wmt-14. InProceedings of the Ninth Workshop on Statistical Machine Translation, pages 97–104, 2014. [DL15]Andrew M. Dai and Quoc V. Le. Semi-supervised sequence learning. InAdvances in neural information processing systems, 2015. [DSC + 16]Yan Duan, John Schulman, Xi Chen, Peter L. Bartlett, Ilya Sutskever, and Pieter Abbeel. Rl 2 : Fast reinforcement learning via slow reinforcement learning.ArXiv, abs/1611.02779, 2016. [DWD + 19]Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs.arXiv preprint arXiv:1903.00161, 2019. [DYY + 19]Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G. Carbonell, Quoc V. Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context.Arxiv, 2019. [EOAG18]Sergey Edunov, Myle Ott, Michael Auli, and David Grangier. Understanding back-translation at scale. arXiv preprint arXiv:1808.09381, 2018. [FAL17]Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks.ArXiv, abs/1703.03400, 2017. [Fyo00]Yaroslav Fyodorov. A natural logic inference system, 2000. [GG19]Hila Gonen and Yoav Goldberg. Lipstick on a pig: Debiasing methods cover up systematic gender biases in word embeddings but do not remove them.arXiv preprint arXiv:1903.03862, 2019. [GLT + 20] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. Realm: Retrieval- augmented language model pre-training.arXiv preprint arXiv:2002.08909, 2020. [Gra16]Alex Graves. Adaptive computation time for recurrent neural networks.Arxiv, 2016. [GSL + 18]Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. Annotation artifacts in natural language inference data.arXiv preprint arXiv:1803.02324, 2018. [GSR19]Sebastian Gehrmann, Hendrik Strobelt, and Alexander M. Rush. Gltr: Statistical detection and visualiza- tion of generated text.arXiv preprint arXiv: 1906.04043, 2019. [GWC + 18]Jiatao Gu, Yong Wang, Yun Chen, Kyunghyun Cho, and Victor OK Li. Meta-learning for low-resource neural machine translation.arXiv preprint arXiv:1808.08437, 2018. [HB20]Daniel Hernandez and Tom Brown. Ai and efficiency, May 2020. [HBFC19]Ari Holtzman, Jan Buys, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. CoRR, abs/1904.09751, 2019. [HLW + 20]Dan Hendrycks, Xiaoyuan Liu, Eric Wallace, Adam Dziedzic, Rishabh Krishnan, and Dawn Song. Pretrained transformers improve out of distribution robustness.arXiv preprint arXiv:2004.06100, 2020. [HNA + 17]Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md. Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. [HR18] Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification.arXiv preprint arXiv:1801.06146, 2018. [HVD15]Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015. [HYC01]Sepp Hochreiter, A Steven Younger, and Peter R Conwell. Learning to Learn Using Gradient Descent. InInternational Conference on Artificial Neural Networks, pages 87–94. Springer, 2001. [HZJ + 19]Po-Sen Huang, Huan Zhang, Ray Jiang, Robert Stanforth, Johannes Welbl, Jack Rae, Vishal Maini, Dani Yogatama, and Pushmeet Kohli. Reducing sentiment bias in language models via counterfactual evaluation.arXiv preprint arXiv:1911.03064, 2019. [IBGC + 14]Mohit Iyyer, Jordan Boyd-Graber, Leonardo Claudino, Richard Socher, and Hal Daume III. A neural ´ network for factoid question answering over paragraphs. InEmpirical Methods in Natural Language Processing, 2014. [IDCBE19]Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, and Douglas Eck. Automatic detection of generated text is easiest when humans are fooled.arXiv preprint arXiv:1911.00650, 2019. [JCWZ17]Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension.arXiv preprint arXiv:1705.03551, 2017. [JN20]Zheng Junyuan and Gamma Lab NYC. Numeric transformer - albert, March 2020. [JVS + 16]Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of language modeling.arXiv preprint arXiv:1602.02410, 2016. [JYS + 19]Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. TinyBERT: Distilling BERT for natural language understanding.arXiv preprint arXiv:1909.10351, 2019. [JZC + 19]Ying Ju, Fubang Zhao, Shijie Chen, Bowen Zheng, Xuefeng Yang, and Yunfeng Liu. Technical report on conversational question answering.arXiv preprint arXiv:1909.10772, 2019. [KKS + 20]Daniel Khashabi, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system.arXiv preprint arXiv:2005.00700, 2020. [KMB20]Sarah E. Kreps, Miles McCain, and Miles Brundage. All the news that’s fit to fabricate: Ai-generated text as a tool of media misinformation, 2020. [KMH + 20]Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. [KPR + 19]Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. Natural ques- tions: a benchmark for question answering research.Transactions of the Association of Computational Linguistics, 2019. [KR16]Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation.Arxiv, 2016. [LB02]Edward Loper and Steven Bird. Nltk: The natural language toolkit, 2002. [LC19]Guillaume Lample and Alexis Conneau. Cross-lingual language model pretraining.arXiv preprint arXiv:1901.07291, 2019. [LCG + 19]Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Sori- cut. ALBERT: A lite BERT for self-supervised learning of language representations.arXiv preprint arXiv:1909.11942, 2019. [LCH + 20]Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. Adversarial training for large neural language models.arXiv preprint arXiv:2004.08994, 2020. [LDL19]Zhongyang Li, Xiao Ding, and Ting Liu. Story ending prediction by transferable bert.arXiv preprint arXiv:1905.07504, 2019. [LDM12]Hector Levesque, Ernest Davis, and Leora Morgenstern. The Winograd schema challenge. InThirteenth International Conference on the Principles of Knowledge Representation and Reasoning, 2012. [LGG + 20]Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation.arXiv preprint arXiv:2001.08210, 2020. [LGH + 15]Xiaodong Liu, Jianfeng Gao, Xiaodong He, Li Deng, Kevin Duh, and Ye-Yi Wang. Representation learning using multi-task deep neural networks for semantic classification and information retrieval. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2015. [LH17]Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. [LHCG19a]Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao. Improving multi-task deep neural networks via knowledge distillation for natural language understanding.arXiv preprint arXiv:1904.09482, 2019. [LHCG19b]Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao. Multi-task deep neural networks for natural language understanding.arXiv preprint arXiv:1901.11504, 2019. [Lin20]Tal Linzen. How can we accelerate progress towards human-like linguistic generalization?arXiv preprint arXiv:2005.00955, 2020. [LLG + 19]Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension.arXiv preprint arXiv:1910.13461, 2019. [LM17]Ke Li and Jitendra Malik. Learning to optimize neural nets.arXiv preprint arXiv:1703.00441, 2017. [LOG + 19]Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692, 2019. [LPP + 20]Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktaschel, Sebastian Riedel, and Kiela Douwe.¨ Retrieval-augmented generation for knowledge-intensive nlp tasks.arXiv preprint arXiv:2005.11401, 2020. [LSP + 18]Peter J. Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer. Generating Wikipedia by summarizing long sequences.arXiv preprint arXiv:1801.10198, 2018. [LWS + 20]Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E. Gonzalez. Train large, then compress: Rethinking model size for efficient training and inference of transformers, 2020. [LXL + 17]Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. Race: Large-scale reading comprehension dataset from examinations.arXiv preprint arXiv:1704.04683, 2017. [LYN + 20]Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. Tttttackling winogrande schemas.arXiv preprint arXiv:2003.08380, 2020. [Mac92]David. MacKay. Information-based objective functions for active data selection.Neural Computation, 1992. [MBXS17]Bryan McCann, James Bradbury, Caiming Xiong, and Richard Socher. Learned in translation: Con- textualized word vectors. InAdvances in Neural Information Processing Systems, pages 6294–6305, 2017. [MCCD13]Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space.arXiv preprint arXiv:1301.3781, 2013. [MCH + 16]Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. A corpus and evaluation framework for deeper understanding of commonsense stories.arXiv preprint arXiv:1604.01696, 2016. [MCKS18]Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering.ArXiv, abs/1809.02789, 2018. [MKAT18]Sam McCandlish, Jared Kaplan, Dario Amodei, and OpenAI Dota Team. An empirical model of large-batch training, 2018. [MKM + 94]Mitchell Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger. The penn treebank: annotating predicate argument structure. InProceedings of the workshop on Human Language Technology, pages 114–119. Association for Computational Linguistics, 1994. [MKXS18]Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. The natural language decathlon: Multitask learning as question answering.arXiv preprint arXiv:1806.08730, 2018. [MPL19]R Thomas McCoy, Ellie Pavlick, and Tal Linzen. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference.arXiv preprint arXiv:1902.01007, 2019. [MWZ + 18]Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting, 2018. [NBR20]Moin Nadeem, Anna Bethke, and Siva Reddy. Stereoset: Measuring stereotypical bias in pretrained language models.arXiv preprint arXiv:2004.09456, 2020. [NK19]Timothy Niven and Hung-Yu Kao. Probing neural network comprehension of natural language arguments. arXiv preprint arXiv:1907.07355, 2019. [Nor09]Peter Norvig. Natural language corpus data, 2009. [NvNvdG19]Malvina Nissim, Rik van Noord, and Rob van der Goot. Fair is better than sensational: Man is to doctor as woman is to doctor.arXiv preprint arXiv:1905.09866, 2019. [NWD + 19]Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. Adversarial nli: A new benchmark for natural language understanding.arXiv preprint arXiv:1910.14599, 2019. [oR16]University of Regensburg. Fascha, 2016. [PFB18]Jason Phang, Thibault Fevry, and Samuel R. Bowman. Sentence encoders on STILTs: Supplementary´ training on intermediate labeled-data tasks.arXiv preprint arXiv:1811.01088, 2018. [PKL + 16]Denis Paperno, German Kruszewski, Angeliki Lazaridou, Quan Ngoc Pham, Raffaella Bernardi, Sandro´ Pezzelle, Marco Baroni, Gemma Boleda, and Raquel Fernandez. The lambada dataset: Word prediction´ requiring a broad discourse context.arXiv preprint arXiv:1606.06031, 2016. [PNZtY18]Matthew E. Peters, Mark Neumann, Luke Zettlemoyer, and Wen tau Yih. Dissecting contextual word embeddings: Architecture and representation, 2018. [Pos18]Matt Post. A call for clarity in reporting BLEU scores.arXiv preprint arXiv:1804.08771, 2018. [PSM14]Jeffrey Pennington, Richard Socher, and Christopher Manning. GloVe: Global vectors for word representation. InProceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), 2014. [QIA20]QIANXIN. Sa-net on albert (ensemble), April 2020. [QMZH19]Yusu Qian, Urwa Muaz, Ben Zhang, and Jae Won Hyun. Reducing gender bias in word-level language models with a gender-equalizing loss function.arXiv preprint arXiv:1905.12801, 2019. [RCM19]Siva Reddy, Danqi Chen, and Christopher D Manning. Coqa: A conversational question answering challenge.Transactions of the Association for Computational Linguistics, 7:249–266, 2019. [RCP + 17]Scott Reed, Yutian Chen, Thomas Paine, Aaron van den Oord, SM Eslami, Danilo Rezende, Oriol¨ Vinyals, and Nando de Freitas. Few-shot autoregressive density estimation: Towards learning to learn distributions.arXiv preprint arXiv:1710.10304, 2017. [RJL18]Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don’t know: Unanswerable questions for squad.arXiv preprint arXiv:1806.03822, 2018. [RL16]Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning.ICLR 2017 (oral), 2016. [RLL + 19]Qiu Ran, Yankai Lin, Peng Li, Jie Zhou, and Zhiyuan Liu. NumNet: Machine reading comprehension with numerical reasoning. InProceedings of EMNLP, 2019. [RNLVD18]Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. Gender bias in coreference resolution.arXiv preprint arXiv:1804.09301, 2018. [RNSS18]Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training, 2018. [Ros12]R.S. Ross. Guide for conducting risk assessments.NIST Special Publication, 2012. [RRBS19]Jonathan S. Rosenfeld, Amir Rosenfeld, Yonatan Belinkov, and Nir Shavit. A constructive prediction of the generalization error across scales, 2019. [RRS20]Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model?arXiv preprint arXiv:2002.08910, 2020. [RSR + 19]Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2019. [RWC + 19]Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners, 2019. [SBBC19]Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale, 2019. [SBC + 19]Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, Miles McCain, Alex Newhouse, Jason Blazakis, Kris McGuffie, and Jasmine Wang. Release strategies and the social impacts of language models, 2019. [SCNP19]Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. The woman worked as a babysitter: On biases in language generation.arXiv preprint arXiv:1909.01326, 2019. [SDCW19]Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter.arXiv preprint arXiv:1910.01108, 2019. [SDSE19]Roy Schwartz, Jesse Dodge, Noah A. Smith, and Oren Etzioni. Green AI.CoRR, abs/1907.10597, 2019. [SHB15]Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data.arXiv preprint arXiv:1511.06709, 2015. [SMM + 17]Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017. [SPP + 19]Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using model parallelism, 2019. [SS20]Timo Schick and Hinrich Schutze. Exploiting cloze questions for few-shot text classification and natural¨ language inference.arXiv preprint arXiv:2001.07676, 2020. [STQ + 19]Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MASS: Masked sequence to sequence pre-training for language generation.arXiv preprint arXiv:1905.02450, 2019. [TFR + 17]Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23–30. IEEE, 2017. [TL05]Peter D. Turney and Michael L. Littman. Corpus-based learning of analogies and semantic relations. CoRR, abs/cs/0508103, 2005. [TL18]Trieu H. Trinh and Quoc V. Le. A simple method for commonsense reasoning. arXiv preprint arXiv:1806.02847, 2018. [TLBS03]Peter D. Turney, Michael L. Littman, Jeffrey Bigham, and Victor Shnayder. Combining independent modules to solve multiple-choice synonym and analogy problems.CoRR, cs.CL/0309035, 2003. [Tur20]Project Turing. Microsoft research blog, Feb 2020. [VBL + 16]Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching Networks for One Shot Learning. InAdvances in neural information processing systems, pages 3630–3638, 2016. [VSP + 17]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in neural information processing systems, 2017. [WPN + 19]Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. Superglue: A stickier benchmark for general-purpose language understand- ing systems. InAdvances in Neural Information Processing Systems, pages 3261–3275, 2019. [WXH + 18]Yiren Wang, Yingce Xia, Tianyu He, Fei Tian, Tao Qin, ChengXiang Zhai, and Tie-Yan Liu. Multi-agent dual learning.ICLR 2019, 2018. [XDH + 19]Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised data augmentation for consistency training, 2019. [YdC + 19]Dani Yogatama, Cyprien de Masson d’Autume, Jerome Connor, Tomas Kocisky, Mike Chrzanowski, Lingpeng Kong, Angeliki Lazaridou, Wang Ling, Lei Yu, Chris Dyer, et al. Learning and evaluating general linguistic intelligence.arXiv preprint arXiv:1901.11373, 2019. [YDY + 19]Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. XLNet: Generalized autoregressive pretraining for language understanding.arXiv preprint arXiv:1906.08237, 2019. [ZHB + 19]Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence?arXiv preprint arXiv:1905.07830, 2019. [ZHR + 19]Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. Defending against neural fake news.arXiv preprint arXiv:1905.12616, 2019. [ZSW + 19a] Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences, 2019. [ZSW + 19b]Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Chris- tiano, and Geoffrey Irving. Fine-tuning language models from human preferences.ArXiv, abs/1909.08593, 2019. <> <> <> <> <> <> Learning both Weights and Connections for Efficient Neural Networks Song Han Jeff Pool Stanford University NVIDIA songhan@stanford.edu jpool@nvidia.com John Tran William J. Dally NVIDIA Stanford University johntran@nvidia.com NVIDIA dally@stanford.edu Abstract Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems. Also, conventional networks fix the architecture before training starts; as a result, training cannot improve the architecture. To address these limitations, we describe a method to reduce the storage and computation required by neural networks by an order of magnitude without affecting their accuracy by learning only the important connections. Our method prunes redundant connections using a three-step method. First, we train the network to learn which connections are important. Next, we prune the unimportant connections. Finally, we retrain the network to fine tune the weights of the remaining connections. On the ImageNet dataset, our method reduced the number of parameters of AlexNet by a factor of9, from 61 million to 6.7 million, without incurring accuracy loss. Similar experiments with VGG-16 found that the total number of parameters can be reduced by13, from 138 million to 10.3 million, again with no loss of accuracy. 1 Introduction Neural networks have become ubiquitous in applications ranging from computer vision [1] to speech recognition [2] and natural language processing [3]. We consider convolutional neural networks used for computer vision tasks which have grown over time. In 1998 LeCun et al.designed a CNN model LeNet-5 with less than 1M parameters to classify handwritten digits [4], while in 2012, Krizhevsky et al.[1] won the ImageNet competition with 60M parameters. Deepface classified human faces with 120M parameters [5], and Coateset al.[6] scaled up a network to 10B parameters. While these large neural networks are very powerful, their size consumes considerable storage, memory bandwidth, and computational resources. For embedded mobile applications, these resource demands become prohibitive. Figure 1 shows the energy cost of basic arithmetic and memory operations in a 45nm CMOS process. From this data we see the energy per connection is dominated by memory access and ranges from 5pJ for 32 bit coefficients in on-chip SRAM to 640pJ for 32bit coefficients in off-chip DRAM [7]. Large networks do not fit in on-chip storage and hence require the more costly DRAM accesses. Running a 1 billion connection neural network, for example, at 20Hz would require(20Hz)(1G)(640pJ) = 12:8Wjust for DRAM access - well beyond the power envelope of a typical mobile device. Our goal in pruning networks is to reduce the energy required to run such large networks so they can run in real time on mobile devices. The model size reduction from pruning also facilitates storage and transmission of mobile applications incorporating DNNs. <

> Figure 1: Energy table for 45nm CMOS process [7]. Memory access is 3 orders of magnitude more energy expensive than simple arithmetic. To achieve this goal, we present a method to prune network connections in a manner that preserves the original accuracy. After an initial training phase, we remove all connections whose weight is lower than a threshold. This pruning converts a dense, fully-connected layer to a sparse layer. This first phase learns the topology of the networks — learning which connections are important and removing the unimportant connections. We then retrain the sparse network so the remaining connections can compensate for the connections that have been removed. The phases of pruning and retraining may be repeated iteratively to further reduce network complexity. In effect, this training process learns the network connectivity in addition to the weights - much as in the mammalian brain [8][9], where synapses are created in the first few months of a child’s development, followed by gradual pruning of little-used connections, falling to typical adult values. 2 Related Work Neural networks are typically over-parameterized, and there is significant redundancy for deep learn- ing models [10]. This results in a waste of both computation and memory. There have been various proposals to remove the redundancy: Vanhouckeet al.[11] explored a fixed-point implementation with 8-bit integer (vs 32-bit floating point) activations. Dentonet al. [12] exploited the linear structure of the neural network by finding an appropriate low-rank approximation of the parameters and keeping the accuracy within 1% of the original model. With similar accuracy loss, Gonget al. [13] compressed deep convnets using vector quantization. These approximation and quantization techniques are orthogonal to network pruning, and they can be used together to obtain further gains [14]. There have been other attempts to reduce the number of parameters of neural networks by replacing the fully connected layer with global average pooling. The Network in Network architecture [15] and GoogLenet [16] achieves state-of-the-art results on several benchmarks by adopting this idea. However, transfer learning, i.e. reusing features learned on the ImageNet dataset and applying them to new tasks by only fine-tuning the fully connected layers, is more difficult with this approach. This problem is noted by Szegedyet al.[16] and motivates them to add a linear layer on the top of their networks to enable transfer learning. Network pruning has been used both to reduce network complexity and to reduce over-fitting. An early approach to pruning was biased weight decay [17]. Optimal Brain Damage [18] and Optimal Brain Surgeon [19] prune networks to reduce the number of connections based on the Hessian of the loss function and suggest that such pruning is more accurate than magnitude-based pruning such as weight decay. However, second order derivative needs additional computation. HashedNets [20] is a recent technique to reduce model sizes by using a hash function to randomly group connection weights into hash buckets, so that all connections within the same hash bucket share a single parameter value. This technique may benefit from pruning. As pointed out in Shiet al. [21] and Weinbergeret al.[22], sparsity will minimize hash collision making feature hashing even more effective. HashedNets may be used together with pruning to give even better parameter savings. <

> Figure 3: Synapses and neurons before and after <

> Figure 2: Three-Step Training Pipeline. pruning. 3 Learning Connections in Addition to Weights Our pruning method employs a three-step process, as illustrated in Figure 2, which begins by learning the connectivity via normal network training. Unlike conventional training, however, we are not learning the final values of the weights, but rather we are learning which connections are important. The second step is to prune the low-weight connections. All connections with weights below a threshold are removed from the network — converting a dense network into a sparse network, as shown in Figure 3. The final step retrains the network to learn the final weights for the remaining sparse connections. This step is critical. If the pruned network is used without retraining, accuracy is significantly impacted. 3.1 Regularization Choosing the correct regularization impacts the performance of pruning and retraining. L1 regularization penalizes non-zero parameters resulting in more parameters near zero. This gives better accuracy after pruning, but before retraining. However, the remaining connections are not as good as with L2 regularization, resulting in lower accuracy after retraining. Overall, L2 regularization gives the best pruning results. This is further discussed in experiment section. 3.2 Dropout Ratio Adjustment Dropout [23] is widely used to prevent over-fitting, and this also applies to retraining. During retraining, however, the dropout ratio must be adjusted to account for the change in model capacity. In dropout, each parameter is probabilistically dropped during training, but will come back during inference. In pruning, parameters are dropped forever after pruning and have no chance to come back during both training and inference. As the parameters get sparse, the classifier will select the most informative predictors and thus have much less prediction variance, which reduces over-fitting. As pruning already reduced model capacity, the retraining dropout ratio should be smaller. Quantitatively, letCi be the number of connections in layeri,Cio for the original network,Cir for the network after retraining,Ni be the number of neurons in layer i. Since dropout works on neurons, andCi varies quadratically withNi , according to Equation 1 thus the dropout ratio after pruning the parameters should follow Equation 2, whereDo represent the original dropout rate,Dr represent the dropout rate during retraining. <> (1) <> (2) 3.3 Local Pruning and Parameter Co-adaptation During retraining, it is better to retain the weights from the initial training phase for the connections that survived pruning than it is to re-initialize the pruned layers. CNNs contain fragile co-adapted features [24]: gradient descent is able to find a good solution when the network is initially trained, but not after re-initializing some layers and retraining them. So when we retrain the pruned layers, we should keep the surviving parameters instead of re-initializing them. Table 1: Network pruning can save 9% to 13% parameters with no drop in predictive performance. <

> Retraining the pruned layers starting with retained weights requires less computation because we don’t have to back propagate through the entire network. Also, neural networks are prone to suffer the vanishing gradient problem [25] as the networks get deeper, which makes pruning errors harder to recover for deep networks. To prevent this, we fix the parameters for CONV layers and only retrain the FC layers after pruning the FC layers, and vice versa. 3.4 Iterative Pruning Learning the right connections is an iterative process. Pruning followed by a retraining is one iteration, after many such iterations the minimum number connections could be found. Without loss of accuracy, this method can boost pruning rate from 5% to 9% on AlexNet compared with single-step aggressive pruning. Each iteration is a greedy search in that we find the best connections. We also experimented with probabilistically pruning parameters based on their absolute value, but this gave worse results. 3.5 Pruning Neurons After pruning connections, neurons with zero input connections or zero output connections may be safely pruned. This pruning is furthered by removing all connections to or from a pruned neuron. The retraining phase automatically arrives at the result where dead neurons will have both zero input connections and zero output connections. This occurs due to gradient descent and regularization. A neuron that has zero input connections (or zero output connections) will have no contribution to the final loss, leading the gradient to be zero for its output connection (or input connection), respectively. Only the regularization term will push the weights to zero. Thus, the dead neurons will be automatically removed during retraining. 4 Experiments We implemented network pruning in Caffe [26]. Caffe was modified to add a mask which disregards pruned parameters during network operation for each weight tensor. The pruning threshold is chosen as a quality parameter multiplied by the standard deviation of a layer’s weights. We carried out the experiments on Nvidia TitanX and GTX980 GPUs. We pruned four representative networks: Lenet-300-100 and Lenet-5 on MNIST, together with AlexNet and VGG-16 on ImageNet. The network parameters and accuracy 1 before and after pruning are shown in Table 1. 4.1 LeNet on MNIST We first experimented on MNIST dataset with the LeNet-300-100 and LeNet-5 networks [4]. LeNet- 300-100 is a fully connected network with two hidden layers, with 300 and 100 neurons each, which achieves 1.6% error rate on MNIST. LeNet-5 is a convolutional network that has two convolutional layers and two fully connected layers, which achieves 0.8% error rate on MNIST. After pruning, the network is retrained with1=10of the original network’s original learning rate. Table 1 shows 1 Reference model is from Caffe model zoo, accuracy is measured without data augmentation Table 2: For Lenet-300-100, pruning reduces the number of weights by 12% and computation by 12%. <> Table 3: For Lenet-5, pruning reduces the number of weights by 12% and computation by 6%. <> <

> Figure 4: Visualization of the first FC layer’s sparsity pattern of Lenet-300-100. It has a banded structure repeated 28 times, which correspond to the un-pruned parameters in the center of the images, since the digits are written in the center. pruning saves 12% parameters on these networks. For each layer of the network the table shows (left to right) the original number of weights, the number of floating point operations to compute that layer’s activations, the average percentage of activations that are non-zero, the percentage of non-zero weights after pruning, and the percentage of actually required floating point operations. An interesting byproduct is that network pruning detects visual attention regions. Figure 4 shows the sparsity pattern of the first fully connected layer of LeNet-300-100, the matrix size is 784x300. It has 28 bands, each band’s width 28, corresponding to the 28x28 input pixels. The colored regions of the figure, indicating non-zero parameters, correspond to the center of the image. Because digits are written in the center of the image, these are the important parameters. The graph is sparse on the left and right, corresponding to the less important regions on the top and bottom of the image. After pruning, the neural network finds the center of the image more important, and the connections to the peripheral regions are more heavily pruned. 4.2 AlexNet on ImageNet We further examine the performance of pruning on the ImageNet ILSVRC-2012 dataset, which has 1.2M training examples and 50k validation examples. We use the AlexNet Caffe model as the reference model, which has 61 million parameters across 5 convolutional layers and 3 fully connected layers. The AlexNet Caffe model achieved a top-1 accuracy of 57.2% and a top-5 accuracy of 80.3%. The original AlexNet took 75 hours to train on NVIDIA Titan X GPU. After pruning, the whole network is retrained with1=100of the original network’s initial learning rate. It took 173 hours to retrain the pruned AlexNet. Pruning is not used when iteratively prototyping the model, but rather used for model reduction when the model is ready for deployment. Thus, the retraining time is less a concern. Table 1 shows that AlexNet can be pruned to 1-9% of its original size without impacting accuracy, and the amount of computation can be reduced by 3%. Table 4: For AlexNet, pruning reduces the number of weights by 9% and computation by 3%. <

> Table 5: For VGG-16, pruning reduces the number of weights by 12% and computation by 5%. <> 4.3 VGG-16 on ImageNet With promising results on AlexNet, we also looked at a larger, more recent network, VGG-16 [27], on the same ILSVRC-2012 dataset. VGG-16 has far more convolutional layers but still only three fully-connected layers. Following a similar methodology, we aggressively pruned both convolutional and fully-connected layers to realize a significant reduction in the number of weights, shown in Table 5. We used five iterations of pruning an retraining. The VGG-16 results are, like those for AlexNet, very promising. The network as a whole has been reduced to 7.5% of its original size (13% smaller). In particular, note that the two largest fully-connected layers can each be pruned to less than 4% of their original size. This reduction is critical for real time image processing, where there is little reuse of fully connected layers across images (unlike batch processing during training). 5 Discussion The trade-off curve between accuracy and number of parameters is shown in Figure 5. The more parameters pruned away, the less the accuracy. We experimented with L1 and L2 regularization, with and without retraining, together with iterative pruning to give five trade off lines. Comparing solid and dashed lines, the importance of retraining is clear: without retraining, accuracy begins dropping much sooner with 1-3% of the original connections, rather than with1=10of the original connections. It’s interesting to see that we have the “free lunch” of reducing 2% the connections without losing accuracy even without retraining; while with retraining we are ably to reduce connections by 9%. <

> Figure 5: Trade-off curve for parameter reduction and loss in top-5 accuracy. L1 regularization performs better than L2 at learning the connections without retraining, while L2 regularization performs better than L1 at retraining. Iterative pruning gives the best result. <

> Figure 6: Pruning sensitivity for CONV layer (left) and FC layer (right) of AlexNet. L1 regularization gives better accuracy than L2 directly after pruning (dotted blue and purple lines) since it pushes more parameters closer to zero. However, comparing the yellow and green lines shows that L2 outperforms L1 after retraining, since there is no benefit to further pushing values towards zero. One extension is to use L1 regularization for pruning and then L2 for retraining, but this did not beat simply using L2 for both phases. Parameters from one mode do not adapt well to the other. The biggest gain comes from iterative pruning (solid red line with solid circles). Here we take the pruned and retrained network (solid green line with circles) and prune and retrain it again. The leftmost dot on this curve corresponds to the point on the green line at 80% (5% pruning) pruned to 8%. There’s no accuracy loss at 9%. Not until 10% does the accuracy begin to drop sharply. Two green points achieve slightly better accuracy than the original model. We believe this accuracy improvement is due to pruning finding the right capacity of the network and hence reducing overfitting. Both CONV and FC layers can be pruned, but with different sensitivity. Figure 6 shows the sensitivity of each layer to network pruning. The figure shows how accuracy drops as parameters are pruned on a layer-by-layer basis. The CONV layers (on the left) are more sensitive to pruning than the fully connected layers (on the right). The first convolutional layer, which interacts with the input image directly, is most sensitive to pruning. We suspect this sensitivity is due to the input layer having only 3 channels and thus less redundancy than the other convolutional layers. We used the sensitivity results to find each layer’s threshold: for example, the smallest threshold was applied to the most sensitive layer, which is the first convolutional layer. Storing the pruned layers as sparse matrices has a storage overhead of only 15.6%. Storing relative rather than absolute indices reduces the space taken by the FC layer indices to 5 bits. Similarly, CONV layer indices can be represented with only 8 bits. Table 6: Comparison with other model reduction methods on AlexNet. Data-free pruning [28] saved only 1-5% parameters with much loss of accuracy. Deep Fried Convnets [29] worked on fully connected layers only and reduced the parameters by less than 4%. [30] reduced the parameters by 4% with inferior accuracy. Naively cutting the layer size saves parameters but suffers from 4% loss of accuracy. [12] exploited the linear structure of convnets and compressed each layer individually, where model compression on a single layer incurred 0.9% accuracy penalty with biclustering + SVD. <

> Figure 7: Weight distribution before and after parameter pruning. The right figure has 10% smaller scale. After pruning, the storage requirements of AlexNet and VGGNet are are small enough that all weights can be stored on chip, instead of off-chip DRAM which takes orders of magnitude more energy to access (Table 1). We are targeting our pruning method for fixed-function hardware specialized for sparse DNN, given the limitation of general purpose hardware on sparse computation. Figure 7 shows histograms of weight distribution before (left) and after (right) pruning. The weight is from the first fully connected layer of AlexNet. The two panels have different y-axis scales. The original distribution of weights is centered on zero with tails dropping off quickly. Almost all parameters are between <>. After pruning the large center region is removed. The network parameters adjust themselves during the retraining phase. The result is that the parameters form a bimodal distribution and become more spread across the x-axis, between <>. 6 Conclusion We have presented a method to improve the energy efficiency and storage of neural networks without affecting accuracy by finding the right connections. Our method, motivated in part by how learning works in the mammalian brain, operates by learning which connections are important, pruning the unimportant connections, and then retraining the remaining sparse network. We highlight our experiments on AlexNet and VGGNet on ImageNet, showing that both fully connected layer and convolutional layer can be pruned, reducing the number of connections by 9% to 13% without loss of accuracy. This leads to smaller memory capacity and bandwidth requirements for real-time image processing, making it easier to be deployed on mobile systems. References [1]Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. InAdvances in neural information processing systems, pages 1097–1105, 2012. [2]Alex Graves and Jurgen Schmidhuber. Framewise phoneme classification with bidirectional lstm and other¨ neural network architectures.Neural Networks, 18(5):602–610, 2005. [3]Ronan Collobert, Jason Weston, Leon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. ´ Natural language processing (almost) from scratch.JMLR, 12:2493–2537, 2011. [4] Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998. [5]Yaniv Taigman, Ming Yang, Marc’Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. InCVPR, pages 1701–1708. IEEE, 2014. [6]Adam Coates, Brody Huval, Tao Wang, David Wu, Bryan Catanzaro, and Ng Andrew. Deep learning with cots hpc systems. In30th ICML, pages 1337–1345, 2013. [7]Mark Horowitz. Energy table for 45nm process, Stanford VLSI wiki. [8] JP Rauschecker. Neuronal mechanisms of developmental plasticity in the cat’s visual system.Human neurobiology, 3(2):109–114, 1983. [9]Christopher A Walsh. Peter huttenlocher (1931-2013).Nature, 502(7470):172–172, 2013. [10] Misha Denil, Babak Shakibi, Laurent Dinh, Nando de Freitas, et al. Predicting parameters in deep learning. InAdvances in Neural Information Processing Systems, pages 2148–2156, 2013. [11]Vincent Vanhoucke, Andrew Senior, and Mark Z Mao. Improving the speed of neural networks on cpus. InProc. Deep Learning and Unsupervised Feature Learning NIPS Workshop, 2011. [12]Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. InNIPS, pages 1269–1277, 2014. [13]Yunchao Gong, Liu Liu, Ming Yang, and Lubomir Bourdev. Compressing deep convolutional networks using vector quantization.arXiv preprint arXiv:1412.6115, 2014. [14]Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding.arXiv preprint arXiv:1510.00149, 2015. [15]Min Lin, Qiang Chen, and Shuicheng Yan. Network in network.arXiv preprint arXiv:1312.4400, 2013. [16]Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions.arXiv preprint arXiv:1409.4842, 2014. [17]Stephen Jose Hanson and Lorien Y Pratt. Comparing biases for minimal network construction with´ back-propagation. InAdvances in neural information processing systems, pages 177–185, 1989. [18]Yann Le Cun, John S. Denker, and Sara A. Solla. Optimal brain damage. InAdvances in Neural Information Processing Systems, pages 598–605. Morgan Kaufmann, 1990. [19]Babak Hassibi, David G Stork, et al. Second order derivatives for network pruning: Optimal brain surgeon. Advances in neural information processing systems, pages 164–164, 1993. [20]Wenlin Chen, James T. Wilson, Stephen Tyree, Kilian Q. Weinberger, and Yixin Chen. Compressing neural networks with the hashing trick.arXiv preprint arXiv:1504.04788, 2015. [21]Qinfeng Shi, James Petterson, Gideon Dror, John Langford, Alex Smola, and SVN Vishwanathan. Hash kernels for structured data.The Journal of Machine Learning Research, 10:2615–2637, 2009. [22]Kilian Weinberger, Anirban Dasgupta, John Langford, Alex Smola, and Josh Attenberg. Feature hashing for large scale multitask learning. InICML, pages 1113–1120. ACM, 2009. [23]Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting.JMLR, 15:1929–1958, 2014. [24]Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? InAdvances in Neural Information Processing Systems, pages 3320–3328, 2014. [25]Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient descent is difficult.Neural Networks, IEEE Transactions on, 5(2):157–166, 1994. [26]Yangqing Jia, et al. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014. [27]Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recogni- tion.CoRR, abs/1409.1556, 2014. [28] Suraj Srinivas and R Venkatesh Babu. Data-free parameter pruning for deep neural networks.arXiv preprint arXiv:1507.06149, 2015. [29]Zichao Yang, Marcin Moczulski, Misha Denil, Nando de Freitas, Alex Smola, Le Song, and Ziyu Wang. Deep fried convnets.arXiv preprint arXiv:1412.7149, 2014. [30]Maxwell D Collins and Pushmeet Kohli. Memory bounded deep convolutional networks.arXiv preprint arXiv:1412.1442, 2014. <> <> <> <> <> <> Learning Efficient Convolutional Networks through Network Slimming Abstract The deployment of deep convolutional neural networks (CNNs) in many real world applications is largely hindered by their high computational cost. In this paper, we propose a novel learning scheme for CNNs to simultaneously 1) reduce the model size; 2) decrease the run-time memory footprint; and 3) lower the number of computing operations, without compromising accuracy. This is achieved by en.forcing channel-level sparsity in the network in a simple but effective way. Different from many existing approaches, the proposed method directly applies to modern CNN architectures, introduces minimum overhead to the training process, and requires no special software/hardware accelerators for the resulting models. We call our approach network slimming, which takes wide and large networks as input models, but during training insignificant channels are automatically identified and pruned afterwards, yielding thin and compact models with comparable accuracy. We empirically demonstrate the effectiveness of our approach with several state-of-the-art CNN models, including VGGNet, ResNet and DenseNet, on various image classification datasets. For VGGNet, a multi-pass version of network slimming gives a 20. reduction in model size and a 5. reduction in computing operations. 1. Introduction In recent years, convolutional neural networks (CNNs) have become the dominant approach for a variety of computer vision tasks, e.g., image classification [22], object detection [8], semantic segmentation [26]. Large-scale datasets, high-end modern GPUs and new network architectures allow the development of unprecedented large CNN models. For instance, from AlexNet [22], VGGNet [31] and GoogleNet [34] to ResNets [14], the ImageNet Classification Challenge winner models have evolved from 8 layers to more than 100 layers. This work was done when Zhuang Liu and Zhiqiang Shen were interns at Intel Labs China. Jianguo Li is the corresponding author. However, larger CNNs, although with stronger representation power, are more resource-hungry. For instance, a 152-layer ResNet [14] has more than 60 million parameters and requires more than 20 Giga float-point-operations (FLOPs) when inferencing an image with resolution 224. 224. This is unlikely to be affordable on resource con.strained platforms such as mobile devices, wearables or Internet of Things (IoT) devices. The deployment of CNNs in real world applications are mostly constrained by 1) Model size: CNNs strong representation power comes from their millions of trainable parameters. Those parameters, along with network structure information, need to be stored on disk and loaded into mem.ory during inference time. As an example, storing a typical CNN trained on ImageNet consumes more than 300MB space, which is a big resource burden to embedded devices. 2) Run-time memory: During inference time, the intermediate activations/responses of CNNs could even take more memory space than storing the model parameters, even with batch size 1. This is not a problem for high-end GPUs, but unaffordable for many applications with low computational power. 3) Number of computing operations: The convolution operations are computationally intensive on high resolution images. A large CNN may take several minutes to process one single image on a mobile device, making it un.realistic to be adopted for real applications. Many works have been proposed to compress large CNNs or directly learn more Efficient CNN models for fast inference. These include low-rank approximation [7], network quantization [3, 12] and binarization [28, 6], weight pruning [12], dynamic inference [16], etc. However, most of these methods can only address one or two challenges mentioned above. Moreover, some of the techniques require specially designed software/hardware accelerators for execution speedup [28, 6, 12]. Another direction to reduce the resource consumption of large CNNs is to sparsify the network. Sparsity can be im.posed on different level of structures [2, 37, 35, 29, 25], which yields considerable model-size compression and inference speedup. However, these approaches generally re. <

> Figure 1: We associate a scaling factor (reused from a batch normalization layer) with each channel in convolutional layers. Sparsity regularization is imposed on these scaling factors during training to automatically identify unimportant channels. The channels with small scaling factor values (in orange color) will be pruned (left side). After pruning, we obtain compact models (right side), which are then fine-tuned to achieve comparable (or even higher) accuracy as normally trained full network. quire special software/hardware accelerators to harvest the gain in memory or time savings, though it is easier than non-structured sparse weight matrix as in [12]. In this paper, we propose network slimming, a simple yet effective network training scheme, which addresses all the aforementioned challenges when deploying large CNNs under limited resources. Our approach imposes L1 regularization on the scaling factors in batch normalization (BN) layers, thus it is easy to implement without introducing any change to existing CNN architectures. Pushing the val.ues of BN scaling factors towards zero with L1 regularization enables us to identify insignificant channels (or neurons), as each scaling factor corresponds to a specific convolutional channel (or a neuron in a fully-connected layer). This facilitates the channel-level pruning at the followed step. The additional regularization term rarely hurt the performance. In fact, in some cases it leads to higher generalization accuracy. Pruning unimportant channels may sometimes temporarily degrade the performance, but this effect can be compensated by the followed fine-tuning of the pruned network. After pruning, the resulting narrower network is much more compact in terms of model size, run.time memory, and computing operations compared to the initial wide network. The above process can be repeated for several times, yielding a multi-pass network slimming scheme which leads to even more compact network. Experiments on several benchmark datasets and different network architectures show that we can obtain CNN models with up to 20x mode-size compression and 5x reduction in computing operations of the original ones, while achieving the same or even higher accuracy. Moreover, our method achieves model compression and inference speedup with conventional hardware and deep learning software packages, since the resulting narrower model is free of any sparse storing format or computing operations. 2. Related Work In this section, we discuss related work from five aspects. Low-rank Decomposition approximates weight matrix in neural networks with low-rank matrix using techniques like Singular Value Decomposition (SVD) [7]. This method works especially well on fully-connected layers, yield.ing 3x model-size compression however without notable speed acceleration, since computing operations in CNN mainly come from convolutional layers. Weight Quantization. HashNet [3] proposes to quantize the network weights. Before training, network weights are hashed to different groups and within each group weight the value is shared. In this way only the shared weights and hash indices need to be stored, thus a large amount of stor.age space could be saved. [12] uses a improved quantization technique in a deep compression pipeline and achieves 35x to 49x compression rates on AlexNet and VGGNet. How.ever, these techniques can neither save run-time memory nor inference time, since during inference shared weights need to be restored to their original positions. [28, 6] quantize real-valued weights into binary/ternary weights (weight values restricted to {-1, 1} or {-1, 0, 1}). This yields a large amount of model-size saving, and significant speedup could also be obtained given bitwise operation libraries. However, this aggressive low-bit approximation method usually comes with a moderate accuracy loss. Weight Pruning / Sparsifying. [12] proposes to prune the unimportant connections with small weights in trained neu.ral networks. The resulting network's weights are mostly zeros thus the storage space can be reduced by storing the model in a sparse format. However, these methods can only achieve speedup with dedicated sparse matrix operation libraries and/or hardware. The run-time memory saving is also very limited since most memory space is consumed by the activation maps (still dense) instead of the weights. In [12], there is no guidance for sparsity during training. [32] overcomes this limitation by explicitly imposing sparse constraint over each weight with additional gate variables, and achieve high compression rates by pruning connections with zero gate values. This method achieves better compression rate than [12], but suffers from the same drawback. Structured Pruning / Sparsifying. Recently, [23] pro.poses to prune channels with small incoming weights in trained CNNs, and then fine-tune the network to regain accuracy. [2] introduces sparsity by random deactivating input-output channel-wise connections in convolutional layers before training, which also yields smaller networks with moderate accuracy loss. Compared with these works, we explicitly impose channel-wise sparsity in the optimization objective during training, leading to smoother channel pruning process and little accuracy loss. [37] imposes neuron-level sparsity during training thus some neurons could be pruned to obtain compact networks. [35] proposes a Structured Sparsity Learning (SSL) method to sparsify different level of structures (e.g. filters, channels or layers) in CNNs. Both methods utilize group sparsity regularization during training to obtain structured sparsity. Instead of resorting to group sparsity on convolutional weights, our approach imposes simple L1 sparsity on channel-wise scaling factors, thus the optimization objective is much simpler. Since these methods prune or sparsify part of the network structures (e.g., neurons, channels) instead of individual weights, they usually require less specialized libraries (e.g. for sparse computing operation) to achieve inference speedup and run-time memory saving. Our network slimming also falls into this category, with absolutely no special libraries needed to obtain the benefits. Neural Architecture Learning. While state-of-the-art CNNs are typically designed by experts [22, 31, 14], there are also some explorations on automatically learning network architectures. [20] introduces sub-modular/super.modular optimization for network architecture search with a given resource budget. Some recent works [38, 1] propose to learn neural architecture automatically with reinforcement learning. The searching space of these methods are extremely large, thus one needs to train hundreds of models to distinguish good from bad ones. Network slimming can also be treated as an approach for architecture learning, despite the choices are limited to the width of each layer. However, in contrast to the aforementioned methods, network slimming learns network architecture through only a single training process, which is in line with our goal of efficiency. 3. Network slimming We aim to provide a simple scheme to achieve channel-level sparsity in deep CNNs. In this section, we first discuss the advantages and challenges of channel-level sparsity, and introduce how we leverage the scaling layers in batch normalization to effectively identify and prune unimportant channels in the network. Advantages of Channel-level Sparsity. As discussed in prior works [35, 23, 11], sparsity can be realized at differ.ent levels, e.g., weight-level, kernel-level, channel-level or layer-level. Fine-grained level (e.g., weight-level) sparsity gives the highest flexibility and generality leads to higher compression rate, but it usually requires special software or hardware accelerators to do fast inference on the sparsified model [11]. On the contrary, the coarsest layer-level sparsity does not require special packages to harvest the inference speedup, while it is less flexible as some whole layers need to be pruned. In fact, removing layers is only effective when the depth is sufficiently large, e.g., more than 50 layers [35, 18]. In comparison, channel-level sparsity provides a nice tradeoff between flexibility and ease of implementation. It can be applied to any typical CNNs or fully-connected networks (treat each neuron as a channel), and the resulting network is essentially a "thinned" version of the unpruned network, which can be Efficiently inferenced on conventional CNN platforms. Challenges. Achieving channel-level sparsity requires pruning all the incoming and outgoing connections associated with a channel. This renders the method of directly pruning weights on a pre-trained model ineffective, as it is unlikely that all the weights at the input or output end of a channel happen to have near zero values. As reported in [23], pruning channels on pre-trained ResNets can only lead to a reduction of 10% in the number of parameters without suffering from accuracy loss. [35] addresses this problem by enforcing sparsity regularization into the training objective. specifically, they adopt group LASSO to push all the filter weights corresponds to the same channel towards zero simultaneously during training. However, this approach re.quires computing the gradients of the additional regularization term with respect to all the filter weights, which is non.trivial. We introduce a simple idea to address the above challenges, and the details are presented below. Scaling Factors and Sparsity-induced Penalty. Our idea is introducing a scaling factor . for each channel, which is multiplied to the output of that channel. Then we jointly train the network weights and these scaling factors, with sparsity regularization imposed on the latter. Finally we prune those channels with small factors, and fine-tune the pruned network. specifically, the training objective of our approach is given by <> (1) where <> denote the train input and target, W denotes the trainable weights, the first sum-term corresponds to the normal training loss of a CNN, <> is a sparsity-induced penalty on the scaling factors, and <> balances the two terms. In our experiment, we choose <>, which is known as <

> Figure 2: Flow-chart of network slimming procedure. The dotted-line is for the multi-pass/iterative scheme. L1-norm and widely used to achieve sparsity. Subgradient descent is adopted as the optimization method for the non-smooth L1 penalty term. An alternative option is to replace the L1 penalty with the smooth-L1 penalty [30] to avoid using sub-gradient at non-smooth point. As pruning a channel essentially corresponds to removing all the incoming and outgoing connections of that chan.nel, we can directly obtain a narrow network (see Figure 1) without resorting to any special sparse computation packages. The scaling factors act as the agents for channel se.lection. As they are jointly optimized with the network weights, the network can automatically identity insignificant channels, which can be safely removed without greatly affecting the generalization performance. Leveraging the Scaling Factors in BN Layers. Batch normalization [19] has been adopted by most modern CNNs as a standard approach to achieve fast convergence and bet.ter generalization performance. The way BN normalizes the activations motivates us to design a simple and efficient method to incorporates the channel-wise scaling fac.tors. Particularly, BN layer normalizes the internal activations using mini-batch statistics. Let z_in and z_out be the input and output of a BN layer, B denotes the current mini-batch, BN layer performs the following transformation: <> where <> and <> are the mean and standard deviation val.ues of input activations over <> and <> are trainable affine transformation parameters (scale and shift) which provides the possibility of linearly transforming normalized activations back to any scales. It is common practice to insert a BN layer after a convolutional layer, with channel-wise scaling/shifting parameters. Therefore, we can directly leverage the . parameters in BN layers as the scaling factors we need for network slimming. It has the great advantage of introducing no overhead to the network. In fact, this is perhaps also the most effective way we can learn meaningful scaling factors for chan.nel pruning. 1), if we add scaling layers to a CNN without BN layer, the value of the scaling factors are not meaning.ful for evaluating the importance of a channel, because both convolution layers and scaling layers are linear transformations. One can obtain the same results by decreasing the scaling factor values while amplifying the weights in the convolution layers. 2), if we insert a scaling layer before a BN layer, the scaling effect of the scaling layer will be completely canceled by the normalization process in BN. 3), if we insert scaling layer after BN layer, there are two consecutive scaling factors for each channel. Channel Pruning and Fine-tuning. After training under channel-level sparsity-induced regularization, we obtain a model in which many scaling factors are near zero (see Figure 1). Then we can prune channels with near-zero scaling factors, by removing all their incoming and outgoing connections and corresponding weights. We prune channels with a global threshold across all layers, which is defined as a certain percentile of all the scaling factor values. For instance, we prune 70% channels with lower scaling factors by choosing the percentile threshold as 70%. By doing so, we obtain a more compact network with less parameters and run-time memory, as well as less computing operations. Pruning may temporarily lead to some accuracy loss, when the pruning ratio is high. But this can be largely compensated by the followed fine-tuning process on the pruned network. In our experiments, the fine-tuned narrow network can even achieve higher accuracy than the original unpruned network in many cases. Multi-pass Scheme. We can also extend the proposed method from single-pass learning scheme (training with sparsity regularization, pruning, and fine-tuning) to a multi-pass scheme. specifically, a network slimming procedure results in a narrow network, on which we could again apply the whole training procedure to learn an even more compact model. This is illustrated by the dotted-line in Figure 2. Experimental results show that this multi-pass scheme can lead to even better results in terms of compression rate. Handling Cross Layer Connections and Pre-activation Structure. The network slimming process introduced above can be directly applied to most plain CNN architectures such as AlexNet [22] and VGGNet [31]. While some adaptations are required when it is applied to modern networks with cross layer connections and the pre-activation design such as ResNet [15] and DenseNet [17]. For these networks, the output of a layer may be treated as the input of multiple subsequent layers, in which a BN layer is placed before the convolutional layer. In this case, the sparsity is achieved at the incoming end of a layer, i.e., the layer selectively uses a subset of channels it received. To harvest the parameter and computation savings at test time, we need to place a channel selection layer to mask out insignificant channels we have identified. 4. Experiments We empirically demonstrate the effectiveness of network slimming on several benchmark datasets. We implement <

> Table 1: Results on CIFAR and SVHN datasets. "Baseline" denotes normal training without sparsity regularization. In column-1, 60% pruned denotes the fine-tuned model with 60% channels pruned from the model trained with sparsity, etc. The pruned ratio of parameters and FLOPs are also shown in column-4&6. Pruning a moderate amount (40%) of channels can mostly lower the test errors. The accuracy could typically be maintained with  60% channels pruned. our method based on the publicly available Torch [5] implementation for ResNets by [10]. The code is available at https://github.com/liuzhuang13/slimming. 4.1. Datasets CIFAR. The two CIFAR datasets [21] consist of natural im. ages with resolution 32.32. CIFAR-10 is drawn from 10 and CIFAR-100 from 100 classes. The train and test sets contain 50,000 and 10,000 images respectively. On CIFAR.10, a validation set of 5,000 images is split from the training set for the search of . (in Equation 1) on each model. We report the final test errors after training or fine-tuning on all training images. A standard data augmentation scheme (shifting/mirroring) [14, 18, 24] is adopted. The input data is normalized using channel means and standard deviations. We also compare our method with [23] on CIFAR datasets. SVHN. The Street View House Number (SVHN) dataset [27] consists of 32x32 colored digit images. Following common practice [9, 18, 24] we use all the 604,388 training images, from which we split a validation set of 6,000 im.ages for model selection during training. The test set con.tains 26,032 images. During training, we select the model with the lowest validation error as the model to be pruned (or the baseline model). We also report the test errors of the models with lowest validation errors during fine-tuning. ImageNet. The ImageNet dataset contains 1.2 million training images and 50,000 validation images of 1000 classes. We adopt the data augmentation scheme as in [10]. We report the single-center-crop validation error of the final model. MNIST. MNIST is a handwritten digit dataset containing 60,000 training images and 10,000 test images. To test the effectiveness of our method on a fully-connected network (treating each neuron as a channel with 1.1 spatial size), we compare our method with [35] on this dataset. 4.2. Network Models On CIFAR and SVHN dataset, we evaluate our method on three popular network architectures: VGGNet[31], ResNet [14] and DenseNet [17]. The VGGNet is originally designed for ImageNet classification. For our experiment a variation of the original VGGNet for CIFAR dataset is taken from [36]. For ResNet, a 164-layer pre-activation ResNet with bottleneck structure (ResNet-164) [15] is used. For DenseNet, we use a 40-layer DenseNet with growth rate 12 (DenseNet-40). On ImageNet dataset, we adopt the 11-layer (8-conv + 3 FC) VGG-A network [31] model with batch normalization from [4]. We remove the dropout layers since we use relatively heavy data augmentation. To prune the neurons in fully-connected layers, we treat them as convolutional channels with 1.1 spatial size. On MNIST dataset, we evaluate our method on the same 3-layer fully-connected network as in [35]. 4.3. Training, Pruning and Fine-tuning Normal Training. We train all the networks normally from scratch as baselines. All the networks are trained using SGD. On CIFAR and SVHN datasets we train using mini-batch size 64 for 160 and 20 epochs, respectively. The initial learning rate is set to 0.1, and is divided by 10 at 50% and 75% of the total number of training epochs. On Im.ageNet and MNIST datasets, we train our models for 60 and 30 epochs respectively, with a batch size of 256, and an initial learning rate of 0.1 which is divided by 10 after 1/3 and 2/3 fraction of training epochs. We use a weight de.cay of 10.4 and a Nesterov momentum [33] of 0.9 without dampening. The weight initialization introduced by [13] is adopted. Our optimization settings closely follow the orig.inal implementation at [10]. In all our experiments, we initialize all channel scaling factors to be 0.5, since this gives higher accuracy for the baseline models compared with de.fault setting (all initialized to be 1) from [10]. Training with Sparsity. For CIFAR and SVHN datasets, when training with channel sparse regularization, the hyper.parameteer ., which controls the tradeoff between empirical loss and sparsity, is determined by a grid search over 10.3, 10.4, 10.5 on CIFAR-10 validation set. For VG-GNet we choose 10.4 and for ResNet and DenseNet 10.5. For VGG-A on ImageNet, we set 10.5 . All other settings are kept the same as in normal training. Pruning. When we prune the channels of models trained with sparsity, a pruning threshold on the scaling factors needs to be determined. Unlike in [23] where different lay.ers are pruned by different ratios, we use a global pruning threshold for simplicity. The pruning threshold is deter.mined by a percentile among all scaling factors , e.g., 40% or 60% channels are pruned. The pruning process is implemented by building a new narrower model and copying the corresponding weights from the model trained with sparsity. Fine-tuning. After the pruning we obtain a narrower and more compact model, which is then fine-tuned. On CIFAR, SVHN and MNIST datasets, the fine-tuning uses the same optimization setting as in training. For ImageNet dataset, due to time constraint, we fine-tune the pruned VGG-A with a learning rate of 10.3 for only 5 epochs. <

> Figure 3: Comparison of pruned models with lower test errors on CIFAR-10 than the original models. The blue and green bars are parameter and FLOP ratios between pruned and original models. 4.4. Results CIFAR and SVHN The results on CIFAR and SVHN are shown in Table 1. We mark all lowest test errors of a model in boldface. Parameter and FLOP reductions. The purpose of network slimming is to reduce the amount of computing re.sources needed. The last row of each model has  60% channels pruned while still maintaining similar accuracy to the baseline. The parameter saving can be up to 10.. The FLOP reductions are typically around 50%. To highlight network slimming's efficiency, we plot the resource savings in Figure 3. It can be observed that VGGNet has a large amount of redundant parameters that can be pruned. On ResNet-164 the parameter and FLOP savings are relatively insignificant, we conjecture this is due to its "bottleneck" structure has already functioned as selecting channels. Also, on CIFAR-100 the reduction rate is typically slightly lower than CIFAR-10 and SVHN, which is possibly due to the fact that CIFAR-100 contains more classes. Regularization Effect. From Table 1, we can observe that, on ResNet and DenseNet, typically when 40% channels are pruned, the fine-tuned network can achieve a lower test er.ror than the original models. For example, DenseNet-40 with 40% channels pruned achieve a test error of 5.19% on CIFAR-10, which is almost 1% lower than the original model. We hypothesize this is due to the regularization effect of L1 sparsity on channels, which naturally provides feature selection in intermediate layers of a network. We will analyze this effect in the next section. <

> Table 3: Results on MNIST. ImageNet. The results for ImageNet dataset are summarized in Table 2. When 50% channels are pruned, the parameter saving is more than 5%, while the FLOP saving is only 30.4%. This is due to the fact that only 378 (out of 2752) channels from all the computation-intensive convolutional layers are pruned, while 5094 neurons (out of 8192) from the parameter-intensive fully-connected layers are pruned. It is worth noting that our method can achieve the savings with no accuracy loss on the 1000-class Im.ageNet dataset, where other methods for Efficient CNNs [2, 23, 35, 28] mostly report accuracy loss. MNIST. On MNIST dataset, we compare our method with the Structured Sparsity Learning (SSL) method [35] in Ta. ble 3. Despite our method is mainly designed to prune channels in convolutional layers, it also works well in pruning neurons in fully-connected layers. In this experiment, we observe that pruning with a global threshold sometimes completely removes a layer, thus we prune 80% of the neurons in each of the two intermediate layers. Our method slightly outperforms [35], in that a slightly lower test error is achieved while pruning more parameters. We provide some additional experimental results in the supplementary materials, including (1) detailed structure of a compact VGGNet on CIFAR-10; (2) wall-clock time and run-time memory savings in practice. (3) comparison with a previous channel pruning method [23]; 4.5. Results for Multi-pass Scheme We employ the multi-pass scheme on CIFAR datasets using VGGNet. Since there are no skip-connections, pruning away a whole layer will completely destroy the models. Thus, besides setting the percentile threshold as 50%, we also put a constraint that at each layer, at most 50% of channels can be pruned. The test errors of models in each iteration are shown in Table 4. As the pruning process goes, we obtain more and <> Table 4: Results for multi-pass scheme on CIFAR-10 and CIFAR.100 datasets, using VGGNet. The baseline model has test errors of 6.34% and 26.74%. Trained and Fine-tuned columns denote the test errors (%) of the model trained with sparsity, and the fine-tuned model after channel pruning, respectively. The parameter and FLOP pruned ratios correspond to the fine-tuned model in that row and the trained model in the next row. more compact models. On CIFAR-10, the trained model achieves the lowest test error in iteration 5. This model achieves 20. parameter reduction and 5. FLOP reduction, while still achieving lower test error. On CIFAR-100, after iteration 3, the test error begins to increase. This is pos.sibly due to that it contains more classes than CIFAR-10, so pruning channels too aggressively will inevitably hurt the performance. However, we can still prune near 90% parameters and near 70% FLOPs without notable accuracy loss. 5. Analysis There are two crucial hyper-parameters in network slimming, the pruned percentage t and the coEfficient of the sparsity regularization term . (see Equation 1). In this section, we analyze their effects in more detail. Effect of Pruned Percentage. Once we obtain a model trained with sparsity regularization, we need to decide what percentage of channels to prune from the model. If we prune too few channels, the resource saving can be very limited. However, it could be destructive to the model if we prune too many channels, and it may not be possible to recover the accuracy by fine-tuning. We train a DenseNet.40 model with 10.5 on CIFAR-10 to show the effect of pruning a varying percentage of channels. The results are summarized in Figure 5. From Figure 5, it can be concluded that the classification performance of the pruned or fine-tuned models degrade only when the pruning ratio surpasses a threshold. The fine. <

> Figure 4: Distributions of scaling factors in a trained VGGNet under various degree of sparsity regularization (controlled by the parameter). With the increase of , scaling factors become sparser. <

> Figure 5: The effect of pruning varying percentages of channels, from DenseNet-40 trained on CIFAR-10 with =10.5 . tuning process can typically compensate the possible accuracy loss caused by pruning. Only when the threshold goes beyond 80%, the test error of fine-tuned model falls behind the baseline model. Notably, when trained with sparsity, even without fine-tuning, the model performs better than the original model. This is possibly due the the regularization effect of L1 sparsity on channel scaling factors. Channel Sparsity Regularization. The purpose of the L1 sparsity term is to force many of the scaling factors to be near zero. The parameter <> in Equation 1 controls its significance compared with the normal training loss. In Figure 4 we plot the distributions of scaling factors in the whole network with different . values. For this experiment we use a VGGNet trained on CIFAR-10 dataset. It can be observed that with the increase of ., the scaling factors are more and more concentrated near zero. When 0, i.e., there's no sparsity regularization, the distribution is relatively flat. When 10.4 , almost all scaling factors fall into a small region near zero. This process can be seen as a feature selection happening in intermediate layers of deep networks, where only channels with non-negligible scaling factors are chosen. We further visualize this process by a heatmap. Figure 6 shows the magnitude of scaling factors from one layer in VGGNet, along the training process. Each channel starts with equal weights; as the training <

> Figure 6: Visulization of channel scaling factorsfi change in scale along the training process, taken from the 11th conv-layer in VG-GNet trained on CIFAR-10. Brighter color corresponds to larger value. The bright lines indicate the selected channels, the dark lines indicate channels that can be pruned. progresses, some channels scaling factors become larger (brighter) while others become smaller (darker). 6. Conclusion We proposed the network slimming technique to learn more compact CNNs. It directly imposes sparsity-induced regularization on the scaling factors in batch normalization layers, and unimportant channels can thus be automatically identified during training and then pruned. On multiple datasets, we have shown that the proposed method is able to significantly decrease the computational cost (up to 20.) of state-of-the-art networks, with no accuracy loss. More importantly, the proposed method simultaneously reduces the model size, run-time memory, computing operations while introducing minimum overhead to the training process, and the resulting models require no special libraries/hardware for Efficient inference. Acknowledgements. Gao Huang is supported by the International Postdoctoral Exchange Fellowship Program of China Postdoctoral Council (No.20150015). Changshui Zhang is supported by NSFC and DFG joint project NSFC 61621136008/DFG TRR-169. References [1] B. Baker, O. Gupta, N. Naik, and R. Raskar. Designing neu.ral network architectures using reinforcement learning. In ICLR, 2017. [2] S. Changpinyo, M. Sandler, and A. Zhmoginov. The power of sparsity in convolutional neural networks. arXiv preprint arXiv:1702.06257, 2017. [3] W. Chen, J. T. Wilson, S. Tyree, K. Q. Weinberger, and Y. Chen. Compressing neural networks with the hashing trick. In ICML, 2015. [4] S. Chintala. Training an object classifier in torch-7 on multiple gpus over imagenet. https://github.com/soumith/imagenet-multiGPU.torch. [5] R. Collobert, K. Kavukcuoglu, and C. Farabet. Torch7: A matlab-like environment for machine learning. In BigLearn, NIPS Workshop, number EPFL-CONF-192376, 2011. [6] M. Courbariaux and Y. Bengio. Binarynet: Training deep neural networks with weights and activations constrained to+ 1 or-1. arXiv preprint arXiv:1602.02830, 2016. [7] E. L. Denton, W. Zaremba, J. Bruna, Y. LeCun, and R. Fer.gus. Exploiting linear structure within convolutional networks for Efficient evaluation. In NIPS, 2014. [8] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich fea.ture hierarchies for accurate object detection and semantic segmentation. In CVPR, pages 580fi587, 2014. [9] I. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and Y. Bengio. Maxout networks. In ICML, 2013. [10] S. Gross and M. Wilber. Training and investigating residual nets. https://github.com/szagoruyko/cifar-torch. [11] S. Han, H. Mao, and W. J. Dally. Deep compression: Com.pressing deep neural network with pruning, trained quanti.zation and huffman coding. In ICLR, 2016. [12] S. Han, J. Pool, J. Tran, and W. Dally. Learning both weights and connections for Efficient neural network. In NIPS, pages 1135fi1143, 2015. [13] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. [14] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016. [15] K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual networks. In ECCV, pages 630fi645. Springer, 2016. [16] G. Huang, D. Chen, T. Li, F. Wu, L. van der Maaten, and K. Q. Weinberger. Multi-scale dense convolutional networks for Efficient prediction. arXiv preprint arXiv:1703.09844, 2017. [17] G. Huang, Z. Liu, K. Q. Weinberger, and L. van der Maaten. Densely connected convolutional networks. In CVPR, 2017. [18] G. Huang, Y. Sun, Z. Liu, D. Sedra, and K. Q. Weinberger. Deep networks with stochastic depth. In ECCV, 2016. [19] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. [20] J. Jin, Z. Yan, K. Fu, N. Jiang, and C. Zhang. Neural network architecture optimization through submodularity and super-modularity. arXiv preprint arXiv:1609.00074, 2016. [21] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. In Tech Report, 2009. [22] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, pages 1097fi1105, 2012. [23] H. Li, A. Kadav, I. Durdanovic, H. Samet, and H. P. Graf. Pruning filters for Efficient convnets. arXiv preprint arXiv:1608.08710, 2016. [24] M. Lin, Q. Chen, and S. Yan. Network in network. In ICLR, 2014. [25] B. Liu, M. Wang, H. Foroosh, M. Tappen, and M. Pensky. Sparse convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni.tion, pages 806fi814, 2015. [26] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431fi 3440, 2015. [27] Y. Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y. Ng. Reading digits in natural images with unsupervised fea.ture learning, 2011. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. [28] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnor.net: Imagenet classification using binary convolutional neu.ral networks. In ECCV, 2016. [29] S. Scardapane, D. Comminiello, A. Hussain, and A. Uncini. Group sparse regularization for deep neural networks. arXiv preprint arXiv:1607.00485, 2016. [30] M. Schmidt, G. Fung, and R. Rosales. Fast optimization methods for l1 regularization: A comparative study and two new approaches. In ECML, pages 286fi297, 2007. [31] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. [32] S. Srinivas, A. Subramanya, and R. V. Babu. Training sparse neural networks. CoRR, abs/1611.06694, 2016. [33] I. Sutskever, J. Martens, G. Dahl, and G. Hinton. On the importance of initialization and momentum in deep learning. In ICML, 2013. [34] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, et al. Going deeper with convolutions. In CVPR, pages 1fi9, 2015. [35] W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li. Learning structured sparsity in deep neural networks. In NIPS, 2016. [36] S. Zagoruyko. 92.5% on cifar-10 in torch. https://github.com/szagoruyko/cifar.torch. [37] H. Zhou, J. M. Alvarez, and F. Porikli. Less is more: Towards compact cnns. In ECCV, 2016. [38] B. Zoph and Q. V. Le. Neural architecture search with rein.forcement learning. In ICLR, 2017. <> <> <> <> <> <> Learning Structured Sparsity in Deep Neural Networks Wei Wen Chunpeng Wu Yandan Wang University of Pittsburgh University of Pittsburgh University of Pittsburgh wew57@pitt.edu chw127@pitt.edu yaw46@pitt.edu Yiran Chen Hai Li University of Pittsburgh University of Pittsburgh yic52@pitt.edu hal66@pitt.edu Abstract High demand for computation resources severely hinders deployment of large-scale Deep Neural Networks (DNN) in resource constrained devices. In this work, we propose aStructured Sparsity Learning(SSL) method to regularize the structures (i.e., filters, channels, filter shapes, and layer depth) of DNNs. SSL can: (1) learn a compact structure from a bigger DNN to reduce computation cost; (2) obtain a hardware-friendly structured sparsity of DNN to efficiently accelerate the DNN’s evaluation. Experimental results show that SSL achieves on average 5.1%and 3.1%speedups of convolutional layer computation of AlexNet against CPU and GPU, respectively, with off-the-shelf libraries. These speedups are about twice speedups of non-structured sparsity; (3) regularize the DNN structure to improve classification accuracy. The results show that for CIFAR-10, regularization on layer depth can reduce 20 layers of a Deep Residual Network ( ResNet ) to 18 layers while improve the accuracy from 91.25% to 92.60%, which is still slightly higher than that of original ResNet with 32 layers. For AlexNet , structure regularization by SSL also reduces the error by%1%. Our source code can be found athttps://github.com/wenwei202/caffe/tree/scnn 1 Introduction Deep neural networks (DNN), especially deep convolutional neural networks (CNN), made remarkable success in visual tasks[1][2][3][4][5] by leveraging large-scale networks learning from a huge volume of data. Deployment of such big models, however, is computation-intensive and memory-intensive. To reduce computation cost, many studies are performed to compress the scale of DNN, including sparsity regularization[6], connection pruning[7][8] and low rank approximation [9][10][11][12][13]. Sparsity regularization and connection pruning approaches, however, often pro- duce non-structured random connectivity in DNN and thus, irregular memory access that adversely impacts practical acceleration in hardware platforms. Figure 1 depicts practical speedup of each layer of a AlexNet , which is non-structurally sparsified by l1-norm. Compared to original model, the accuracy loss of the sparsified model is controlled within 2%. Because of the poor data locality associated with the scattered weight distribution, the achieved speedups are either very limited or negative even the actual sparsity is high, say, >95%. We define sparsity as the ratio of zeros in this paper. In recently proposed low rank approximation approaches, the DNN is trained first and then each trained weight tensor is decomposed and approximated by a product of smaller factors. Finally, fine-tuning is performed to restore the model accuracy. Low rank approximation is able to achieve practical speedups because it coordinates model parameters in dense matrixes and avoids the locality problem of non-structured sparsity regularization. However, low rank approximation can only obtain <

> Figure 1: Evaluation speedups of AlexNet on GPU platforms and the sparsity. conv1 refers to convolutional layer 1, and so forth. Baseline is profiled by GEMM of cuBLAS. The sparse matrixes are stored in the format of Compressed Sparse Row (CSR) and accelerated by cuSPARSE. the compact structure within each layer, and the structures of the layers are fixed during fine-tuning such that costly reiterations of decomposing and fine-tuning are required to find an optimal weight approximation for performance speedup and accuracy retaining. Inspired by the facts that (1) there is redundancy across filters and channels [11]; (2) shapes of filters are usually fixed as cuboid but enabling arbitrary shapes can potentially eliminate unnecessary computation imposed by this fixation; and (3) depth of the network is critical for classification but deeper layers cannot always guarantee a lower error because of the exploding gradients and degradation problem [5], we propose Structured Sparsity Learning (SSL) method to directly learn a compressed structure of deep CNNs by group Lasso regularization during the training. SSL is a generic regularization to adaptively adjust multiple structures in DNN, including structures of filters, channels, and filter shapes within each layer, and structure of depth beyond the layers. SSL combines structure regularization (on DNN for classification accuracy) with locality optimization (on memory access for computation efficiency), offering not only well-regularized big models with improved accuracy but greatly accelerated computation (e.g. 5.1% on CPU and 3.1% on GPU for AlexNet ). 2 Related works Connection pruning and weight sparsifying. Hanet al.[7][8] reduced number of parameters of AlexNet by 9% andVGG-16by 13% using connection pruning. Since most reduction is achieved on fully-connected layers, the authors obtained 3% to 4% layer-wise speedup for fully-connected layers. However, no practical speedups of convolutional layers are observed because of the issue shown in Figure 1. As convolution is the computational bottleneck and many new DNNs use fewer fully-connected layers,e.g., only 3.99% parameters of ResNet -152in [5] are from fully-connected layers, compression and acceleration on convolutional layers become essential. Liuet al.[6] achieved >90% sparsity of convolutional layers in AlexNet with 2% accuracy loss, and bypassed the issue shown in Figure 1 by hardcoding the sparse weights into program, achieving layer-wise 4.59% speedup on a CPU. In this work, we also focus on convolutional layers. Compared to the above techniques, our SSL method can coordinate sparse weights in adjacent memory space and achieve higher speedups with the same accuracy. Note that hardware and program optimizations can further boost the system performance on top of the level of SSL but are not covered in this work. Low rank approximation. Denilet al.[9] predicted 95% parameters in a DNN by exploiting the redundancy across filters and channels. Inspired by it, Jaderberget al.[11] achieved 4.5% speedup on CPUs for scene text character recognition and Dentonet al.[10] achieved 2% speedups on both CPUs and GPUs for the first two layers. Both of the works usedLow Rank Approximation(LRA) with%1% accuracy drop. [13][12] improved and extended LRA to larger DNNs. However, the network structure compressed by LRA is fixed; reiterations of decomposing, training/fine-tuning, and cross-validating are still needed to find an optimal structure for accuracy and speed trade-off. As number of hyper-parameters in LRA method increases linearly with layer depth [10][13], the search space increases linearly or even polynomially for very deep DNNs. Comparing to LRA, our contributions are: (1) SSL can dynamically optimize the compactness of DNN structure with only one hyper-parameter and no reiterations; (2) besides the redundancy within the layers, SSL also exploits the necessity of deep layers and reduce them; (3) DNN filters regularized by SSL have lower rank approximation, so it can work together with LRA for more efficient model compression. Model structure learning.Group Lasso [14] is an efficient regularization to learn sparse structures. Kimet al.[15] used group Lasso to regularize the structure of correlation tree for multi-task regression problem and reduced prediction errors. Liuet al.[6] utilized group Lasso to constrain the scale <> <> Figure 2: The proposed structured sparsity learning (SSL) for DNNs. Weights in filters are split W(l) into multiple groups. Through group Lasso regularization, a more compact DNN is obtained by :,c l ,:,: removing some groups. The figure illustrates the filter-wise, channel-wise, shape-wise, and depth-wise structured sparsity that were explored in the work. <> of the structure of LRA. To adapt DNN structure to different databases, Fenget al.[16] learned the appropriate number of filters in DNN. Different from these prior arts, we apply group Lasso to regularize multiple DNN structures (filters, channels, filter shapes, and layer depth). Our source code can be found at https://github.com/wenwei202/caffe/tree/scnn. 3 Structured Sparsity Learning Method for DNNs We focus mainly on theStructured Sparsity Learning(SSL) on convolutional layers to regularize the structure of DNNs. We first propose a generic method to regularize structures of DNN in Section 3.1, 1 and then specify the method to structures of filters, channels, filter shapes and depth in section 3.2. Variants of formulations are also discussed from computational efficiency viewpoint in Section 3.3. 3.1 Proposed structured sparsity learning for generic structures Suppose weights of convolutional layers in a DNN form a sequence of 4-D tensors <>, where <> and <> are the dimensions of the l-th weight tensor along the axes of filter, channel, spatial height and spatial width, respectively. L denotes the number of convolutional layers. Then the proposed generic optimization target of a DNN with structured sparsity regularization can be formulated as: 1 <> (1) Here W represents the collection of all weights in the <> is the loss on data <> is non-structured regularization applying on every weight,e.g., l2-norm; and <> is the structured sparsity regularization on each layer. Because Group Lasso can effectively zero out all weights in some groups [14][15], we adopt it in our SSL. The regularization of group Lasso on a set of weights Pw can be represented as <>, where <> is a group of partial weights in w and G is the total number of groups. Different groups may overlap. Here <>, where <> the number of weights in <>. 3.2 Structured sparsity learning for structures of filters, channels, filter shapes and depth In SSL, the learned “structure” is decided by the way of splitting groups ofw(g) . We investigate and formulate thefiler-wise,channel-wise,shape-wise, and depth-wise structured sparsity in Figure 2. For simplicity, the <> term of Eq. (1) is omitted in the following formulation expressions. Penalizing unimportant filers and channels. Suppose <> is then l-th filter and <> is the cl-th channel of all filters in the l-th layer. The optimization target of learning the filter-wise and channel-wise structured sparsity can be defined as <> (2) As indicated in Eq. (2), our approach tends to remove less important filters and channels. Note that zeroing out a filter in the l-th layer results in a dummy zero output feature map, which in turn makes a corresponding channel in the (l+ 1)-th layer useless. Hence, we combine the filter-wise and channel-wise structured sparsity in the learning simultaneously. Learning arbitrary shapes of filers. As illustrated in Figure 2, <> denotes the vector of :;c l ;m l ;k all corresponding weights located at spatial position of <> in the 2D filters across the cl-th channel. Thus, we defineW(l) as the shape fiber related to learning arbitrary filter shape <> because a homogeneous non-cubic filter shape can be learned by zeroing out some shape fibers. The l optimization target of learning shapes of filers becomes: <> (3) Regularizing layer depth. We also explore the depth-wise sparsity to regularize the depth of DNNs in order to improve accuracy and reduce computation cost. The corresponding optimization target is Different from other discussed sparsification techniques, zeroing out all the filters in a layer will cut off the message propagation in the DNN so that the output neurons cannot perform any classification. Inspired by the structure of highway networks [17] and deep residual networks [5], we propose to leverage the shortcuts across layers to solve this issue. As illustrated in Figure 2, even when SSL removes an entire unimportant layers, feature maps will still be forwarded through the shortcut. 3.3 Structured sparsity learning for computationally efficient structures All proposed schemes in section 3.2 can learn a compact DNN for computation cost reduction. Moreover, some variants of the formulations of these schemes can directly learn structures that can be efficiently computed. 2D-filter-wise sparsity for convolution. 3D convolution in DNNs essentially is a composition of 2D convolutions. To perform efficient convolution, we explored a fine-grain variant of filter-wise sparsity, namely,2D-filter-wise sparsity, to spatially enforce group Lasso on each 2D filter ofW(l)nl ;c l ;:;: . The saved convolution is proportional to the percentage of the removed 2D filters. The fine-grain version of filter-wise sparsity can more efficiently reduce the computation associated with convolution: Because the group sizes are much smaller and thus the weight updating gradients are shaper, it helps group Lasso to quickly obtain a high ratio of zero groups for a large-scale DNN. Combination of filter-wise and shape-wise sparsity for GEMM. Convolutional computation in DNNs is commonly converted to modality of general Matrix Multiplication (GEMM) by lowering weight tensors and feature tensors to matrices [18]. For example, in Caffe [19], a 3D filter <> is reshaped to a row in the weight matrix where each column is the collection of weights <> related to shape-wise sparsity. Combining filter-wise and shape-wise sparsity can directly reduce the dimension of weight matrix in GEMM by removing zero rows and columns. In this context, we use row-wise and column-wise sparsity as the interchangeable terminology of filter-wise and shape-wise sparsity, respectively. 4 Experiments We evaluated the effectiveness of our SSL using published models on three databases – MNIST, CIFAR-10, and ImageNet. Without explicit explanation, SSL starts with the network whose weights are initialized by the baseline, and speedups are measured in matrix-matrix multiplication by Caffe in a single-thread Intel Xeon E5-2630 CPU . Table 1: Results after penalizing unimportant filters and channels inLeNet <

> 4.1 LeNet and multilayer perceptron on MNIST In the experiment of MNIST, we examined the effectiveness of SSL in two types of networks: LeNet[20] implemented by Caffe and amultilayer perceptron(MLP) network. Both networks were trained without data augmentation. LeNet:When applying SSL toLeNet, we constrain the network with filter-wise and channel-wise sparsity in convolutional layers to penalize unimportant filters and channels. Table 1 summarizes the remained filters and channels,floating-point operations(FLOP), and practical speedups. In the table,LeNet 1is the baseline and the others are the results after applying SSL in different strengths of structured sparsity regularization. The results show that our method achieves the similar error (0.1%) with much fewer filters and channels, and saves significant FLOP and computation time. To demonstrate the impact of SSL on the structures of filters, we present all learned conv1 filters in Figure 3. It can be seen that most filters inLeNet 2are entirely zeroed out except for five most important detectors of stroke patterns that are sufficient for feature extraction. The accuracy of LeNet 3(that further removes the weakest and redundant stroke detector) drops only 0.2% from that ofLeNet 2. Compared to the random and blurry filter patterns inLeNet 1that resulted from the high freedom of parameter space, the filters inLeNet 2 & 3are regularized and converge to smoother and more natural patterns. This explains why our proposed SSL obtains the same-level accuracy but has much less filters. The smoothness of the filters are also observed in the deeper layers. The effectiveness of the shape-wise sparsity on LeNet is summarized in Table 2. The baselineLeNet 1 has conv1 filters with a regular 5x5 square (size = 25) whileLeNet 5reduces the dimension that can be constrained by a 2x4 rectangle (size = 7). The 3D shape of conv2 filters in the baseline is also regularized to the 2D shape inLeNet 5within only one channel, indicating that only one filter in conv1 is needed. This fact significantly saves FLOP and computation time. <

> Figure 3: Learned conv1 filters in LeNet 1(top),LeNet 2(middle) and LeNet 3(bottom) MLP:Besides convolutional layers, our proposed SSL can be extended to learn the structure (i.e.the number of neurons) of fully-connected layers. We enforce the group Lasso regularization on all the input (or output) connections of each neuron. A neuron whose input connections are all zeroed out can degenerate to a bias neuron in the next layer; similarly, a neuron can degenerate to a removable dummy neuron if all of its output connections are zeroed out. Figure 4(a) summarizes the learned structure and FLOP of differentMLPnetworks. The results show that SSL can not only remove hidden neurons but also discover the sparsity of images. For example, Figure 4(b) depicts the number of connections of each input neuron inMLP 2, where 40.18% of input neurons have zero connections and they concentrate at the boundary of the image. Such a distribution is consistent with our intuition: Table 2: Results after learning filter shapes inLeNet <

> Figure 4: The normalized reconstructure error of weight matrix vs. the percent of ranks.Principal Component Analysis(PCA) is utilized to explore the redundancy among filters.% ranks of eigenvectors corresponding to the largest eigenvalues are selected as basis to perform low rank approximation. Left:LeNet2 in Table 1; middle: ConvNet2 in Table 4; right: AlexNet 4 in Table 5. Dash lines indicate baselines and solid lines indicate results of SSL. 170 detectors of stroke patterns which are sufficient for feature extraction. The accuracy ofLeNet 3 171 (that further removes one weakest and one redundant stroke detector) compared withLeNet 2drops 172 only 0.2%. Although the training processes of three networks are independent, the corresponding 173 regularized filters inLeNet 2andLeNet 3demonstrate very high similarity and represent certain level 174 of alikeness to those inLeNet 1. Comparing with random and blurry filter patterns inLeNet 1resulted 175 from the high freedom of parameter space, the filters inLeNet 2 & 3are regularized through the 176 filter-wise and channel-wise sparsity and therefore converge at smoother and more natural patterns. 177 This explains why our proposed SSL obtains the same-level accuracy but having much less filters. 178 These regularity and similarity phenomena are also observed in deeper layers. Different from low 179 rank decomposition which only explore the redundancy and does not change the rank, SSL can reduce 180 the redundancy as shown in Figure 4. 181 We also explore the effectiveness of the shape-wise sparsity onLeNetin Table 2. The baselineLeNet 182 1has a regular5⇥5square size of conv1 filters, whileLeNet 5reduces the dimension to less than 183 2⇥4. And the 3D shape of filters inconv2ofLeNet 1are regularized to 2D shape ofLeNet 5with 184 only one channel, indicating that only one filter in conv1 is needed. This saves significant FLOP and 185 computing time. 186 MLP:Besides convolutional layers, our proposed SSL can be extended to learn the structure (i.e. 187 the number of neurons) in fully-connected layers. Here, the baselineMLPnetwork composed of 188 two hidden layers with 500 and 300 neurons respectively obtains a test error of 1.43%. We enforced 189 the group Lasso regularization on all the input (or output) connections of every neuron, including 190 those of the input layer. Note that a neuron with all the input connections zeroed out degenerate 191 to a bias neuron in the next layer; similarly, a neuron degenerates to a removable dummy neuron 192 if all of its output connections are zeroed out. As such, the computation ofGEneral Matrix Vector 193 (GEMV) product in fully-connected layers can be significantly reduced. Table 3 summarizes the Table 3: Learning the number of neurons in multi-layer perceptron <> Figure 4: (a) Results of learning the number of neurons inMLP. (b) the connection numbers of input <

> handwriting digits are usually written in the center and pixels close to the boundary contain little discriminative classification information. 4.2 ConvNet and ResNet on CIFAR-10 We implemented the ConvNet of [1] and deep residual networks( ResNet ) [5] on CIFAR-10. When regularizing filters, channels, and filter shapes, the results and observations of both networks are similar to that of the MNIST experiment. Moreover, we simultaneously learn the filter-wise and shape-wise sparsity to reduce the dimension of weight matrix in GEMM ofConvNet. We also learn the depth-wise sparsity of ResNet to regularize the depth of the DNNs. ConvNet:We use the network from Alex Krizhevskyet al.[1] as the baseline and implement it using Caffe. All the configurations remain the same as the original implementation except that we added a dropout layer with a ratio of 0.5 in the fully-connected layer to avoid over-fitting.ConvNetis trained without data augmentation. Table 3 summarizes the results of threeConvNetnetworks. Here, the row/column sparsity of a weight matrix is defined as the percentage of all-zero rows/columns. Figure 5 shows their learned conv1 filters. In Table 3, SSL can reduce the size of weight matrix inConvNet 2by 50%, 70.7% and 36.1% for each convolutional layer and achieve good speedups without accuracy drop. Surprisingly, without SSL, four conv1 filters of the baseline are actually all-zeros as shown in Figure 5, demonstrating the great potential of filter sparsity. When SSL is applied, half of conv1 filters inConvNet 2can be zeroed out without accuracy drop. On the other hand, inConvNet 3, SSL achieves 1.0% (0.16%) lower error with a model even smaller than the baseline. In this scenario, SSL performs as a structure regularization to dynamically learn a better network structure (including the number of filters and filer shapes) to reduce the error. <

> Figure 5: Learned conv1 filters inConvNet 1(top),ConvNet 2(middle) andConvNet 3(bottom) ResNet :To investigate the necessary depth of DNNs required by SSL, we use a 20-layer deep residual networks ( ResNet -20) proposed in [5] as the baseline. The network has 19 convolutional layers and 1 fully-connected layer.Identity shortcuts are utilized to connect the feature maps with the same dimension while 1%1 convolutional layers are chosen as shortcuts between the feature maps with different dimensions. Batch normalization [21] is adopted after convolution and before activation. We use the same data augmentation and training hyper-parameters as that in [5]. The final error of baseline is 8.82%. In SSL, the depth of ResNet -20is regularized by depth-wise sparsity. Group Lasso regularization is only enforced on the convolutional layers between each pair of shortcut endpoints, excluding the first convolutional layer and all convolutional shortcuts. After SSL converges, layers <

> Figure 6: Error vs. layer number after depth regularization by SSL. in [ 1412 5] with # layers.SSL- ResNet -#is the depth-regularized ResNet by SSL with # layers, including the last fully-connected layer indicates the convolutional layers with an output map size of 32,64 32, and so forth with all zero weights are removed and the net is finally fine-tuned with a base learning rate of 0.01, Figure 6 plots the trend of the error vs. the number of layers under different strengths of depth regularizations. Compared with original ResNet in [5], SSL learns a ResNet with 14 layers (SSL- ResNet -14) that reaching a lower error than the one of the baseline with 20 layers ( ResNet -20); SSL- ResNet -18and ResNet -32achieve an error of 7.40% and 7.51%, respectively. This result implies that SSL can work as a depth regularization to improve classification accuracy. Note that SSL can efficiently learn shallower DNNs without accuracy loss to reduce computation cost; however, it does not mean the depth of the network is not important. The trend in Figure 6 shows that the test error generally declines as more layers are preserved. A slight error rise of SSL-ResNet-20 from SSL- ResNet -18shows the suboptimal selection of the depth in the group of “32x32”. 4.3 AlexNet on ImageNet To show the generalization of our method to large scale DNNs, we evaluate SSL using AlexNet with ILSVRC 2012.CaffeNet[19] – the replication of AlexNet [1] with mirror changes, is used in our experiment. All training images are rescaled to the size of 256x256. A 227%227 image is randomly cropped from each scaled image and mirrored for data augmentation and only the center crop is used for validation. The final top-1 validation error is 42.63%. In SSL, AlexNet is first trained with structure regularization; when it converges, zero groups are removed to obtain a DNN with the new structure; finally, the network is fine-tuned without SSL to regain the accuracy. We first studied 2D-filter-wise and shape-wise sparsity by exploring the trade-offs between computation complexity and classification accuracy. Figure 7(a) shows the 2D-filter sparsity (the ratio between the removed 2D filters and total 2D filters) and the saved FLOP of 2D convolutions vs. the validation error. In Figure 7(a), deeper layers generally have higher sparsity as the group size shrinks <

> Figure 7: (a) 2D-filter-wise sparsity and FLOP reduction vs. top-1 error. Vertical dash line shows the error of original AlexNet ; (b) The reconstruction error of weight tensor vs. dimensionality.Principal Component Analysis(PCA) is utilized to perform dimensionality reduction to exploit filter redundancy. The eigenvectors corresponding to the largest eigenvalues are selected as basis of lower-dimensional space. Dash lines denote the results of the baselines and solid lines indicate the ones of the AlexNet 5 in Table 4; (c) Speedups of‘1 -norm and SSL on various CPU and GPU platforms (In labels of x-axis, T# is number of the maximum physical threads in Xeon CPU). AlexNet 1and AlexNet 2in Table 4 are used as test benches. and the number of 2D filters grows. 2D-filter sparsity regularization can reduce the total FLOP by 30%–40% without accuracy loss or reduce the error of AlexNet by%1% down to 41.69% by retaining the original number of parameters. Shape-wise sparsity also obtains similar results – In Table 4, for example, AlexNet 5achieves on average 1.4%layer-wise speedup on both CPU and GPU without accuracy loss after shape regularization; The top-1 error can also be reduced down to 41.83% if the parameters are retained. In Figure 7(a), the obtained DNN with the lowest error has a very low sparsity, indicating that the number of parameters in a DNN is still important to maintain learning capacity. In this case, SSL works as a regularization to add restriction of smoothness to the model in order to avoid over-fitting. Figure 7(b) compares the results of dimensionality reduction of weight tensors in the baseline and our SSL-regularized AlexNet . The results show that the smoothness restriction enforces parameter searching in lower-dimensional space and enables lower rank approximation of the DNNs. Therefore, SSL can work together with low rank approximation to achieve even higher model compression. Besides the above analyses, the computation efficiencies of structured sparsity and non-structured sparsity are compared in Caffe using standard off-the-shelf libraries,i.e., Intel Math Kernel Library on CPU and CUDA cuBLAS and cuSPARSE on GPU. We use SSL to learn a AlexNet with high column-wise and row-wise sparsity as the representative of structured sparsity method.‘1 -norm is selected as the representative of non-structured sparsity method instead of connection pruning in [7] because‘1 -norm get a higher sparsity on convolutional layers as the results of AlexNet 3and AlexNet 4depicted in Table 4. Speedups achieved by SSL are measured by subroutines of GEMM where nonzero rows and columns in each weight matrix are concatenated in consecutive memory space. Note that compared to GEMM, the overhead of concatenation can be ignored. To measure the speedups of‘1 -norm, sparse weight matrices are stored in the format of Compressed Sparse Row (CSR) and computed by sparse-dense matrix multiplication subroutines. Table 4 compares the obtained sparsity and speedups of‘1 -norm and SSL on CPU (Intel Xeon) and GPU (GeForce GTX TITAN Black) under approximately the same errors,e.g., with acceptable or no accuracy loss. For a fair comparison, after‘1 -norm regularization, the DNN is also fine- tuned by disconnecting all zero-weighted connections so that 1.39% accuracy is recovered for the AlexNet 1. Our experiments show that the DNNs require a very high non-structured sparsity to achieve a reasonable speedup (The speedups are even negative when the sparsity is low). SSL, however, can always achieve positive speedups. With an acceptable accuracy loss, our SSL achieves on average 5.1% and 3.1% layer-wise acceleration on CPU and GPU, respectively. Instead,‘1 -norm achieves on average only 3.0% and 0.9% layer-wise acceleration on CPU and GPU, respectively. We note that at the same accuracy, our average speedup is indeed higher than that of [6] which adopts heavy hardware customization to overcome the negative impact of non-structured sparsity. Figure 7(c) shows the speedups of‘1 -norm and SSL on various platforms, including both GPU (Quadro, Tesla Table 4: Sparsity and speedup of AlexNet on ILSVRC 2012 <

> and Titan) and CPU (Intel Xeon E5-2630). SSL can achieve on average%3%speedup on GPU while non-structured sparsity obtain no speedup on GPU platforms. On CPU platforms, both methods can achieve good speedups and the benefit grows as the processors become weaker. Nonetheless, SSL can always achieve averagely%2%speedup compared to non-structured sparsity. 5 Conclusion In this work, we have proposed aStructured Sparsity Learning(SSL) method to regularize filter, channel, filter shape, and depth structures in deep neural networks (DNN). Our method can enforce the DNN to dynamically learn more compact structures without accuracy loss. The structured compactness of the DNN achieves significant speedups for the DNN evaluation both on CPU and GPU with off-the-shelf libraries. Moreover, a variant of SSL can be performed as structure regularization to improve classification accuracy of state-of-the-art DNNs. Acknowledgments This work was supported in part by NSF XPS-1337198 and NSF CCF-1615475. The authors thank Drs. Sheng Li and Jongsoo Park for valuable feedback on this work. References [1]Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. InAdvances in Neural Information Processing Systems, pages 1097–1105. 2012. [2]Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. InThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. [3]Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recogni- tion.arXiv preprint arXiv:1409.1556, 2014. [4]Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions.arXiv preprint arXiv:1409.4842, 2015. [5]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015. [6]Baoyuan Liu, Min Wang, Hassan Foroosh, Marshall Tappen, and Marianna Pensky. Sparse convolutional neural networks. InThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. [7]Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. InAdvances in Neural Information Processing Systems, pages 1135–1143. 2015. [8]Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding.arXiv preprint arXiv:1510.00149, 2015. [9] Misha Denil, Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato, and Nando de Freitas. Predicting parameters in deep learning. InAdvances in Neural Information Processing Systems, pages 2148–2156. 2013. [10]Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. InAdvances in Neural Information Processing Systems, pages 1269–1277. 2014. [11]Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions.arXiv preprint arXiv:1405.3866, 2014. [12]Yani Ioannou, Duncan P. Robertson, Jamie Shotton, Roberto Cipolla, and Antonio Criminisi. Training cnns with low-rank filters for efficient image classification.arXiv preprint arXiv:1511.06744, 2015. [13]Cheng Tai, Tong Xiao, Xiaogang Wang, and Weinan E. Convolutional neural networks with low-rank regularization.arXiv preprint arXiv:1511.06067, 2015. [14]Ming Yuan and Yi Lin. Model selection and estimation in regression with grouped variables.Journal of the Royal Statistical Society. Series B (Statistical Methodology), 68(1):49–67, 2006. [15]Seyoung Kim and Eric P Xing. Tree-guided group lasso for multi-task regression with structured sparsity. InProceedings of the 27th International Conference on Machine Learning, 2010. [16]Jiashi Feng and Trevor Darrell. Learning the structure of deep convolutional networks. InThe IEEE International Conference on Computer Vision (ICCV), 2015. [17]Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Highway networks.arXiv preprint arXiv:1505.00387, 2015. [18]Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, and Evan Shelhamer. cudnn: Efficient primitives for deep learning.arXiv preprint arXiv:1410.0759, 2014. [19]Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding.arXiv preprint arXiv:1408.5093, 2014. [20]Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998. [21]Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift.arXiv preprint arXiv:1502.03167, 2015. <> <> <> <> <> <> MIXED PRECISION TRAINING Sharan Narang % , Gregory Diamos, Erich Elsen y Baidu Research fsharan, gdiamosg@baidu.com Paulius Micikevicius % , Jonah Alben, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, Hao Wu NVIDIA fpauliusm, alben, dagarcia, bginsburg, mhouston, okuchaiev, gavenkatesh, skywg@nvidia.com ABSTRACT Increasing the size of a neural network typically improves accuracy but also in- creases the memory and compute requirements for training the model. We intro- duce methodology for training deep neural networks using half-precision float- ing point numbers, without losing model accuracy or having to modify hyper- parameters. This nearly halves memory requirements and, on recent GPUs, speeds up arithmetic. Weights, activations, and gradients are stored in IEEE half- precision format. Since this format has a narrower range than single-precision we propose three techniques for preventing the loss of critical information. Firstly, we recommend maintaining a single-precision copy of weights that accumulates the gradients after each optimizer step (this copy is rounded to half-precision for the forward- and back-propagation). Secondly, we propose loss-scaling to pre- serve gradient values with small magnitudes. Thirdly, we use half-precision arith- metic that accumulates into single-precision outputs, which are converted to half- precision before storing to memory. We demonstrate that the proposed methodology works across a wide variety of tasks and modern large scale (exceeding 100 million parameters) model architectures, trained on large datasets. 1 INTRODUCTION Deep Learning has enabled progress in many different applications, ranging from image recognition (He et al., 2016a) to language modeling (Jozefowicz et al., 2016) to machine translation (Wu et al., 2016) and speech recognition (Amodei et al., 2016). Two trends have been critical to these results - increasingly large training data sets and increasingly complex models. For example, the neural network used in Hannun et al. (2014) had 11 million parameters which grew to approximately 67 million for bidirectional RNNs and further to 116 million for the latest forward only Gated Recurrent Unit (GRU) models in Amodei et al. (2016). Larger models usually require more compute and memory resources to train. These requirements can be lowered by using reduced precision representation and arithmetic. Performance (speed) of any program, including neural network training and inference, is limited by one of three factors: arithmetic bandwidth, memory bandwidth, or latency. Reduced precision addresses two of these limiters. Memory bandwidth pressure is lowered by using fewer bits to to store the same number of values. Arithmetic time can also be lowered on processors that offer higher throughput for reduced precision math. For example, half-precision math throughput in recent GPUs is 2% to 8% higher than for single-precision. In addition to speed improvements, reduced precision formats also reduce the amount of memory required for training. Modern deep learning training systems use single-precision (FP32) format. In this paper, we address the training with reduced precision while maintaining model accuracy. Specifically, we train various neural networks using IEEE half-precision format (FP16). Since FP16 format has a narrower dynamic range than FP32, we introduce three techniques to prevent model accuracy loss: maintain- ing a master copy of weights in FP32, loss-scaling that minimizes gradient values becoming zeros, and FP16 arithmetic with accumulation in FP32. Using these techniques we demonstrate that a wide variety of network architectures and applications can be trained to match the accuracy FP32 training. Experimental results include convolutional and recurrent network architectures, trained for classification, regression, and generative tasks. Applications include image classification, image generation, object detection, language modeling, machine translation, and speech recognition. The proposed methodology requires no changes to models or training hyper-parameters. 2 RELATED WORK There have been a number of publications on training Convolutional Neural Networks (CNNs) with reduced precision. Courbariaux et al. (2015) proposed training with binary weights, all other tensors and arithmetic were in full precision. Hubara et al. (2016a) extended that work to also binarize the activations, but gradients were stored and computed in single precision. Hubara et al. (2016b) considered quantization of weights and activations to 2, 4 and 6 bits, gradients were real numbers. Rastegari et al. (2016) binarize all tensors, including the gradients. However, all of these approaches lead to non-trivial loss of accuracy when larger CNN models were trained for ILSVRC classification task (Russakovsky et al., 2015). Zhou et al. (2016) quantize weights, activations, and gradients to different bit counts to further improve result accuracy. This still incurs some accuracy loss and requires a search over bit width configurations per network, which can be impractical for larger models. Mishra et al. improve on the top-1 accuracy achieved by prior weight and activation quantizations by doubling or tripling the width of layers in popular CNNs. However, the gradients are still computed and stored in single precision, while quantized model accuracy is lower than that of the widened baseline. Gupta et al. (2015) demonstrate that 16 bit fixed point representation can be used to train CNNs on MNIST and CIFAR-10 datasets without accuracy loss. It is not clear how this approach would work on the larger CNNs trained on large datasets or whether it would work for Recurrent Neural Networks (RNNs). There have also been several proposals to quantize RNN training. He et al. (2016c) train quantized variants of the GRU (Cho et al., 2014) and Long Short Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997) cells to use fewer bits for weights and activations, albeit with a small loss in accuracy. It is not clear whether their results hold for larger networks needed for larger datasets Hubara et al. (2016b) propose another approach to quantize RNNs without altering their structure. Another approach to quantize RNNs is proposed in Ott et al. (2016). They evaluate binary, ternary and exponential quantization for weights in various different RNN models trained for language modelling and speech recognition. All of these approaches leave the gradients unmodified in single- precision and therefore the computation cost during back propagation is unchanged. The techniques proposed in this paper are different from the above approaches in three aspects. First, all tensors and arithmetic for forward and backward passes use reduced precision, FP16 in our case. Second, no hyper-parameters (such as layer width) are adjusted. Lastly, models trained with these techniques do not incur accuracy loss when compared to single-precision baselines. We demonstrate that this technique works across a variety of applications using state-of-the-art models trained on large scale datasets. 3 IMPLEMENTATION We introduce the key techniques for training with FP16 while still matching the model accuracy of FP32 training session: single-precision master weights and updates, loss-scaling, and accumulating FP16 products into FP32. Results of training with these techniques are presented in Section 4. 3.1 FP32 MASTER COPY OF WEIGHTS In mixed precision training, weights, activations and gradients are stored as FP16. In order to match the accuracy of the FP32 networks, an FP32 master copy of weights is maintained and updated with the weight gradient during the optimizer step. In each iteration an FP16 copy of the master weights is used in the forward and backward pass, halving the storage and bandwidth needed by FP32 training. Figure 1 illustrates this mixed precision training process. While the need for FP32 master weights is not universal, there are two possible reasons why a number of networks require it. One explanation is that updates (weight gradients multiplied by the learning rate) become too small to be represented in FP16 - any value whose magnitude is smaller than2%24 becomes zero in FP16. We can see in Figure 2b that approximately 5% of weight gradient values have exponents smaller than%24. These small valued gradients would become zero in the optimizer when multiplied with the learning rate and adversely affect the model accuracy. Using a single-precision copy for the updates allows us to overcome this problem and recover the accuracy. Another explanation is that the ratio of the weight value to the weight update is very large. In this case, even though the weight update is representable in FP16, it could still become zero when addition operation right-shifts it to align the binary point with the weight. This can happen when the magnitude of a normalized weight value is at least 2048 times larger that of the weight update. Since FP16 has 10 bits of mantissa, the implicit bit must be right-shifted by 11 or more positions to potentially create a zero (in some cases rounding can recover the value). In cases where the ratio is larger than 2048, the implicit bit would be right-shifted by 12 or more positions. This will cause the weight update to become a zero which cannot be recovered. An even larger ratio will result in this effect for de-normalized numbers. Again, this effect can be counteracted by computing the update in FP32. To illustrate the need for an FP32 master copy of weights, we use the Mandarin speech model (described in more detail in Section 4.3) trained on a dataset comprising of approximately 800 hours of speech data for 20 epochs. As shown in 2a, we match FP32 training results when updating an FP32 master copy of weights after FP16 forward and backward passes, while updating FP16 weights results in 80% relative accuracy loss. Even though maintaining an additional copy of weights increases the memory requirements for the weights by 50% compared with single precision training, impact on overall memory usage is much smaller. For training memory consumption is dominated by activations, due to larger batch sizes and activations of each layer being saved for reuse in the back-propagation pass. Since activations are also stored in half-precision format, the overall memory consumption for training deep neural networks is roughly halved. 3.2 LOSS SCALING FP16 exponent bias centers the range of normalized value exponents to[%14;15]while gradient values in practice tend to be dominated by small magnitudes (negative exponents). For example, consider Figure 3 showing the histogram of activation gradient values, collected across all layers during FP32 training of Multibox SSD detector network (Liu et al., 2015a). Note that much of the FP16 representable range was left unused, while many values were below the minimum representable range and became zeros. Scaling up the gradients will shift them to occupy more of the representable range and preserve values that are otherwise lost to zeros. This particular network diverges when gradients are not scaled, but scaling them by a factor of 8 (increasing the exponents by 3) is sufficient to match the accuracy achieved with FP32 training. This suggests that activation gradient values below2%27 in magnitude were irrelevant to the training of this model, but values in the[2 %27 ;2%24 )range were important to preserve. One efficient way to shift the gradient values into FP16-representable range is to scale the loss value computed in the forward pass, prior to starting back-propagation. By chain rule back-propagation ensures that all the gradient values are scaled by the same amount. This requires no extra operations during back-propagation and keeps the relevant gradient values from becoming zeros. Weight gradients must be unscaled before weight update to maintain the update magnitudes as in FP32 training. It is simplest to perform this unscaling right after the backward pass but before gradient clipping or any other gradient-related computations, ensuring that no hyper-parameters (such as gradient clipping threshold, weight decay, etc.) have to be adjusted. There are several options to choose the loss scaling factor. The simplest one is to pick a constant scaling factor. We trained a variety of networks with scaling factors ranging from 8 to 32K (many networks did not require a scaling factor). A constant scaling factor can be chosen empirically or, if gradient statistics are available, directly by choosing a factor so that its product with the maximum absolute gradient value is below 65,504 (the maximum value representable in FP16). There is no downside to choosing a large scaling factor as long as it does not cause overflow during back-propagation - overflows will result in infinities and NaNs in the weight gradients which will irreversibly damage the weights after an update. Note that overflows can be efficiently detected by inspecting the computed weight gradients, for example, when weight gradient values are unscaled. One option is to skip the weight update when an overflow is detected and simply move on to the next iteration. <

> Figure 2: Figure 2a shows the results of three experiments; baseline (FP32), pseudo FP16 with FP32 master copy, pseudo FP16 without FP32 master copy. Figure 2b shows the histogram for the exponents of weight gradients for Mandarin speech recognition training with FP32 weights. The gradients are sampled every 4,000 iterations during training for all the layers in the model. <

> Figure 3: Histogram of activation gradient values during the training of Multibox SSD network. Note that the bins on the x-axis cover varying ranges and there’s a separate bin for zeros. For example, 2% of the values are in the[2 %34 ;2%32 )range, 2% of values are in the[2 %24 ;2%23 )range, and 67% of values are zero. 3.3 ARITHMETIC PRECISION By and large neural network arithmetic falls into three categories: vector dot-products, reductions, and point-wise operations. These categories benefit from different treatment when it comes to reduced precision arithmetic. To maintain model accuracy, we found that some networks require that FP16 vector dot-product accumulates the partial products into an FP32 value, which is converted to FP16 before writing to memory. Without this accumulation in FP32, some FP16 models did not match the accuracy of the baseline models. Whereas previous GPUs supported only FP16 multiply- add operation, NVIDIA Volta GPUs introduce Tensor Cores that multiply FP16 input matrices and accumulate products into either FP16 or FP32 outputs (NVIDIA, 2017). Large reductions (sums across elements of a vector) should be carried out in FP32. Such reductions mostly come up in batch-normalization layers when accumulating statistics and softmax layers. Both of the layer types in our implementations still read and write FP16 tensors from memory, performing the arithmetic in FP32. This did not slow down the training process since these layers are memory-bandwidth limited and not sensitive to arithmetic speed. Point-wise operations, such as non-linearities and element-wise matrix products, are memory- bandwidth limited. Since arithmetic precision does not impact the speed of these operations, either FP16 or FP32 math can be used. 4 RESULTS We have run experiments for a variety of deep learning tasks covering a wide range of deep learning models. We conducted the following experiments for each application: %Baseline (FP32): Single-precision storage is used for activations, weights and gradients. All arithmetic is also in FP32. %Mixed Precision (MP): FP16 is used for storage and arithmetic. Weights, activations and gradients are stored using in FP16, an FP32 master copy of weights is used for updates. Loss-scaling is used for some applications. Experiments with FP16 arithmetic used Tensor Core operations with accumulation into FP32 for convolutions, fully-connected layers, and matrix multiplies in recurrent layers. The Baseline experiments were conducted on NVIDIA’s Maxwell or Pascal GPU. Mixed Precision experiments were conducted on Volta V100 that accumulates FP16 products into FP32. The mixed precision speech recognition experiments (Section 4.3) were conducted using Maxwell GPUs using FP16 storage only. This setup allows us to emulate the TensorCore operations on non-Volta hard- ware. A number of networks were trained in this mode to confirm that resulting model accuracies are equivalent to MP training run on Volta V100 GPUs. This is intuitive since MP arithmetic was accumulating FP16 products into FP32 before converting the result to FP16 on a memory write. 4.1 CNN S FOR ILSVRC CLASSIFICATION We trained several CNNs for ILSVRC classification task (Russakovsky et al., 2015) using mixed precision: Alexnet, VGG-D, GoogLeNet, Inception v2, Inception v3, and pre-activation Resnet-50. In all of these cases we were able to match the top-1 accuracy of baseline FP32 training session using identical hyper-parameters. Networks were trained using Caffe (Jia et al., 2014) framework modified to use Volta TensorOps, except for Resnet50 which used PyTorch (Paszke et al., 2017). Training schedules were used from public repositories, when available (training schedule for VGG- D has not been published). Top-1 accuracy on ILSVRC validation set are shown in Table 1. Baseline (FP32) accuracy in a few cases is different from published results due to single-crop testing and a simpler data augmentation. Our data augmentation in Caffe included random horizontal flipping and random cropping from 256x256 images, Resnet50 training in PyTorch used the full augmentation in the training script from PyTorch vision repository. Table 1: ILSVRC12 classification top-1 accuracy. <

> Loss-scaling technique was not required for successful mixed precision training of these networks. While all tensors in the forward and backward passes were in FP16, a master copy of weights was updated in FP32 as outlined in Section 3.1. 4.2 DETECTION CNN'S Object detection is a regression task, where bounding box coordinate values are predicted by the network (compared to classification, where the predicted values are passed through a softmax layer to convert them to probabilities). Object detectors also have a classification component, where prob- abilities for an object type are predicted for each bounding box. We trained two popular detection approaches: Faster-RCNN (Ren et al., 2015) and Multibox-SSD (Liu et al., 2015a). Both detectors used VGG-16 network as the backbone. Models and training scripts were from public repositories (Girshick; Liu). Mean average precision (mAP) was computed on Pascal VOC 2007 test set. Faster- RCNN was trained on VOC 2007 training set, whereas SSD was trained on a union of VOC 2007 and 2012 data, which is the reason behind baseline mAP difference in Table 2. Table 2: Detection network average mean precision. <> As can be seen in table 2, SSD detector failed to train in FP16 without loss-scaling. By losing small gradient values to zeros, as described in Section 3.2, poor weights are learned and training diverges. As described in Section 3.2, loss-scaling factor of 8 recovers the relevant gradient values and mixed-precision training matches FP32 mAP. 4.3 SPEECH RECOGNITION We explore mixed precision training for speech data using the DeepSpeech 2 model for both English and Mandarin datasets. The model used for training on the English dataset consists of two 2D con- volution layers, three recurrent layers with GRU cells, 1 row convolution layer and Connectionist temporal classification (CTC) cost layer (Graves et al., 2006). It has approximately 115 million parameters. This model is trained on our internal dataset consisting of 6000 hours of English speech. The Mandarin model has a similar architecture with a total of 215 million parameters. The Man- darin model was trained on 2600 hours of our internal training set. For these models, we run the Baseline and Pseudo FP16 experiments. All the models were trained for 20 epochs using Nesterov Stochastic Gradient Descent (SGD). All hyper-parameters such as learning rate, annealing schedule and momentum were the same for baseline and pseudo FP16 experiments. Table 3 shows the results of these experiments on independent test sets. Table 3: Character Error Rate (CER) using mixed precision training for speech recognition. English results are reported on the WSJ ’92 test set. Mandarin results are reported on our internal test set. <> Similar to classification and detection networks, mixed precision training works well for recurrent neural networks trained on large scale speech datasets. These speech models are the largest models trained using this technique. Also, the number of time-steps involved in training a speech model are unusually large compared to other applications using recurrent layers. As shown in table 3, Pseudo FP16 results are roughly 5 to 10% better than the baseline. This suggests that the half-precision storage format may act as a regularizer during training. <> Figure 4: English to French translation network training perplexity, 3x1024 LSTM model with attention. Ref1, ref2 and ref3 represent three different FP32 training runs. 4.4 MACHINE TRANSLATION For language translation we trained several variants of the model in TensorFlow tutorial for English to French translation (Google). The model used word-vocabularies, 100K and 40K entries for English and French, respectively. The networks we trained had 3 or 5 layers in the encoder and decoder, each. In both cases a layer consisted of 1024 LSTM cells. SGD optimizer was used to train on WMT15 dataset. There was a noticeable variation in accuracy of different training sessions with the same settings. For example, see the three FP32 curves in Figure 4, which shows the 3-layer model. Mixed-precision with loss-scaling matched the FP32 results, while no loss-scaling resulted in a slight degradation in the results. The 5-layer model exhibited the same training behavior. 4.5 LANGUAGE MODELING We trained English language model, designated as big LSTM (Jozefowicz et al., 2016), on the 1 billion word dataset. The model consists of two layers of 8192 LSTM cells with projection to a 1024-dimensional embedding. This model was trained for 50 epochs using the Adagrad optimizer. The the vocabulary size is 793K words. During training, we use a sampled softmax layer with 8K negative samples. Batch size aggregated over 4 GPUs is 1024. To match FP32 perplexity training this network with FP16 requires loss-scaling, as shown in Figure 5. Without loss scaling the training perplexity curve for FP16 training diverges, compared with the FP32 training, after 300K iterations. Scaling factor of 128 recovers all the relevant gradient values and the accuracy of FP16 training matches the baseline run. 4.6 DCGAN RESULTS Generative Adversarial Networks (GANs) combine regression and discrimination tasks during train- ing. For image tasks, the generator network regresses pixel colors. In our case, the generator predicts three channels of 8-bit color values each. The network was trained to generate 128x128 pixel im- ages of faces, using DCGAN methodology (Radford et al., 2015) and CelebFaces dataset (Liu et al., 2015b). The generator had 7 layers of fractionally-strided convolutions, 6 with leaky ReLU activa- tions, 1 withtanh. The discriminator had 6 convolutions, and 2 fully-connected layers. All used leaky ReLU activations except for the last layer, which used sigmoid. Batch normalization was ap- plied to all layers except the last fully-connected layer of the discriminator. Adam optimizer was used to train for 100K iterations. An set of output images in Figure 6. Note that we show a randomly selected set of output images, whereas GAN publications typically show a curated set of outputs by excluding poor examples. Unlike other networks covered in this paper, GANs do not have a widely- accepted quantification of their result quality. Qualitatively the outputs of FP32 and mixed-precision training appear comparable. This network did not require loss-scaling to match FP32 results. <

> Figure 5: bigLSTM training perplexity <

> Figure 6: An uncurated set of face images generated by DCGAN. FP32 training (left) and mixed- precision training (right). 5 CONCLUSIONS AND FUTURE WORK Mixed precision training is an important technique that allows us to reduce the memory consumption as well as time spent in memory and arithmetic operations of deep neural networks. We have demonstrated that many different deep learning models can be trained using this technique with no loss in accuracy without any hyper-parameter tuning. For certain models with a large number of small gradient values, we introduce the gradient scaling method to help them converge to the same accuracy as FP32 baseline models. DNN operations benchmarked with DeepBench 1 on Volta GPU see 2-6x speedups compared to FP32 implementations if they are limited by memory or arithmetic bandwidth. Speedups are lower when operations are latency-limited. Full network training and inference speedups depend on library and framework optimizations for mixed precision and are a focus of future work (experiments in this paper were carried out with early versions of both libraries and frameworks). We would also like to extend this work to include generative models like text-to-speech systems and deep reinforcement learning applications. Furthermore, automating loss-scaling factor selection would further simplify training with mixed precision. Loss-scaling factor could be dynamically increased or decreased by inspecting the weight gradients for overflow, skipping weight updates when an overflow is detected. REFERENCES D. Amodei, R. Anubhai, E. Battenberg, C. Case, J. Casper, B. Catanzaro, J. Chen, M. Chrzanowski, A. Coates, G. Diamos, et al. Deep speech 2: End-to-end speech recognition in english and mandarin. InProceedings of The 33rd International Conference on Machine Learning, pages 173–182, 2016. K. Cho, B. Van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio.¨ Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. M. Courbariaux, Y. Bengio, and J.-P. David. Binaryconnect: Training deep neural networks with binary weights during propagations. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 3123–3131. Curran Associates, Inc., 2015. URLhttp://papers.nips.cc/paper/ 5647-binaryconnect-training-deep-neural-networks-with-binary-weights-during-propagations. pdf. R. Girshick. Faster r-cnn github repository. https://github.com/rbgirshick/ py-faster-rcnn. Google. Tensorflow tutorial: Sequence-to-sequence models. URL https://www. tensorflow.org/tutorials/seq2seq. A. Graves, S. Fernandez, F. Gomez, and J. Schmidhuber. Connectionist temporal classification:´ labelling unsegmented sequence data with recurrent neural networks. InProceedings of the 23rd international conference on Machine learning, pages 369–376. ACM, 2006. S. Gupta, A. Agrawal, K. Gopalakrishnan, and P. Narayanan. Deep learning with limited numerical precision. InProceedings of the 32nd International Conference on Machine Learning (ICML-15), pages 1737–1746, 2015. A. Hannun, C. Case, J. Casper, B. Catanzaro, G. Diamos, E. Elsen, R. Prenger, S. Satheesh, S. Sen- gupta, A. Coates, et al. Deep speech: Scaling up end-to-end speech recognition.arXiv preprint arXiv:1412.5567, 2014. K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016a. K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual networks. InECCV, 2016b. Q. He, H. Wen, S. Zhou, Y. Wu, C. Yao, X. Zhou, and Y. Zou. Effective quantization methods for recurrent neural networks.arXiv preprint arXiv:1611.10176, 2016c. S. Hochreiter and J. Schmidhuber. Long short-term memory.Neural Comput., 9(8):1735–1780, Nov. 1997. ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URLhttp://dx.doi.org/10. 1162/neco.1997.9.8.1735. I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio. Binarized neural networks. In Advances in Neural Information Processing Systems, pages 4107–4115, 2016a. I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio. Quantized neural net- works: Training neural networks with low precision weights and activations. arXiv preprint arXiv:1609.07061, 2016b. S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reduc- ing internal covariate shift. In F. R. Bach and D. M. Blei, editors,ICML, volume 37 of JMLR Workshop and Conference Proceedings, pages 448–456. JMLR.org, 2015. URLhttp: //dblp.uni-trier.de/db/conf/icml/icml2015.html#IoffeS15. Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding.arXiv preprint arXiv:1408.5093, 2014. R. Jozefowicz, O. Vinyals, M. Schuster, N. Shazeer, and Y. Wu. Exploring the limits of language modeling, 2016. URLhttps://arxiv.org/pdf/1602.02410.pdf. A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convo- lutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Wein- berger, editors, Advances in Neural Information Processing Systems 25, pages 1097– 1105. Curran Associates, Inc., 2012. URL http://papers.nips.cc/paper/ 4824-imagenet-classification-with-deep-convolutional-neural-networks. pdf. W. Liu. Ssd github repository.https://github.com/weiliu89/caffe/tree/ssd. W. Liu, D. Anguelov, D. Erhan, C. Szegedy, and S. E. Reed. Ssd: Single shot multibox detec- tor.CoRR, abs/1512.02325, 2015a. URLhttp://dblp.uni-trier.de/db/journals/ corr/corr1512.html#LiuAESR15. Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face attributes in the wild. InProceedings of International Conference on Computer Vision (ICCV), 2015b. A. Mishra, E. Nurvitadhi, J. Cook, and D. Marr. Wrpn: Wide reduced-precision networks.arXiv preprint arXiv:1709.01134, year=2017. NVIDIA. Nvidia tesla v100 gpu architecture. https://images.nvidia.com/content/ volta-architecture/pdf/Volta-Architecture-Whitepaper-v1.0.pdf, 2017. J. Ott, Z. Lin, Y. Zhang, S.-C. Liu, and Y. Bengio. Recurrent neural networks with limited numerical precision.arXiv preprint arXiv:1608.06902, 2016. A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer. Automatic differentiation in pytorch. 2017. A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolu- tional generative adversarial networks. CoRR, abs/1511.06434, 2015. URLhttp://dblp. uni-trier.de/db/journals/corr/corr1511.html#RadfordMC15. M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi.XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks, pages 525–542. Springer International Publishing, Cham, 2016. ISBN 978-3-319-46493-0. doi: 10.1007/978-3-319-46493-032. URLhttps://doi. org/10.1007/978-3-319-46493-0_32. S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. InNeural Information Processing Systems (NIPS), 2015. O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Chal- lenge.International Journal of Computer Vision (IJCV), 115(3):211–252, 2015. doi: 10.1007/ s11263-015-0816-y. K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recogni- tion.arXiv preprint arXiv:1409.1556, 2014. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Ra- binovich. Going deeper with convolutions. InComputer Vision and Pattern Recognition (CVPR), 2015. URLhttp://arxiv.org/abs/1409.4842. C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the inception architec- ture for computer vision. InThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao, K. Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation.arXiv preprint arXiv:1609.08144, 2016. S. Zhou, Z. Ni, X. Zhou, H. Wen, Y. Wu, and Y. Zou. Dorefa-net: Training low bitwidth con- volutional neural networks with low bitwidth gradients. CoRR, abs/1606.06160, 2016. URL http://arxiv.org/abs/1606.06160. <> <> <> <> <> <> Learning to Generalize SECTION VI / MODEL NEURAL NETWORKS FOR COMPUTATION AND LEARNING MANFRED OPPER Neural Computation Research Group Aston University Birmingham B4 7ET, United Kingdom Introduction Neural networks learn from examples. This statement is obviously true for the brain, but also artificial networks (or neural networks), which have become a powerful new tool for many pattern-recognition problems, adapt their synaptic couplings to a set of examples. Neural nets usually consist of many simple computing units which are combined in an architecture which is often independent from the problem. The parameters which control the interaction among the units can be changed during the learning phase and these are often called synaptic couplings. After the learning phase, a network adopts some ability to generalize from the examples; it can make predictions about inputs which it has not seen before; it has begun to understand a Theories that try to understand the ability of neural networks to generalize from learned examples are discussed. Also, an approach that is based on ideas from statistical physics which aims to model typical learning behavior is compared with a worst-case framework. rule. To what extent is it possible to understand the complexity of learning from examples by mathematical models and their solutions? This question is the focus of this article. I concentrate on the use of neural networks for classification. Here, one can take characteristic features (e.g., the pixels of an image) as an input pattern to the network. In the simplest case, it should decide whether a given pattern belongs (at least more likely) to a certain class of objects and respond with the output 1 or 1. To learn the under.lying Classification rule, the network is trained on a set of patterns together with the Classification labels, which are provided by a trainer. A heuristic strategy for training is to tune the parameters of the machine (the couplings of the network) using a learning algorithm, in such a way that the errors made on the set of training examples are small, in the hope that this helps to reduce the errors on new data. How well will the trained network be able to classify an in. put that it has not seen before? This performance on new data defines the generalization ability of the network. This ability will be affected by the problem of realizability: The network may not be sufficiently complex to learn the rule completely or there may be ambiguities in Classification. Here, I concentrate on a second problem arising from the fact that learning will mostly not be exhaustive and the in.formation about the rule contained in the examples is not complete. Hence, the performance of a network may vary from one training set to another. In order to treat the generalization ability in a quantitative way, a common model assumes that all input patterns, those from the training set and the new one on which the network is tested, have a pre.assigned probability distribution (which characterizes the feature that must be classified), and they are produced in.dependently at random with the same probability distribution from the network's environment. Sometimes the probability distribution used to extract the examples and the Classification of these examples is called the rule. The network's performance on novel data can now be quantified by the so-called generalization error, which is the probability of misclassifying the test input and can be measured by repeating the same learning experiment many times with different data. Within such a probabilistic framework, neural networks are often viewed as statistical adaptive models which should give a likely explanation of the observed data. In this frame.work, the learning process becomes mathematically related to a statistical estimation problem for optimal network parameters. Hence, mathematical statistics seems to be a most appropriate candidate for studying a neural network's behavior. In fact, various statistical approaches have been ap.plied to quantify the generalization performance. For ex.ample, expressions for the generalization error have been obtained in the limit, where the number of examples is large compared to the number of couplings (Seung et al., 1992; for the case of realizable rules they are also independent of the specific algorithm, as long as the training examples are perfectly learned. Because it is able to cover even bad situations which are unfavorable for improvement of the learning process, it is not surprising that this theory may in some cases provide too pessimistic results which are also too crude to reveal interesting behavior in the intermediate region of the learning curve. In this article, I concentrate mainly on a different approach, which has its origin in statistical physics rather than in mathematical statistics, and compare its results with the worst-case results. This method aims at studying the typical rather than the worst-case behavior and often enables the exact calculations of the entire learning curve for models of simple networks which have many parameters. Since both biological and artificial neural networks are composed of many elements, it is hoped that such an approach may actually reveal some relevant and interesting structures. At first, it may seem surprising that a problem should simplify when the number of its constituents becomes large. However, this phenomenon is well-known for macroscopic physical systems such as gases or liquids which consist of a huge number of molecules. Clearly, it is not possible to study the complete microscopic state of such a system, which is described by the rapidly fluctuating positions and velocities of all particles. On the other hand, macroscopic quantities such as density, temperature, and pressure are usually collective properties influenced by all elements. For such quantities, fluctuations are averaged out in the thermodynamic limit of a large number of particles and the collective properties become, to some extent, independent of the microstate. Similarly, the generalization ability of a neu.ral network is a collective property of all the network parameters, and the techniques of statistical physics allow, at least for some simple but nontrivial models, for exact computations in the thermodynamic limit. Before explaining these ideas in detail, I provide a short description of feed-forward neural networks. Amari and Murata, 1993). In such a case, one can expect that learning is almost exhaustive, such that the statistical fluctuations of the parameters around their optimal values are small. However, in practice the number of parameters is artificial Neural Networks often large so that the network can be flexible, and it is not clear how many examples are needed for the asymptotic theory to become valid. The asymptotic theory may actually miss interesting behavior of the so-called learning curve, which displays the progress of generalization ability with an increasing amount of training data. A second important approach, which was introduced into mathematical statistics in the 1970s by Vapnik and Chervonenkis (VC) (Vapnik, 1982, 1995), provides exact bounds for the generalization error which are valid for any number of training examples. Moreover, they are entirely independent of the underlying distribution of inputs, and Based on highly idealized models of brain function, artificial neural networks are built from simple elementary computing units, which are sometimes termed neurons after their biological counterparts. Although hardware implementations have become an important research topic, neu.ral nets are still simulated mostly on standard computers. Each computing unit of a neural net has a single output and several ingoing connections which receive the outputs of other units. To every ingoing connection (labeled by the index i) a real number is assigned, the synaptic weight w_i, which is the basic adjustable parameter of the network. To compute a unit's output, all incoming values xi are multi.plied by the weights wi and then added. Figure 1a shows an example of such a computation with three couplings. Finally, the result, <>, is passed through an activation function which is typically of the shape of the red curve in Fig. 1a (a sigmoidal function), which allows for a soft, ambiguous Classification between 1 and 1. Other important cases are the step function (green curve) and the linear function (yellow curve; used in the output neuron for problems of fitting continuous functions). In the following, to keep matters simple, I restrict the discussion mainly to the step function. Such simple units can develop a remarkable computational power when connected in a suitable architectures. An important network type is the feedforward architecture shown in Fig. 1b, which has two layers of computing units and adjustable couplings. The input nodes (which do not compute) are coupled to the so-called hidden units, which feed their outputs into one or more output units. With such an architecture and sigmoidal activation functions, any continuous function of the inputs can be arbitrarily closely approximated when the number of hidden units is sufficiently large. <

> FIGURE 1 (a) Example of the computation of an elementary unit (neuron) in a neural network. The numerical values assumed by the incoming inputs to the neuron and the weights of the synapses by which the inputs reach the neuron are indicated. The weighted sum of the inputs corresponds to the value of the abscissa at which the value of the activation function is calculated (bottom graph). Three functions are shown: sigmoid, linear, and step. (b) Scheme of a feedforward network. The arrow indicates the direction of propagation of information. LEARNING TO GENERALIZE The Perceptron The simplest type of network is the perceptron (Fig. 2a). There are N inputs, N synaptic couplings <>, and the output is simply <> It has a single-layer architecture and the step function (green curve in Fig. 1a) as its activation function. Despite . its simple structure, it can for many learning problems give a nontrivial generalization performance and may be used as a first step to an unknown Classification task. As can be seen by comparing Figs. 2a and 1b, it is also a building block for the more complex multilayer networks. Hence, understanding its performance theoretically may also pro.vide insight into the more complex machines. To learn a set of examples, a network must adjust its couplings appropriately (I often use the word couplings for their numerical strengths, the weights <>, for <>). Remarkably, for the perceptron there exists a simple learning algorithm which always enables the network to find those parameter values whenever the examples can be learnt by a perceptron. In Rosenblatt's algorithm, the input patterns are presented sequentially (e.g., in cycles) to the network and the <

> FIGURE 2 (a) The perceptron. (b) Classification of inputs by a perceptron with two inputs. The arrow indicates the vector composed of the weights of the network, and the line per.pendicular to this vector is the boundary between the classes of input. output is tested. Whenever a pattern is not classified correctly, all couplings are altered simultaneously. We increase by a fixed amount all weights for which the input unit and the correct value of the output neuron have the same sign but we decrease them for the opposite sign. This simple algorithm is reminiscent of the so-called Hebbian learning rule, a physiological model of a learning processes in the real brain. It assumes that synaptic weights are increased when two neurons are simultaneously active. Rosenblatt's theorem states that in cases in which there exists a choice of the wi which classify correctly all of the examples (i.e., perfectly learnable perceptron), this algorithm finds a solution in a finite number of steps, which is at worst equal to A N3, where A is an appropriate constant. It is often useful to obtain an intuition of a perceptron's Classification performance by thinking in terms of a geo.metric picture. We may view the numerical values of the in.puts as the coordinates of a point in some (usually) high-dimensional space. The case of two dimensions is shown in Fig. 2b. A corresponding point is also constructed for the couplings wi. The arrow which points from the origin of the coordinate system to this latter point is called the weight vector or coupling vector. An application of linear algebra to the computation of the network shows that the line which is perpendicular to the coupling vector is the boundary be.tween inputs belonging to the two different classes. Input points which are on the same side as the coupling vector are classified as 1 (the green region in Fig. 2b) and those on the other side as 1 (red region in Fig. 2b). Rosenblatt's algorithm aims to determine such a line when it is possible. This picture generalizes to higher dimensions, for which a hyperplane plays the same role of the line of the previous two-dimensional example. We can still obtain an intuitive picture by projecting on two-dimensional planes. In Fig. 3a, 200 input patterns with random coordinates (randomly labeled red and blue) in a 200-dimensional input space are projected on the plane spanned by two arbitrary coordinate axes. If we instead use a plane for projection which contains the coupling vector (determined from a variant of Rosenblatt's algorithm) we obtain the view shown in Fig. 3b, in which red and green points are clearly separated and there is even a gap between the two clouds. It is evident that there are cases in which the two sets of points are too mixed and there is no line in two dimensions (or no hyperplane in higher dimensions which separates them). In these cases, the rule is too complex to be perfectly learned by a perceptron. If this happens, we must attempt to determine the choice of the coupling which minimizes the number of errors on a given set of examples. Here, Rosenblatt's algorithm does not work and the problem of finding the minimum is much more difficult from the algorithmic point. The training error, which is the number of errors made on the training set, is usually a non-smooth function of the network couplings (i.e., it may have large variations for small changes of the couplings). Hence, in general, in addition to the perfectly learnable perceptron case in which the final error is zero, minimizing the training error is usually a difficult task which could take a large amount of computer time. However, in practice, iterative approaches, which are based on the minimization of other smooth cost functions, are used to train a neural network (Bishop, 1995). As previously shown, perceptrons are only able to realize a very restricted type of Classification rules, the so-called linearly separable ones. Hence, independently from the issue of finding the best algorithm to learn the rule, one may ask the following question: In how many cases will the perceptron be able to learn a given set of training examples perfectly if the output labels are chosen arbitrarily? In order to answer this question in a quantitative way, it is convenient to introduce some concepts such as capacity, VC dimension and Worst-Case Generalization. <

> FIGURE 3 (a) Projection of 200 random points (with ran.dom labels) from a 200-dimensional space onto the first two coordinate axes (x1 and x2). (b) Projection of the same points onto a plane which contains the coupling vector of a perfectly trained perceptron. LEARNING TO GENERALIZE <>, where the function <> vanishes for a 2 and it is positive for a 2. Such a threshold phenomenon is an example of a phase transition (i.e., a sharp change of behavior) which can occur in the thermodynamic limit of a large network size. and worst-case generalization, which can be used in the case of the perceptron and have a more general meaning. In the case of perceptrons, this question was answered in the 1960s by Cover (1965). He calculated for any set of in.put patterns, e.g., m, the fraction of all the 2m possible map.pings that can be linearly separated and are thus learnable by perceptrons. This fraction is shown in Fig. 4 as a function of the number of examples per coupling for different numbers of input nodes (couplings) N. Three regions can be distinguished: Region in which m/N 1: Simple linear algebra shows that it is always possible to learn all mappings when the number m of input patterns is less than or equal to the number N of couplings (there are simply enough adjustable parameters). Region in which m/N 1: For this region, there are examples of rules that cannot be learned. However, when the number of examples is less than twice the number of couplings (m/N 2), if the network is large enough almost all mappings can be learned. If the output labels for each of the m inputs are chosen randomly 1 or 1 with equal probability, the probability of finding a nonrealizable coupling goes to zero exponentially when N goes to infinity at fixed ratio m/N. Region in which m/N 2: For m/N 2 the probability for a mapping to be realizable by perceptrons decreases to zero rapidly and it goes to zero exponentially when N goes to infinity at fixed ratio m/N (it is proportional to <

> FIGURE 4 Fraction of all mappings of m input patterns which are learnable by perceptrons as a function of m/N for different numbers of couplings N: N 10 (in green), N 20 (in blue), and N 100 (in red). fraction of realizable mappings Generally, the point at which such a transition takes place defines the so-called capacity of the neural network. Although the capacity measures the ability of a network to learn random mappings of the inputs, it is also related to its ability to learn a rule (i.e., to generalize from examples). The question now is, how does the network perform on a new example after having been trained to learn m example on the training set? To obtain an intuitive idea of the connection between capacity and ability to generalize, we assume a training set of size m and a single pattern for test. Suppose we define a possible rule by an arbitrary learnable mapping from inputs to outputs. If m 1 is much larger than the capacity, then for most rules the labels on the m training pat.terns which the perceptron is able to recognize will nearly uniquely determine the couplings (and consequently the answer of the learning algorithm on the test pattern), and the rule can be perfectly understood from the examples. Be.low capacity, in most cases there are two different choices of couplings which give opposite answers for the test pat.tern. Hence, a correct Classification will occur with probability 0.5 assuming all rules to be equally probable. Figure 5 displays the two types of situations for m^3 and N^2. This intuitive connection can be sharpened. Vapnik and Chervonenkis established a relation between a capacity such as quantity and the generalization ability that is valid for general classifiers (Vapnik, 1982, 1995). The VC dimension is defined as the size of the largest set of inputs for which all mappings can be learned by the type of classifier. It equals N for the perceptron. Vapnik and Chervonenkis were able to show that for any training set of size m <

> FIGURE 5 Classification rules for four patterns based on a perceptron. The patterns colored in red represent the training examples, and triangles and circles represent different class la.bels. The question mark is a test pattern. (a) There are two possible ways of classifying the test point consistent with the examples; (b) only one Classification is possible. larger than the VC dimension DVC, the growth of the number of realizable mappings is bounded by an expression which grows much slower than 2m (in fact, only like a polynomial in m). They proved that a large difference between training er.ror (i.e., the minimum percentage of errors that is done on the training set) and generalization error (i.e., the probability of producing an error on the test pattern after having learned the examples) of classifiers is highly improbable if the number of examples is well above DVC. This theorem implies a small expected generalization error for perfect learning of the training set results. The expected generalization error is bounded by a quantity which increases proportionally to DVC and decreases (neglecting logarithmic corrections in m) inversely proportional to m. than DVC is also necessary for good generalization. The VC results should, in practice, enable us to select the network with the proper complexity which guarantees the smallest bound on the generalization error. For example, in order to find the proper size of the hidden layer of a network with two layers, one could train networks of different sizes on the same data. The relation among these concepts can be better under.stood if we consider a family of networks of increasing complexity which have to learn the same rule. A qualitative picture of the results is shown in Fig. 6. As indicated by the blue curve in Fig. 6, the minimal training error will decrease for increasing complexity of the nets. On the other hand, the VC dimension and the complexity of the networks in.crease with the increasing number of hidden units, leading to an increasing expected difference (confidence interval) between training error and generalization error as indicated by the red curve. The sum of both (green curve) will have a minimum, giving the smallest bound on the generalization error. As discussed later, this procedure will in some cases lead to not very realistic estimates by the rather pessimistic bounds of the theory. In other words, the rigorous bounds, which are obtained from an arbitrary network and rule, are much larger than those determined from the results for most of the networks and rules. Conversely, one can construct a worst-case distribution Typical Scenario: The Approach of input patterns, for which a size of the training set larger of Statistical Physics When the number of examples is comparable to the size of the network, which for a perceptron equals the VC dimension, the VC theory states that one can construct malicious situations which prevent generalizations. However, in gen.eral, we would not expect that the world acts as an adver.sary. Therefore, how should one model a typical situation? As a first step, one may construct rules and pattern dis.tributions which act together in a nonadversarial way. The teacher�student paradigm has proven to be useful in such a situation. Here, the rule to be learned is modeled by a sec.ond network, the teacher network; in this case, if the teacher and the student have the same architecture and the same <

> FIGURE 6 As the complexity of the network varies (i.e., of the number of hidden units, as shown schematically below), the generalization error (in red), calculated from the sum of the training error (in green) and the confidence interval (in blue) according to the theory of Vapnik Chervonenkis, shows a minimum; this corresponds to the network with the best generalization ability. number of units, the rule is evidently realizable. The correct class labels for any inputs are given by the outputs of the teacher. Within this framework, it is often possible to ob.tain simple expressions for the generalization error. For a perceptron, we can use the geometric picture to visualize the generalization error. A misClassification of a new in.put vector by a student perceptron with coupling vector ST occurs only if the input pattern is between the separating planes (dashed region in Fig. 7) defined by ST and the vector of teacher couplings TE. If the inputs are drawn randomly from a uniform distribution, the generalization error is directly proportional to the angle between ST and TE. Hence, the generalization error is small when teacher and student vectors are close together and decreases to zero when both coincide. In the limit, when the number of examples is very large all the students which learn the training examples perfectly will not differ very much from and their couplings will be close to those of the teacher. Such cases with a small generalization error have been successfully treated by asymptotic methods of statistics. On the other hand, when the number of examples is relatively small, there are many different students which are consistent with the teacher regarding the training examples, and the uncertainty about LEARNING TO GENERALIZE <

> FIGURE 7 For a uniform distribution of patterns, the generalization error of a perceptron equals the area of the shaded region divided by the area of the entire circle. ST and TE represent the coupling vectors of the student and teacher, respectively. the true couplings of the teacher is large. Possible generalization errors may range from zero (if, by chance, a learning algorithm converges to the teacher) to some worst-case value. We may say that the constraint which specifies the macrostate of the network (its training error) does not spec.ify the microstate uniquely. Nevertheless, it makes sense to speak of a typical value for the generalization error, which is defined as the value which is realized by the majority of the students. In the thermodynamic limit known from statistical physics, in which the number of parameters of the network is taken to be large, we expect that in fact almost all students belong to this majority, provided the quantity of interest is a cooperative effect of all components of the system. As the geometric visualization for the generalization error of the perceptron shows, this is actually the case. The following approach, which was pioneered by Elizabeth Gardner (Gardner, 1988; Gardner and Derrida, 1989), is based on the calculation of V(e), the volume of the space of couplings which both perfectly implement m training examples and have a given generalization error e. For an intuitive picture, consider that only discrete values for the couplings are allowed; then <> would be proportional to the number of students. The typical value of the generalization error is the value of e, which maximizes V(e). It should be kept in mind that V(e) is a random number and fluctuates from one training set to another. A correct treatment of this randomness requires involved mathematical techniques (Mzard et al., 1987). To obtain a picture which is quite often qualitatively correct, we may replace it by its average over many realizations of training sets. From elementary probability theory we see that this average number can be found by calculating the volume A of the space of all students with generalization error e, irrespective of their behavior on the training set, and multiplying it by the probability B that a student with generalization error e gives m times the correct answers on independent drawings of the input patterns. Since A increases exponentially with the number of couplings N (like typical volumes in N-dimensional spaces) and B decreases exponentially with m (because it becomes more improbable to be correct m times for any e 0), both factors can balance each other when m increases like m aN. a is an effective measure for the size of the training set when N goes to infinity. In order to have quantities which remain finite as N Sq, it is also useful to take the logarithm of V(e) and divide by N, which transforms the product into a sum of two terms. The first one (which is often called the entropic term) increases with increasing generalization error (green curve in Fig. 8). This is true because there are many networks which are not similar to the teacher, but there is only one network equal to the teacher. For almost all networks (remember, the entropic term does not include the effect of the training examples) e 0.5, i.e., they are correct half of the time by random guessing. On the other hand, the second term (red curve in Fig. 8) decreases with increasing generalization er.ror because the probability of being correct on an input pattern increases when the student network becomes more similar to the teacher. It is often called the energetic contribution because it favors highly ordered (toward the teacher) network states, reminiscent of the states of physical systems at low energies. Hence, there will be a maximum (Fig. 8, ar.row) of <> at some value of e which by definition is the typical generalization error. The development of the learning process as the number of examples aN increases can be understood as a competition between the entropic term, which favors disordered network configurations that are not similar to the teacher, and the energetic term. The latter term dominates when the number of examples is large. It will later be shown that such a competition can lead to a rich and interesting behavior as the number of examples is varied. The result for the learning curve (Gyrgyi and Tishby, 1990; Sompolinsky et al., FIGURE 8 Logarithm of the average volume of students that have learned m examples and give e generalization error (green curve). The blue and red curves represent the energetic and entropic contributions, respectively. <

> student is free to ask the teacher questions, i.e., if the stu.dent can choose highly informative input patterns. For the simple perceptron a fruitful query strategy is to select a new input vector which is perpendicular to the current coupling vector of the student (Kinzel and Ruj�n, 1990). Such an input is a highly ambiguous pattern because small changes in the student couplings produce different Classification answers. For more complicated networks it may be difficult to obtain similar ambiguous inputs by an explicit construction. A general algorithm has been proposed (Seung et al., 1992a) which uses the principle of maximal disagreement in a committee of several students as a selection process for training patterns. Using an appropriate randomized train.ing strategy, different students are generated which all learn the same set of examples. Next, any new input vector is only <

> FIGURE 9 Learning curves for typical student perceptrons. a m/N is the ratio between the number of examples and the coupling number. 1990) of a perceptron obtained by the statistical physics approach (treating the random sampling the proper way) is shown by the red curve of Fig. 9. In contrast to the worst-case predictions of the VC theory, it is possible to have some generalization ability below VC dimension or capacity. As we might have expected, the generalization error decreases monotonically, showing that the more that is learned, the more that is understood. Asymptotically, the error is pro-accepted for training when the disagreement of its classification between the students is maximal. For a committee of two students it can be shown that when the number of examples is large, the information gain does not decrease but reaches a positive constant. This results in a much faster decrease of the generalization error. Instead of being in.versely proportional to the number of examples, the de.crease is now exponentially fast. portional to N and inversely proportional to m, in agree-monotonically decreasing learning curve, the possibility ment with the VC predictions. This may not be true for that some concrete learning algorithm may result in a set more complicated networks. of student couplings which are untypical in the sense of our theory cannot be ruled out. For bad students, even non-monotic generalization behavior is possible. The problem Bad Students and Good Students Although the typical student perceptron has a smooth, Query Learning Soon after Gardner's pioneering work, it was realized that the approach of statistical physics is closely related to ideas in information theory and Bayesian statistics (Levin et al., 1989; Gyfirgyi and Tishby, 1990; Opper and Haussler, 1991), for which the reduction of an initial uncertainty about the true state of a system (teacher) by observing data is a cen.tral topic of interest. The logarithm of the volume of rele.vant microstates as defined in the previous section is a di.rect measure for such uncertainty. The moderate progress in generalization ability displayed by the red learning curve of Fig. 9 can be understood by the fact that as learning progresses less information about the teacher is gained from a new random example. Here, the information gain is defined as the reduction of the uncertainty when a new example is learned. The decrease in information gain is due to the in.crease in the generalization performance. This is plausible because inputs for which the majority of student networks give the correct answer are less informative than those for which a mistake is more likely. The situation changes if the of a concrete learning algorithm can be made to fit into the statistical physics framework if the algorithm minimizes a certain cost function. Treating the achieved values of the new cost function as a macroscopic constraint, the tools of statistical physics apply again. As an example, it is convenient to consider a case in which the teacher and the student have a different architectures: In one of the simplest examples one tries to learn a Classification problem by interpreting it as a regression problem, i.e., a problem of fitting a continuous function through data points. To be specific, we study the situation in which the teacher network is still given by a perceptron which computes binary valued outputs of the form y i wixi , 1, but as the student we choose a network with a linear transfer function (the yellow curve in Fig. 1a) <> and try to fit this linear expression to the binary labels of the teacher. If the number of couplings is sufficiently large (larger than the number of examples) the linear function LEARNING TO GENERALIZE (unlike the sign) is perfectly able to fit arbitrary continuous output values. This linear fit is an attempt to explain the data in a more complicated way than necessary, and the couplings have to be �nely tuned in order to achieve this goal. We find that the student trained in such a way does not generalize well (Opper and Kinzel, 1995). In order to compare the Classifications of teacher and student on a new random input after training, we have finally converted the student�s output into a Classification label by taking the sign of its output. As shown in the red curve of Fig. 10, after an initial improvement of performance the generalization error increases again to the random guessing value e 0.5 at a 1 (Fig. 10, red curve). This phenomenon is called overfitting. For a 1 (i.e., for more data than parameters), it is no longer possible to have a perfect linear fit through the data, but a fit with a minimal deviation from a linear function leads to the second part of the learning curve. e de.creases again and approaches 0 asymptotically for a Sq. This shows that when enough data are available, the details of the training algorithm are less important. The dependence of the generalization performance on the complexity of the assumed data model is well-known. If function class is used that is too complex, data values can be perfectly fitted but the predicted function will be very sen.sitive to the variations of the data sample, leading to very unreliable predictions on novel inputs. On the other hand, functions that are too simple make the best fit almost insen.sitive to the data, which prevents us from learning enough from them. It is also possible to calculate the worst-case generalization ability of perceptron students learning from a perceptron teacher. The largest generalization error is obtained (Fig. 7) when the angle between the coupling vectors of teacher and student is maximized under the constraint that the student learns all examples perfectly. Although it may not be easy to construct a learning algorithm which per.forms such a maximization in practice, the resulting gener.alization error can be calculated using the statistical phys.ics approach (Engel and Van den Broeck, 1993). The result is in agreement with the VC theory: There is no prediction better than random guessing below the capacity. Although the previous algorithms led to a behavior which is worse than the typical one, we now examine the op.posite case of an algorithm which does better. Since the generalization ability of a neural network is related to the fact that similar input vectors are mapped onto the same out.put, one can assume that such a property can be enhanced if the separating gap between the two classes is maximized, which defines a new cost function for an algorithm. This optimal margin perceptron can be practically realized and when applied to a set of data leads to the projection of Fig. 11. As a remarkable result, it can be seen that there is a relatively large fraction of patterns which are located at the gap. These points are called support vectors (SVs). In order to understand their importance for the generalization abil.ity, we make the following gedankenexperiment and assume that all the points which lie outside the gap (the nonsupport vectors) are eliminated from the training set of examples. From the two-dimensional projection of Fig. 11, we may conjecture that by running the maximal margin algorithm on the remaining examples (the SVs) we cannot create a larger gap between the points. Hence, the algorithm will converge to the same separating hyperplane as before. This intuitive picture is actually correct. If the SVs of a training set were known beforehand (unfortunately, they are only identi�ed after running the algorithm), the margin classifier would have to be trained only on the SVs. It would au.tomatically classify the rest of the training inputs correctly. FIGURE 11 Learning with a margin classifier and m 300 examples in an N 150-dimensional space. Bias/Variance trade-off Hence, if in an actual Classification experiment the number of SVs is small compared to the number of non-SVs, we may expect a good generalization ability. The learning curve for a margin classifier (Opper and Kinzel, 1995) learning from a perceptron teacher (calculated by the statistical physics approach) is shown in Fig. 10 (blue curve). The concept of a margin classifier has recently ber of consistent students is small; nevertheless, the few re.maining ones must still differ in a finite fraction of bits from each other and from the teacher so that perfect generalization is still impossible. For a slightly above ac only the couplings of the teacher survive. been generalized to the so-called support vector machines (Vapnik, 1995), for which the inputs of a perceptron are re.placed by suitable features which are cleverly chosen nonlinear functions of the original inputs. In this way, nonlinear separable rules can be learned, providing an interesting alternative to multilayer networks. Learning with Errors The example of the Ising perceptron teaches us that it will not always be simple to obtain zero training error. Moreover, an algorithm trying to achieve this goal may get stuck in local minima. Hence, the idea of allowing errors explic. itly in the learning procedure, by introducing an appropriate noise, can make sense. An early analysis of such a sto- The Ising Perceptron The approach of statistical physics can develop a specific predictive power in situations in which one would like to un.derstand novel network models or architectures for which currently no ef�cient learning algorithm is known. As the simplest example, we consider a perceptron for which the couplings wj are constrained to binary values 1 and 1 (Gardner and Derrida, 1989; Gy�rgyi, 1990; Seung et al., 1992b). For this so-called Ising perceptron (named after Ernst Ising, who studied coupled binary-valued elements as a model for a ferromagnet), perfect learning of examples is equivalent to a difficult combinatorial optimization prob.lem (integer linear programming), which in the worst case is believed to require a learning time that increases expo.nentially with the number of couplings N. To obtain the learning curve for the typical student, we can proceed as before, replacing V(e) by the number of student configurations that are consistent with the teacher which results in changing the entropic term appropriately. When the examples are provided by a teacher network of the same binary type, one can expect that the generalization error will decrease monotonically to zero as a function of a. The learning curve is shown as the blue curve in Fig. 9. For sufficiently small a, the discreteness of the couplings has al.most no effect. However, in contrast to the continuous case, perfect generalization does not require infinitely many examples but is achieved already at a finite number ac 1.24. This is not surprising because the teacher�s couplings con.tain only a finite amount of information (one bit per coupling) and one would expect that it does not take much more than about N examples to learn them. The remark.able and unexpected result of the analysis is the fact that the transition to perfect generalization is discontinuous. The generalization error decreases immediately from a non.zero value to zero. This gives an impression about the com.plex structure of the space of all consistent students and also gives a hint as to why perfect learning in the Ising per.ceptron is a difficult task. For a slightly below ac, the num.chastic training procedure and its generalization ability for the learning in so-called Boolean networks (with elemen.tary computing units different from the ones used in neural networks) can be found in Carnevali and Patarnello (1987). A stochastic algorithm can be useful to escape local min.ima of the training error, enabling a better learning of the training set. Surprisingly, such a method can also lead to better generalization abilities if the Classification rule is also corrupted by some degree of noise (Gy�rgyi and Tishby, 1990). A stochastic training algorithm can be realized by the Monte Carlo metropolis method, which was invented to generate the effects of temperature in simulations of physical systems. Any changes of the network couplings which lead to a decrease of the training error during learning are allowed. However, with some probability that in.creases with the temperature, an increase of the training error is also accepted. Although in principle this algorithm may visit all the network's configurations, for a large sys.tem, with an overwhelming probability, only states close to some fixed training error will actually appear. The method of statistical physics applied to this situation shows that for sufficiently large temperatures (T) we often obtain a quali.tatively correct picture if we repeat the approximate calcu.lation for the noise-free case and replace the relative number of examples a by the effective number a/T. Hence, the learning curves become essentially stretched and good generalization ability is still possible at the price of an increase in necessary training examples. Within the stochastic framework, learning (with errors) can now also be realized for the Ising perceptron, and it is interesting to study the number of relevant student con�gu.rations as a function of e in more detail (Fig. 12). The green curve is obtained for a small value of a where a strong maxi.mum with high generalization error exists. By increasing a, this maximum decreases until it is the same as the second maximum at e 0.5, indicating a transition like that of the blue learning curve in Fig. 9. For larger a, the state of per.fect generalization should be the typical state. Neverthe.less, if the stochastic algorithm starts with an initial state <

> FIGURE 12 Logarithm of the number of relevant Ising stu.dents for different values of a. which has no resemblance to the (unknown) teacher (i.e., with e 0.5), it will spend time that increases exponentially with N in the smaller local maximum, the metastable state. Hence, a sudden transition to perfect generalization will be observable only in examples which correspond to the blue curve of Fig. 12, where this metastable state disappears. For large vales of a (yellow curve), the stochastic algorithm will converge always to the state of perfect generalization. On the other hand, since the state with e 0.5 is always metastable, a stochastic algorithm which starts with the teacher�s couplings will never drive the student out of the state of perfect generalization. It should be made clear that the sharp phase transitions are the result of the thermody.namic limit, where the macroscopic state is entirely domi.nated by the typical configurations. For simulations of any finite system a rounding and softening of the transitions will be observed. More Sophisticated Computations Are Needed for Multilayer Networks As a first step to understand the generalization perfor.mance of multilayer networks, one can study an architectures which is simpler than the fully connected one of Fig. 1b. The tree architecture of Fig. 13 has become a popular model. Here, each hidden unit is connected to a different set of the input nodes. A further simpli�cation is the replacement of adaptive couplings from the hidden units to the output node by a prewired fixed function which maps the states of the hidden units to the output. Two such functions have been studied in great detail. For the first one, the output gives just the majority vote of the hidden unitsfithatis, if themajority of the hidden units is negative, then the total output is negative, and vice versa. This network is called a committee machine. For the second type of network, the parity machine, the output is the par.ity of the hidden outputsfithat is, a minus results from an odd number of negative hidden units and a plus from an even number. For both types of networks, the capacity has been calculated in the thermodynamic limit of a large number N of (first layer) couplings (Barkai et al., 1990; Monas.son and Zecchina, 1995). By increasing the number of hid.den units (but always keeping it much smaller than N), the capacity per coupling (and the VC dimension) can be made arbitrarily large. Hence, the VC theory predicts that the ability to generalize begins at a size of the training set which increases with the capacity. The learning curves of the typical parity machine (Fig. 14) being trained by a par.ity teacher for (from left to right) one, two, four, and six hidden units seem to partially support this prediction. Below a certain number of examples, only memorization of the learned patterns occurs and not generalization. Then, a transition to nontrivial generalization takes place (Han.sel et al., 1992; Opper, 1994). Far beyond the transition, the decay of the learning curves becomes that of a simple per.ceptron (black curve in Fig. 14) independent of the number of hidden units, and this occurs much faster than for the bound given by VC theory. This shows that the typical learning curve can in fact be determined by more than one <

> complexity parameter. In contrast, the learning curve of the committee machine with the tree architecture of Fig. 13 (Schwarze and Hertz, 1992) is smooth and resembles that of the simple perceptron. As the number of hidden units is increased (keeping N fixed and very large), the generalization error increases, but despite the diverging VC di.mension the curves converge to a limiting one having an asymptotic decay which is only twice as slow as that of the perceptron. This is an example for which typical and worst-case generalization behaviors are entirely different. Recently, more light has been shed on the relation be.tween average and worst-case scenarios of the tree com-the same similarity to every teacher perceptron. Although this symmetric state allows for some degree of generalization, it is not able to recover the teacher�s rule completely. After a long plateau, the symmetry is broken and each of the student perceptrons specializes to one of the teacher perceptrons, and thus their similarity with the others is lost. This leads to a rapid (but continuous) decrease in the generalization error. Such types of learning curves with plateaus can actually be observed in applications of fully connected multilayer networks. Outlook mittee. A reduced worst-case scenario, in which a tree committee teacher was to be learned from tree committee students under an input distribution, has been analyzed from a statistical physics perspective (Urbanczik, 1996). As expected, few students show a much worse generalization ability than the typical one. Moreover, such students may also be difficult to find by most reasonable learning algorithms because bad students require very �ne tuning of their couplings. Calculation of the couplings with finite pre.cision requires many bits per coupling that increases faster than exponentially with a and which for sufficiently large a will be beyond the capability of practical algorithms. Hence, it is expected that, in practice, a bad behavior will not be observed. Transitions of the generalization error such as those observed for the tree parity machine are a characteristic feature of large systems which have a symmetry that can be spontaneously broken. To explain this, consider the sim.plest case of two hidden units. The output of this parity ma.chine does not change if we simultaneously change the sign of all the couplings for both hidden units. Hence, if the teacher�s couplings are all equal to 1, a student with all couplings equal to 1 acts exactly as the same classifier. If there are few examples in the training set, the entropic contribution will dominate the typical behavior and the typical students will display the same symmetry. Their coupling vectors will consist of positive and negative random numbers. Hence, there is no preference for the teacher or the reversed one and generalization is not possible. If the number of examples is large enough, the symmetry is broken and there are two possible types of typical students, one with more positive and the other one with more negative couplings. Hence, any of the typical students will show some similarity with the teacher (or it's negative image) and generalization occurs. A similar type of symmetry break.ing also leads to a continuous phase transition in the fully connected committee machine. This can be viewed as a committee of perceptrons, one for each hidden unit, which share the same input nodes. Any permutation of these perceptrons obviously leaves the output invariant. Again, if few examples are learned, the typical state reflects the symmetry. Each student perceptron will show approximately The worst-case approach of the VC theory and the typical case approach of statistical physics are important theories for modeling and understanding the complexity of learning to generalize from examples. Although the VC approach plays an important role in a general theory of learnability, its practical applications for neural networks have been limited by the overall generality of the approach. Since only weak assumptions about probability distributions and machines are considered by the theory, the estimates for generalization errors have often been too pessimistic. Recent developments of the theory seem to overcome these problems. By using modified VC dimensions, which depend on the data that have actually occurred and which in favorable cases are much smaller than the general dimensions, more realistic results seem to be possible. For the support vector machines (Vapnik, 1995) (generalizations of the margin classifiers which allow for nonlinear boundaries that separate the two classes), Vapnik and collaborators have shown the effectiveness of the modified VC results for selecting the optimal type of model in practical applications. The statistical physics approach, on the other hand, has revealed new and unexpected behavior of simple network models, such as a variety of phase transitions. Whether such transitions play a cognitive role in animal or human brains is an exciting topic. Recent developments of the theory aim to understand dynamical problems of learning. For ex.ample, online learning (Saad, 1998), in which the problems of learning and generalization are strongly mixed, has en.abled the study of complex multilayer networks and has stimulated research on the development of optimized algorithms. In addition to an extension of the approach to more complicated networks, an understanding of the robustness of the typical behavior, and an interpolation to the other extreme, the worst-case scenario is an important subject of research. Acknowledgments I thank members of the Department of Physics of Complex Sys.tems at the Weizmann Institute in Rehovot, Israel, where parts of this article were written, for their warm hospitality. References Cited AMARI, S., and MURATA, N. (1993). Statistical theory of learning curves under entropic loss. Neural Comput. 5, 140. BARKAI, E., HANSEL, D., and KANTER, I. (1990). Statistical me.chanics of a multilayered neural network. Phys. Rev. Lett. 65, 2312. BISHOP, C. M. (1995). Neural Networks for Pattern Recognition. Clarendon/Oxford Univ. Press, Oxford/New York. CARNEVALI, P., and PATARNELLO, S. (1987). Exhaustive thermo.dynamical analysis of Boolean learning networks. Europhys. Lett. 4, 1199. COVER, T. M. (1965). Geometrical and statistical properties of systems of linear inequalities with applications in pattern rec.ognition. IEEE Trans. El. Comp. 14, 326. ENGEL, A., and VAN DEN BROECK, C. (1993). Systems that can learn from examples: Replica calculation of uniform conver.gence bound for the perceptron. Phys. Rev. Lett. 71, 1772. GARDNER, E. (1988). The space of interactions in neural networks. J. Phys. A 21, 257. GARDNER, E., and DERRIDA, B. (1989). Optimal storage proper.ties of neural network models. J. Phys. A 21, 271. GY�RGYI, G. (1990). First order transition to perfect generalization in a neural network with binary synapses. Phys. Rev. A 41, 7097. GY�RGYI, G., and TISHBY, N. (1990). Statistical theory of learning a rule. In Neural Networks and Spin Glasses: Proceedings of the STATPHYS 17 Workshop on Neural Networks and Spin Glasses (W. K. Theumann and R. Koberle, Eds.). World Scien.ti�c, Singapore. HANSEL, D., MATO, G., and MEUNIER, C. (1992). Memorization without generalization in a multilayered neural network. Eu.rophys. Lett. 20, 471. KINZEL, W., and RUJ�N, P. (1990). Improving a network generalization ability by selecting examples. Europhys. Lett. 13, 473. LEVIN, E., TISHBY,N.,andSOLLA, S. (1989). A statistical approach to learning and generalization in neural networks. In Proceed.ings of the Second Workshop on Computational Learning The.ory (R. Rivest, D. Haussler, and M. Warmuth, Eds.). Morgan Kaufmann, San Mateo, CA. M�ZARD, M., PARISI, G., and VIRASORO, M. A. (1987). Spin glass theory and beyond. In Lecture Notes in Physics, Vol. 9. World Scienti�c, Singapore. MONASSON, R., and ZECCHINA, R. (1995). Weight space structure and internal representations: A direct approach to learning and generalization in multilayer neural networks. Phys. Rev. Lett. 75, 2432. OPPER, M. (1994). Learning and generalization in a two-layer neural network: The role of the Vapnik�Chervonenkis dimension. Phys. Rev. Lett. 72, 2113. LEARNING TO GENERALIZE OPPER, M., and HAUSSLER, M. (1991). Generalization perfor.mance of Bayes optimal Classification algorithm for learning a perceptron. Phys. Rev. Lett. 66, 2677. OPPER, M., and KINZEL, W. (1995). Statistical mechanics of generalization. In Physics of Neural Networks III (J. L. van Hem-men, E. Domany, and K. Schulten, Eds.). Springer-Verlag, New York. SAAD, D. (Ed.) (1998). Online Learning in Neural Networks. Cambridge Univ. Press, New York. SCHWARZE, H., and HERTZ, J. (1992). Generalization in a large committee machine. Europhys. Lett. 20, 375. SCHWARZE, H., and HERTZ, J. (1993). Generalization in fully con.nected committee machines. Europhys. Lett. 21, 785. SEUNG, H. S., SOMPOLINSKY, H., and TISHBY, N. (1992a). Statis.tical mechanics of learning from examples. Phys. Rev. A 45, 6056. SEUNG, H. S., OPPER, M., and SOMPOLINSKY, H. (1992b). Query by committee. In The Proceedings of the Vth Annual Workshop on Computational Learning Theory (COLT92), p. 287. Associ.ation for Computing Machinery, New York. SOMPOLINSKY, H., TISHBY, N., and SEUNG, H. S. (1990). Learning from examples in large neural networks. Phys. Rev. Lett. 65, 1683. URBANCZIK, R. (1996). Learning in a large committee machine: Worst case and average case. Europhys. Lett. 35, 553. VALLET, F., CAILTON, J., and REFREGIER, P. (1989). Linear and nonlinear extension of the pseudo-inverse solution for learning Boolean functions. Europhys. Lett. 9, 315. VAPNIK, V. N. (1982). Estimation of Dependencies Based on Em.pirical Data. Springer-Verlag, New York. VAPNIK, V. N. (1995). The Nature of Statistical Learning Theory. Springer-Verlag, New York. VAPNIK, V. N., and CHERVONENKIS, A. (1971). On the uniform convergence of relative frequencies of events to their probabil.ities. Theory Probability Appl. 16, 254. General References ARBIB, M. A. (Ed.) (1995). The Handbook of Brain Theory and Neural Networks. MIT Press, Cambridge, MA. BERGER, J. O. (1985). Statistical Decision Theory and Bayesian Analysis. Springer-Verlag, New York. HERTZ,J.A.,KROGH,A.,andPALMER, R. G. (1991). Introduction to the Theory of Neural Computation. Addison-Wesley, Red.wood City, CA. MINSKY, M., and PAPERT, S. (1969). Perceptrons. MIT Press, Cambridge, MA. WATKIN, T. L. H., RAU, A., and BIEHL, M. (1993). The statistical mechanics of learning a rule. Rev. Modern Phys. 65, 499. <> <> <> <> <> <> Model Compression and Acceleration for Deep Neural Networks The principles, progress, and challenges In recent years, deep neural networks (DNNs) have received increased attention, have been applied to different applications, and achieved dramatic accuracy improvements in many tasks. These works rely on deep networks with millions or even billions of parameters, and the availability of graphics process.ing units (GPUs) with very high computation capability plays a key role in their success. For example, Krizhevsky et al. [1] achieved breakthrough results in the 2012 ImageNet Challenge using a network containing 60 million parameters with five convolutional layers and three fully connected layers. Usually, it takes two to three days to train the whole model on the ImagetNet data set with an NVIDIA K40 machine. In another example, the top face-verification results from the Labeled Faces in the Wild (LFW) data set were obtained with networks containing hundreds of millions of parameters, using a mix of convolutional, locally connected, and fully connected layers [2], [3]. It is also very time-consuming to train such a model to obtain a reasonable performance. In architectures that only rely on fully connected layers, the number of parameters can grow to billions [4]. Introduction As larger neural networks with more layers and nodes are considered, reducing their storage and computational cost becomes critical, especially for some real-time applications such as online learning and incremental learning. In addition, recent years witnessed significant progress in virtual reality, augmented reality, and smart wearable devices, creating unprecedented opportunities for researchers to tackle fundamental challenges in deploying deep-learning systems to portable devices with limited resources [e.g., memory, central processing units (CPUs), energy, bandwidth]. Efficient deep-learning methods can have a significant impact on distributed systems, embedded devices, and field-programmable gate ar.ray (FPGA) for artificial intelligence (AI). For example, the residual network-50 (ResNet-50) [5], which has 50 convolutional layers, needs more than 95 megabytes of memory for storage, and numerous floating number multiplications for calculating each image. After discarding some redundant weights, the network still works as usual but saved more than 75% of parameters and 50% computational time. For devices like cell phones and FPGAs with only several megabyte resources, how to compact the models used on them is also important. Achieving these goals calls for joint solutions from many disciplines, including but not limited to machine learning, optimization, computer architecture, data compression, indexing, and hardware design. In this article, we review recent works on compressing and accelerating DNNs, which attracted much attention from the deep-learning community and has already achieved significant progress in past years. We classify these approaches into four categories: 1) Parameter pruning and sharing: The parameter pruning and sharing-based methods explore the redundancy in the model parameters and try to remove the redundant and noncritical ones. 2) Low-rank factorization: Low-rank factorization-based techniques use matrix/tensor decomposition to estimate the informative parameters of the deep convolutional neural networks (CNNs). 3) Transferred/compact convolutional filters: The trans.ferred/compact convolutional filters-based approaches design special structural convolutional filters to reduce the storage and computation complexity. 4) Knowledge distillation (KD): The KD methods learn a dis.tilled model and train a more compact neural network to reproduce the output of a larger network. In Table 1, we briefly summarize these four types of methods. Generally, the parameter pruning and sharing, low-rank factorization, and KD approaches can be used in DNNs with fully connected layers and convolutional layers, achieving comparable performances. On the other hand, methods using transferred/compact filters are designed for models with convolutional layers only. Low-rank factorization and transferred/compact filters-based As larger neural networks with more layers and approaches provide an end-to-end pipeline nodes are considered, and can be easily implemented in a CPU/ reducing their storage GPU environment, which is straight for and computational ward, while parameter pruning and sharing cost becomes critical, use different methods such as vector quan. especially for some real- tization, binary coding, and sparse constraints to perform the task. Usually, it will time applications such take several steps to achieve the goal. as online learning and Regarding training protocols, models incremental learning. based on parameter pruning/sharing low-rank factorization can be extracted from pretrained ones or trained from scratch, while the transferred/ compact filter and KD models can only support training from scratch. These methods are independently designed and complement each other. For example, transferred layers and parameter pruning and sharing can be used together, and model quantization and binarization can be used together with low-rank approximations to achieve further speedup. We will de.scribe the details of each theme and their properties, strengths, and drawbacks in the following sections. Parameter pruning and sharing An early work that showed that network pruning is effective in reducing the network complexity and addressed the overfitting problem is [6]. Since then, it has been widely studied to compress DNN models, trying to remove parameters that are not crucial to the model performance. These techniques can be further classified into three categories: model quantization and binarization, parameter sharing, and structural matrix. Quantization and binarization Network quantization compresses the original network by reducing the number of bits required to represent each weight. Gong et al. [6] and Wu et al. [7] applied k-means scalar quantization to the parameter values. Vanhoucke et al. [8] showed that 8-bit quantization of the parameters can result in significant speedup with minimal loss of accuracy. The work in [9] used Theme Name Parameter pruning and sharing Low-rank factorization Transferred/compact convolutional filters KD Description Reducing redundant parameters that are not sensitive to the performance Using matrix/tensor decomposition to estimate the informative parameters Designing special structural convolutional filters to save parameters Training a compact neural network with distilled knowledge of a large model Applications Convolutional layer and fully connected layer Convolutional layer and fully connected layer Only for convolutional layer Convolutional layer and fully connected layer More Details Robust to various settings, can achieve good performance, can support both train.ing from scratch and pretrained model Standardized pipeline, easily implement.ed, can support both training from scratch and pretrained model Algorithms are dependent on applications, usually achieve good performance, only support training from scratch Model performances are sensitive to applications and network structure, only support training from scratch 16-bit fixed-point representation in stochastic rounding-based CNN training, which significantly reduced memory usage and float- point operations with little loss in classification accuracy. The method proposed in [10] first pruned the unimportant connections and retrained the sparsely connected networks. Then it quantized the link weights using weight-sharing, and then applied Huffman coding to the quantized weights as well as the codebook to further reduce the rate. As shown in Figure 1, it starts by learn.ing the connectivity via normal network train.ing, followed by pruning the small-weight connections. Finally, the network is retrained to learn the final weights for the remaining sparse connections. This work achieves the state-of-the-art performance among all parameter quantization-based methods. It was shown in [11] that Hessian weight could be used to measure the importance of network parameters and proposed to minimize Hessian-weighted quantization errors on average for clustering network parameters. A novel quantization framework was introduced in [12], which reduced the precision of network weights to ternary values. In the extreme case of 1-bit representation of each weight, i.e., binary weight neural networks, there are also many works that directly train CNNs with binary weights; for instance, Binary-Connect [13], BinaryNet [14], and XNORNetworks [15]. The main idea is to directly learn binary weights or activations dur.ing the model training. The systematic study in [16] showed that networks trained with backpropagation could be robust against (robust against or resilient to) specific weight distortions, includ.ing binary weights. Drawbacks However, the accuracy of such binary nets is significantly low.ered when dealing with large CNNs such as GoogleNet. Another drawback of these binary nets is that existing binarization schemes are based on simple matrix approximations and ignore the effect of binarization on the accuracy loss. To address this issue, the work in [17] proposed a proximal Newton algorithm with diagonal Hessian approximation that directly mini.mizes the loss with respect to the binary weights. The work in [18] significantly reduced the time on float-point multiplication in the training stage by stochastically binarizing weights and con-sharing has been used converting multiplications in the hidden state both to reduce network computation to sign changes complexity and to address the overfitting issue. Pruning and sharing Network pruning and sharing has been used both to reduce network complexity and to address the overfitting issue. An early approach to pruning was biased weight decay [19]. The optimal brain damage [20] and the optimal brain surgeon [21] methods reduced the number of connections based on the Hessian of the loss function, and their works suggested that such pruning gave higher accuracy than magnitude-based pruning such as the weight decay meth.od. Those methods supported training from scratch. A recent trend in this direction is to prune redundant, non-informative weights in a pretrained CNN model. For example, Srinivas and Babu [22] explored the redundancy among neurons and proposed a data-free pruning method to remove redundant neurons. Han et al. [23] proposed to reduce the total number of parameters and operations in the entire network. Chen et al. [24] proposed a HashedNets model that used a low-cost hash function to group weights into hash buckets for parameter sharing. The deep compression method in [10] removed the redundant connections and quantized the weights and then used Huffman coding to encode the quantized weights. In [25], a simple regularization method based on soft weight-sharing was proposed, which included both quantization and pruning in one simple (re)train.ing procedure. It is worth noting that the aforementioned pruning schemes typically produce connection pruning in CNNs. There is also growing interest in training compact CNNs with sparsity constraints. Those sparsity constraints are <

> FIGURE 1. The three-stage compression method proposed in [10]: pruning, quantization, and encoding. The input is the original model, and the output is the compression model. nels, or even layers. In filter-level pruning, all of the afore.mentioned works used l_2-norm regularizers. The work in [29] used l1-norm to select and prune unimportant filters. Drawbacks There are some potential issues of the pruning and sharing works. First, pruning with l1 or l2 regularization requires more iterations to converge. Furthermore, all pruning criteria require manual setup of sensitivity for layers, which demands fine-tuning of the parameters and could be cumbersome for some applications. Designing the structural matrix In architectures that contain only fully connected layers, the number of parameters can grow up to billions [4]. Thus, it is critical to explore this redundancy of parameters in fully connected layers, which is often the bottleneck in terms of memory consumption. These network layers use the nonlinear transforms <>, where v () is an element-wise nonlinear operator, x is the input vector, and M is the mn matrix of <> parameters. When M is a large general dense matrix, the cost of storing mn parameters and computing matrix-vector products in Om( n) time. Thus, an intuitive way to prune parameters is to impose x as a parameterized structural matrix. An mn matrix <> that can be described using much fewer parameters than mn is called a structured matrix. Typically, the structure should not only reduce the memory cost but also dramatically accelerate the inference and training stage via fast matrix-vector multiplication and gradient computations. Following this direction, the work in [30] proposed a sim.ple and efficient approach based on circulant projections, while maintaining competitive error rates. Given a vector <>, a circulant matrix <> is defined as Thus the memory cost becomes <> instead of <>. <> This circulant structure also enables the use of fast Fourier transform (FFT) to speed up the computation. Given a d-dimensional vector r, the 1-layer circulant neural network in (1) has time complexity of <>. In [31], a novel adaptive fastfood transform was introduced to reparameterize the matrix-vector multiplication of fully connected layers. The adaptive fastfood translation invariant property form matrix <> was defined as of the representations to input image, which is the key <>. (2) to the success of training due to exploring the very deep models without Here, <> are random diago-SG and severe overfitting. nal matrices. <> is a random permutation matrix and H denotes the Walsh-Hadamard matrix. Reparameterizing a fully connected layer with d inputs and n outputs using the adaptive fastfood transform reduces the storage and the computational costs from <> to <> and from <> to <> And <>, respectively. The work in [32] showed the effectiveness of the new notion of parsimony in the theory of structured matrices. Their pro.posed method can be extended to various other structured matrix classes, including block and multilevel Toeplitz-like [33] matrices related to multidimensional convolution [34]. Drawbacks One potential problem of this kind of approach is that the structural constraint will cause loss in accuracy since the constraint might bring bias to the model. On the other hand, how to find a proper structural matrix is difficult. There is no theoretical way from which to derive it. Low-rank factorization and sparsity As convolution operations constitute the bulk of all computations in CNNs, simplifying the convolution layer would have a direct impact on the overall speedup. The convolution kernels in a typical CNN is a four-dimensional tensor. The key observation is that there might be a significant amount of redundancy in the tensor. Ideas based on tensor decomposition seem to be a particularly promising way to remove the redundancy. Regarding to the fully connected layer, it can be viewed as a two-dimensional (2-D) matrix and the low-rankness can also help. Using low-rank filters to accelerate convolution has a long history. Typical examples include high-dimensional discrete cosine transform (DCT) and wavelet systems constructed from one-dimensional (1-D) DCT transform and 1-D wave.lets, respectively, using tensor products. In the context of dictionary learning, Rigamonti et al. [35] suggested learning separable 1-D filters. In [36], a few low-rank approximation <> and clustering schemes for the convolutional kernels were proposed. They achieved 2# speedup for a single convolutional layer with 1% drop in classification accuracy. The work in [37] suggested using different tensor decomposition schemes, reporting a 45. # speedup with 1% drop in accuracy <> case. For the scheme in [39], the decomposition always exists and can achieve better performance than general CP. Table 2 lists a performance comparison of both methods. The actual speedup and compression rates are used to mea.sure the performances. We can see that the BN version can achieve slightly bet.ter performance while the CP version gives higher compression rates. Original Framework Low-Rank Note that the fully connected layers <

> FIGURE 2. A typical framework of the low-rank regularization method. (a) is the original convolutional layer, and (b) is the low-rank constraint convolutional layer with rank-K. in text recognition. In both works, the approximation was done layer by layer. After one layer was approximated by the low-rank filters, the parameters of that layer were fixed, and the layers above were fine-tuned based on a reconstruction error criterion. These are typical low-rank methods for compressing 2-D convolutional layers, which is described in Figure 2. In [38], canonical polyadic (CP) decomposition of the kernel tensors was proposed. Their work used nonlinear least squares to compute the CP decomposition, which was also based on the tensor decomposition idea. In [39], a new algorithm for computing the low-rank tensor decomposition and a new method for training low-rank constrained CNNs from scratch were proposed. It used batch normalization (BN) to transform the activations of the internal hidden units, and it was shown to be an effective way to deal with the exploding or vanishing gradients. In principle, both the CP decomposition scheme and the decomposition scheme in [39] (BN low-rank) can be used to train CNNs from scratch. For the CP decomposition, finding the best low-rank approximation is an ill-posed problem, and the best rank-K approximation may not exist in the general in fully connected layers. For instance, Misha et al. [40] reduced the number of dynamic parameters in deep models using the low-rank method. Reference [41] explored a low-rank matrix factorization of the final weight layer in a DNN for acoustic modeling. Drawbacks Low-rank approaches are straightforward for model compression and acceleration. The idea complements recent advances in deep learning such as dropout, rectified units, and maxout. However, the implementation is not that easy since it involves a decomposition operation, which is computationally expensive. Another issue is that current methods perform low-rank approximation layer by layer, and thus cannot perform global parameter compression, which is important as different layers hold different information. Finally, factorization requires extensive model retraining to achieve convergence when com.pared to the original model. Transferred/compact convolutional filters CNNs are parameter-efficient due to exploring the translation invariant property of the representations to input image, which is the key to the success of training very deep models without severe overfitting. Although a strong theory is currently missing, a large amount of empirical evidence sup.ports the notion that both the translation invariant property and convolutional weight-sharing are important for good predictive performance. The idea of using transferred convolutional filters to compress CNN models is motivated by recent works in [42], which introduced the equivariant group theory. Let x be an input, <> be a network or layer, and <> be the transform matrix. The concept of equivariance is defined as <>, (3) which says that transforming the input x by the transform <> and then passing it through the network or layer <> should give the same result as first mapping x through the network and then transforming the representation. Note that, Model AlexNet BN low-rank CP low-rank VGG-16 BN low-rank CP low-rank GoogleNet BN low-rank CP low-rank <

> in [42], the transforms <> and Tl ()$ are not necessarily the same as they operate on different objects. According to this theory, it is reasonable to apply the transform to layers or filters <> to compress the whole network models. From empirical observation, deep CNNs also benefit from using a large set of convolutional filters by applying a certain transform <> to a small set of base filters since it acts as a regularizer for the model. Following this trend, there are many recent works proposed to build a convolutional layer from a set of base filters [42] [45]. What they have in common is that the transform <> lies in the family of functions that only operate in the spatial domain of the convolutional filters. For example, the work in [44] found that the lower convolution layers of CNNs learned redundant filters to extract both positive and negative phase information of an input signal, and defined <> to be the simple negation function - <>. (4) Here, Wx is the basis convolutional filter - and Wx is the filter consisting of the shifts whose activation is opposite to that of Wx and selected after max-pooling operation. By doing this, the work in [44] can easily achieve 2# compression rate on all the convolutional layers. It is also shown that the negation transform acts as a strong regularizer to improve the classification accuracy. The intuition is that the learning algorithm with pair-wise positive-negative constraint can lead to useful convolutional filters instead of redundant ones. In [45], it was observed that magnitudes of the responses from convolutional kernels had a wide diversity of pattern representations in the network, and it was not proper to discard weaker signals with a single threshold. Thus, a multibias nonlinearity activation function was proposed to generate more patterns in the feature space at low computational cost. The transform <> was define as <>, (5) where d were the multibias factors. The work in [46] considered a combination of rotation by a multiple of 90% and horizontal/vertical flipping with <>, (6) where WTi was the transformation matrix that rotated the original filters with angle i ! {90,180,270}. In [42], the transform was generalized to any angle learned from data, and i was directly obtained from data. Both [46] and [42] can achieve good classification performance. Reference [43] defined <> as the set of translation functions applied to 2-D filters <>, (7) The basic idea of KD is to distill knowledge from a Drawbacks large teacher model into There are several issues that need to be a small one by learning addressed for approaches that apply transfer information to convolutional filters. First, the class distributions these methods can achieve competitive performance for wide/flat architectures (like output by the teacher via softened softmax. where <> denoted the translation of the first operand by <> xy along its spatial dimensions, with proper zero padding at borders to maintain the shape. The proposed framework can be used to 1) improve the classification accuracy as a regularized version of maxout networks and 2) to achieve parameter efficiency by flexibly varying their architectures to compress networks. Table 3 briefly compares the performance of different methods with transferred convolutional filters, using VGG-Net (16 layers) as the baseline model. The results are report.ed on the CIFAR-10 and CIFAR-100 data sets with top-five error rates. It is observed that they can achieve reduction in parameters with little or no drop in classification accuracy. VGGNet) but not narrow/special ones (like GoogleNet and ResNet). Second, the trans.fer assumptions sometimes are too strong to guide the algorithm, making the results unstable on some data sets. Using a compact filter for convolution can directly reduce the computation cost. The key idea is to replace the loose and overparametric filters with compact blocks to improve the speed, which significantly accelerate CNNs on several benchmarks. Decomposing 33 convolution into two 1x1 convolutions was used in [47], which achieved state-of-the-art acceleration performance on object recognition. SqueezeNet [48] was proposed to replace 33# convolution with 1x1 convolution, which created a compact neural network with approximately 50 fewer parameters and comparable accuracy when compared to AlexNet. KD To the best of our knowledge, exploiting knowledge transfer to compress model was first proposed by Caruana et al. [49]. They trained a compressed model with pseudo-data labeled by an ensemble of strong classifiers and reproduced the output of the original larger network. However, their work is limited to shal.low models. The idea has been recently adopted in [50] as KD to compress deep and wide networks into shallower ones, where <> the compressed model mimicked the function learned by the complex model. The basic idea of KD is to distill knowledge from a large teacher model into a small one by learning the class distributions output by the teacher via softened softmax. The work in [51] introduced a KD compression framework, which eased the training of deep networks by following a student-teacher paradigm, in which the student was penalized according to a softened version of the teacher's output. The framework compressed an ensemble of deep networks (teacher) into a stu.dent network of similar depth. To do so, the student was trained to predict the output of the teacher, as well as the true classifica.tion labels. Despite its simplicity, KD demonstrates promising results in various image classification tasks. The work in [52] aimed to address the network compression problem by taking advantage of depth neural networks. It proposed an approach to train thin and deep networks, called FitNets, to compress wide and shallower (but still deep) networks. The method was rooted in KD and extended the idea to allow for thinner and deeper student models. To learn from the intermediate representations of the teacher network, FitNet made the student mimic the full feature maps of the teacher. However, such assumptions are too strict since the capacities of teacher and student may differ greatly. In certain circumstances, FitNet may adversely affect the performance and convergence. All the aforementioned methods are validated on the MNIST, CIFAR-10, CIFAR-100, SVHN, and AFLW benchmark data sets, and simulation results show that these methods match or outperform the teacher's performance, while requiring notably fewer parameters and multiplications. There are several extensions along this direction of distillation knowledge. The work in [53] trained a parametric student model to approximate a Monte Carlo teacher. The proposed framework used online training and used DNNs for the student model. Different from previous works, which represented the knowledge using the softened label probabilities, [54] represented the knowledge by using the neurons in the higher hidden layer, which preserved as much information as the label probabilities, but are more compact. The work in [55] accelerated the experimentation process by instantaneously transferring the knowledge from a previous network to each new deeper or wider network. The techniques are based on the concept of function-preserving transformations between neural net.work specifications. Zagoruyko et al. [56] proposed attention transfer to relax the assumption of FitNet. They transferred the attention maps that are summaries of the full activations. Drawbacks KD-based approaches can make deeper models thinner and help significantly reduce the computational cost. However, there are a few disadvantages. One of them is that KD can only be applied to classification tasks with softmax loss function, which hinders its usage. Another drawback is that the model assumptions sometimes are too strict to make the performance competitive with other types of approaches. Other types of approaches We first summarize the works utilizing attention-based methods. Note that attention-based systems [57] can reduce computations significantly by learning to selectively focus or �attend to� a few, task-relevant input regions. The work in [57] introduced the dynamic capacity network that combined two types of modules: the small subnetworks with low capacity, and the large ones with high capacity. The low-capacity subnetworks were active on the whole input to first find the task-relevant areas in the input, and then the attention mechanism was used to direct the high-capacity subnetworks to focus on the task-relevant regions in the input. By doing this, the size of the CNN model could be significantly reduced. Following this direction, the work in to measure the quality some important neurons. It proposed a new and acceleration are the The standard criteria [58] introduced the conditional computation idea, which only computes the gradient for of model compression type of general-purpose neural network component: a sparsely gated mixture-of-experts compression and the (MoE) layer. The MoE consisted of a number speedup rates. of experts, each a simple feed-forward neural network, and a trainable gating network that selected a sparse combination of the experts to process each input. In [59], dynamic DNNs (D2NNs) were introduced, which were a type of feed-forward DNN that selected and executed a subset of D2NN neurons based on the input. There have been other attempts to reduce the number of parameters of neural networks by replacing the fully connected layer with global average pooling[43], [60]. Network architectures, such as GoogleNet or network in network, can achieve state-of-the-art results on several benchmarks by adopting this idea. However, transfer learning, i.e., reusing features learned on the ImageNet data set and applying them to new tasks, is more difficult with this approach. This problem was noted by Szegedy et al. [60] and motivated them to add a linear layer on top of their networks to enable transfer learning. The work in [61] targeted the ResNet-based model with a spatially varying computation time, called stochastic depth, which enabled the seemingly contradictory setup to train short networks and used deep networks at test time. It started with very deep networks and, while during training, for each mini-batch, randomly dropped a subset of layers and bypassed them with the identity function. This model is end-to-end trainable, deterministic, and can be viewed as a black-box feature extractor. Following this direction, the work in [62] proposed a pyramidal residual network with stochastic depth. Other approaches to reduce the convolutional overheads include using FFT-based convolutions [63] and fast convolution using the Winograd algorithm [64]. Those works only aim to speedup the computation but not reduce the memory storage. Benchmarks, evaluation, and databases In the past five years, the deep-learning community has made great efforts in benchmark models. One of the most well-known models used in compression and acceleration for CNNs is Alexnet [1], which occasionally has been used for assessing the performance of compression. Other popular standard models include LeNets [65], All-CNN-nets [66], and many others. LeNet-300-100 is a fully connected network with two hidden layers, with 300 and 100 neurons each. LeNet-5 is Proposing some general/ unified approaches is one direction that can be taken regarding the use of CNNs in small platforms. about how to choose different compression approaches and possible challenges/solutions in this area. General suggestions There is no golden rule to measure which one of the four kinds of approaches is the best. How a convolutional network that has two convolutional layers and two fully connected layers. Recently, more state-of-the-art architectures are used as baseline models in many works, including network in networks [67], VGGNets [68], and ResNets [69]. Table 4 summarizes the baseline mod.els commonly used in several typical compression methods. The standard criteria to measure the quality of model compression and acceleration are the compression and the speedup rates. Assume that a is the number of the parameters in the original model M and a* is that of the compressed model M*, then the compression rate a (,MM*) of M* over Mis <> (8) Another widely used measurement is the index space saving defined in several papers [70], [71] as <>, (9) where a and a are the number of the dimension of the index space in the original model and that of the compressed model, respectively. Similarly, given the running time s of M and s* of M* , the speedup rate d (,MM*) is defined as <> (10) Most work used the average training time per epoch to mea.sure the running time, while in [70] and [71], the average testing time was used. Generally, the compression rate and speedup rate are highly correlated, as smaller models often results in faster computation for both the training and the testing stages. Good compression methods are expected to achieve almost the same performance as the original model with much smaller parameters and less computational time. However, for differ.ent applications with varying CNN designs, the correlation between parameter size and computational time may be different. For example, it is observed that, for deep CNNs with fully connected layers, most of the parameters are in the fully connected layers; while for image classification tasks, float-point operations are mainly in the first few convolutional layers since each filter is convolved with the whole image, which is usually very large at the beginning. Different applications should focus on different layers. Discussion and challenges In this article, we summarized recent works on compress.ing and accelerating DNNs. Here we discuss more details to choose the proper approaches is really de.pendent on the applications and requirements. Here, we provide some general suggestions. If the applications needs compacted models from pretrained models, one can choose either pruning and sharing or low-rank factorization-based methods. If end-to-end solutions are needed for the problem, the low-rank and transferred convolutional filters approaches are preferred. For applications in some specific domains, methods with human prior (like the transferred convolutional filters and structural matrix) sometimes have benefits. For example, when conducting medical images classification, transferred convolutional filters should work well as medical images (like organs) do have the rotation transformation property. Usually, the approaches of pruning and sharing could give a reasonable compression rate while not hurting the accuracy. Thus, for applications that require stable model accuracy, it is better to utilize pruning and sharing. If a problem involves small- or medium-size data sets, one can try the KD approaches. The compressed student model can take the benefit of transferring knowledge from the teacher model, making it a robust data set that is not large. As we mentioned in the �Introduction,� techniques of the four themes are orthogonal. It makes sense to combine two or three of them to maximize the compression/speedup rates. For some specific applications, like object detection, which requires both convolutional and fully connected layers, one can compress the convolutional layers with low-rank factorization and the fully connected layers with a pruning method. <> Technique challenges Techniques for deep model compression and acceleration are still in the early stages, and the following challenges still need to be addressed. Most of the current state-of-the-art approaches are built on well-designed CNN models, which have limited freedom to change the configuration (e.g., network structural, hyperparameters). To handle more complicated tasks, it should provide more plausible ways to configure the compressed models. Good compression methods are expected to achieve almost the same performance as the original model with much smaller parameters and less computational time. approaches. Instead of directly reducing and transferring parameters from the teach.er models, passing selectivity knowledge of neurons could be helpful. One can derive a way to select essential neurons related to the task. The intuition is that, if a neuron is activated in certain regions or samples, this implies these regions or samples share Pruning is an effective way to compress and accelerate CNNs. Current pruning techniques are mostly designed to eliminate connections between neurons. On the other hand, a pruning channel can directly reduce the feature map width and shrink the model into a thinner one. It is efficient but also challenging because removing channels might dramatically change the input of the following layer. It is important to focus on how to address this issue. As we mentioned previously, methods of structural matrix and transferred convolutional filters impose prior human knowledge to the model, which could significantly affect the performance and stability. It is critical to investigate how to control the impact of the imposed prior knowledge. The methods of KD provide many benefits such as directly accelerating the model without special hardware or implementations. It is still worth it to develop KD-based approaches and explore how to improve the performance. Hardware constraints in various of small platforms (e.g., mobile, robotic, self-driving cars) are still a major problem that hinder the extension of deep CNNs. How to make full use of the limited computational source available and how to design special compression methods for such platforms are still challenges that need to be addressed. Possible solutions To solve the hyperparameters configuration problem, we can rely on the recent learning-to-learn strategy [72], [73]. This framework provides a mechanism, allowing the algorithm to automatically learn how to exploit structure in the problem of interest. There are two different ways to combine the learning.to-learn module with the model compression. The first designs compression and learning-to-learn simultaneously, while the second way first configures the model with learn-to-learning and then prunes the parameters. Channel pruning provides the efficiency benefit on both CPUs and GPUs because no special implementation is required. But it is also challenging to handle the input con.figuration. One possible solution is to use the training-based channel pruning methods [74], which focus on imposing sparse constraints on weights during training, and could adaptively determine hyperparameters. However, training from scratch for such a method is costly for very deep CNNs. Exploring new types of knowledge in the teacher models and transferring it to the student models is useful for the KD some common properties that may relate to the task. Performing such steps is time-consuming, thus efficient implementation is important. For methods with convolutional filters and the structural matrix, we can conclude that the transformation lies in the family of functions that only operations on the spatial dimen.sions. Hence, to address the imposed prior issue, one solution is to provide a generalization of the aforementioned approach.es in two aspects: 1) instead of limiting the transformation to belong to a set of predefined transformations, let it be the whole family of spatial transformations applied to 2-D filters or the matrix, and 2) learn the transformation jointly with all of the model parameters. Proposing some general/unified approaches is one direction that can be taken regarding the use of CNNs in small platforms. Yuhen et al. [75] presented a feature map dimensionality reduc.tion method by excavating and removing redundancy in feature maps generated by different filters, which could also preserve intrinsic information of the original network. The idea can be extended to make CNNs more applicable for different platforms. The work in [76] proposed a one-shot whole network compression scheme consisting of three components: rank selection, low-rank tensor decomposition, and fine-tuning to make deep CNNs work in mobile devices. From the systematic side, Facebook released the platform Caffe2 [77], which employed a particularly lightweight and modular framework and included mobile-specif.ic optimizations based on the hardware design. Caffe2 can help developers and researchers train large machine-learning models and deliver AI on mobile devices. Acknowledgments We would like to thank the reviewers and broader community for their feedback on this survey. In particular, we would like to thank Hong Zhao from the Department of Automation of Tsinghua University for her help on modifying this article. This research is supported by National Science Foundation of China, grant number 61401169. The corresponding author of this article is Pan Zhou. Authors Yu Cheng (chengyu@us.ibm.com) received his bachelor�s degree in automation from Tsinghua University, Beijing, China, in 2010 and his Ph.D. degree in computer science from Northwestern University, Evanston, Illinois in 2015. Currently, he is a research staff member at AI Foundations Lab, IBM T.J. Watson Research Center, Yorktown Heights, New York. His research is focused on deep learning in general, with specific interests in deep generative models and deep models compression. He also has published many works regarding the applications of deep learning in computer vision and natural language processing. Duo Wang (d-wang15@mails.tsinghua.edu.cn) received the B.S. degree in automation from the Harbin Institute of Technology, China, in 2015, where he is currently pursuing his Ph.D. degree in the Department of Automation, Tsinghua University. His research interests are deep/machine learning and their applications in computer vision and robotics vision. Pan Zhou (panzhou@hust.edu.cn) received his B.S. degree in the Advanced Class of Huazhong University of Science and Technology (HUST), Wuhan China, and his M.S. degree in electronics and information engineering from the same university in 2006 and 2008, respectively. He received his Ph.D. degree from the School of Electrical and Computer Engineering at the Georgia Institute of Technology, Atlanta in 2011. Currently, he is an associate professor with School of Electronic Information and Communications, HUST. His research interests include big data analytics and machine learning, security and privacy, and information networks. Tao Zhang (taozhang@mail.tsinghua.edu.cn) received his B.S., M.S., and Ph.D. degrees from Tsinghua University, Beijing, China, in 1993, 1995, and 1999, respectively, and his Ph.D. degree from Saga University, Japan, in 2002, all in control engineering. He is a professor with the Department of Automation, Tsinghua University. His current research interests include artificial intelligence, robotics, image processing, control theory, and control of spacecraft. References [1] A. Krizhevsky, I. Sutskever, and G. Hinton, �Imagenet classification with deep convolutional neural networks,� in Proc. Conf. Neural Information Processing Systems, 2012, pp. 1097�1105. [2] Y. Taigman, M. Yang, M. Ranzato, and L. Wolf, �Deepface: Closing the gap to human-level performance in face verification,� in Proc. IEEE Conf. Computer Vision Pattern Recognition, 2014, pp. 1701�1708. [3] Y. Sun, X. Wang, and X. Tang, �Deeply learned face representations are sparse, selective, and robust,� in Proc. IEEE Conf. Computer Vision Pattern Recognition, 2015, pp. pp. 2892�2900. [4] J. Dean, G. Corrado, R. Monga, K. Chen, M. Devin, Q. Le, M. Mao, M. Ranzato, A. Senior, P. Tucker, K. Yang, and A. Ng, �Large scale distributed deep networks,� in Proc. Conf. Neural Information Processing Systems, 2012, pp. 1223�1231. [5] K. He, X. Zhang, S. Ren, and J. Sun, �Deep residual learning for image recogni.tion,� Computing Res. Repository, vol. abs/1512.03385, 2015. [Online]. Available: https://arxiv.org/pdf/1512.03385.pdf [6] Y. Gong, L. Liu, M. Yang, and L. D. Bourdev, �Compressing deep convolutional networks using vector quantization,� Computing Res. Repository, vol. abs/1412.6115, 2014. [Online]. Available: https://arxiv.org/pdf/1412.6115.pdf [7] Y. W. Q. H. Jiaxiang Wu, C. Leng, and J. Cheng, �Quantized convolutional neu.ral networks for mobile devices,� in Proc. IEEE Conf. Computer Vision Pattern Recognition, 2016, pp. 4820�4828. [8] V. Vanhoucke, A. Senior, and M. Z. Mao, �Improving the speed of neural net.works on cpus,� in Proc. Conf. Neural Information Processing Systems Deep Learning and Unsupervised Feature Learning Workshop, 2011. [9] S. Gupta, A. Agrawal, K. Gopalakrishnan, and P. Narayanan, �Deep learning with limited numerical precision,� in Proc. 32nd Int. Conf. Machine Learning, 2015, vol. 37, pp. 1737�1746. [10] S. Han, H. Mao, and W. J. Dally, �Deep compression: Compressing deep neural networks with pruning, trained quantization and Huffman coding,� in Proc. Int. Conf. Learning Representations, 2016. [11] Y. Choi, M. El-Khamy, and J. Lee, �Towards the limit of network quantization,� Computing Res. Repository, vol. abs/1612.01543, 2016. [Online]. Available: https://arxiv.org/abs/1612.01543 [12] C. Zhu, S. Han, H. Mao, and W. J. Dally, �Trained ternary quantization,� arXiv Preprint, arXiv:1612.01064, 2016. [13] M. Courbariaux, Y. Bengio, and J. David, �Binaryconnect: Training deep neu.ral networks with binary weights during propagations,� in Proc. Advances Neural Information Processing Systems Annu. Conf., 2015, pp. 3123�3131. [14] M. Courbariaux and Y. Bengio, �Binarynet: Training deep neural networks with weights and activations constrained to +1 or .1,� Computing Res. Repository, vol. abs/1602.02830, 2016. [Online]. Available: https://arxiv.org/abs/1602.02830 [15] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi, �Xnor-net: Imagenet classification using binary convolutional neural networks,� in Proc. European Conf. Computer Vision, 2016, pp. 525�542. [16] P. Merolla, R. Appuswamy, J. V. Arthur, S. K. Esser, and D. S. Modha, �Deep neural networks are robust to weight binarization and other non-linear distortions,� Computing Res. Repository, vol. abs/1606.01981, 2016. [Online]. Available: https:// arxiv.org/abs/1606.01981 [17] L. Hou, Q. Yao, and J. T. Kwok, �Loss-aware binarization of deep networks,� Computing Res. Repository, vol. abs/1611.01600, 2016. [Online]. Available: https:// arxiv.org/abs/1611.01600 [18] Z. Lin, M. Courbariaux, R. Memisevic, and Y. Bengio, �Neural networks with few multiplications,� Computing Res. Repository, vol. abs/1510.03009, 2015. [Online]. Available: https://arxiv.org/abs/1510.03009 [19] S. J. Hanson and L. Y. Pratt, �Comparing biases for minimal network con.struction with back-propagation,� Adv. Neural Inform. Process. Syst. 1, 1989, pp. 177�185. [20] Y. L. Cun, J. S. Denker, and S. A. Solla, �Advances in neural information pro.cessing systems 2,� in Optimal Brain Damage, D. S. Touretzky, Ed. San Mateo, CA: Morgan Kaufmann, 1990, pp. 598�605. [21] B. Hassibi, D. G. Stork, and S. C. R. Com, �Second order derivatives for network pruning: Optimal brain surgeon,� in Advances in Neural Information Processing Systems, vol. 5. San Mateo, CA: Morgan Kaufmann, 1993, pp. 164� 171. [22] S. Srinivas and R. V. Babu, �Data-free parameter pruning for deep neural net.works,� in Proc. British Machine Vision Conf., 2015, pp. 31.1�31.12. [23] S. Han, J. Pool, J. Tran, and W. J. Dally, �Learning both weights and connections for efficient neural networks,� in Proc. 28th Int. Conf. Neural Information Processing Systems, 2015, pp. 1135�1143. [24] W. Chen, J. Wilson, S. Tyree, K. Q. Weinberger, and Y. Chen, �Compressing neural networks with the hashing trick,� in Proc. Machine Learning Research Workshop Conf., 2015, pp. 2285�2294. [25] K. Ullrich, E. Meeds, and M. Welling, �Soft weight-sharing for neural network compression,� Computing Res. Repository, vol. abs/1702.04008, 2017. [Online]. Available: https://arxiv.org/abs/1702.04008 [26] V. Lebedev and V. S. Lempitsky, �Fast convnets using group-wise brain dam.age,� in Proc. IEEE Conf. Computer Vision Pattern Recognition, 2016, pp. 2554� 2564. [27] H. Zhou, J. M. Alvarez, and F. Porikli, �Less is more: Towards compact CNNs,� in Proc. European Conf. Computer Vision, 2016, pp. 662�677. [28] W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li, �Learning structured sparsity in deep neural networks,� Adv. Neural Inform. Process. Syst., vol. 29, pp. 2074�2082, 2016. [29] H. Li, A. Kadav, I. Durdanovic, H. Samet, and H. P. Graf, �Pruning filters for efficient convnets,� Computing Res. Repository, vol. abs/1608.08710, 2016. [Online]. Available: https://arxiv.org/abs/1608.08710 [30] Y. Cheng, F. X. Yu, R. Feris, S. Kumar, A. Choudhary, and S.-F. Chang, �An exploration of parameter redundancy in deep networks with circulant projections,� in Proc. Int. Conf. Computer Vision, 2015, pp. 2857�2865. [31] Z. Yang, M. Moczulski, M. Denil, N. de Freitas, A. Smola, L. Song, and Z. Wang, �Deep fried convnets,� in Proc. Int. Conf. Computer Vision, 2015, pp. 1476� 1483. [32] V. Sindhwani, T. Sainath, and S. Kumar. (2015). Structured transforms for small-footprint deep learning. Advances in Neural Information Processing Systems, 28, pp. 3088�3096. [Online]. Available: http://papers.nips.cc/paper/5869.structured-transforms-for-small-footprint-deep-learning.pdf [33] J. Chun and T. Kailath, Generalized Displacement Structure for Block-Toeplitz, Toeplitz-Block, and Toeplitz-Derived Matrices. Berlin, Germany: Springer, 1991, pp. 215�236. [34] M. V. Rakhuba and I. V. Oseledets. (2015). Fast multidimensional convolution in low-rank tensor formats via cross approximation. SIAM J. Sci. Comput., 37(2). [Online]. Available: http://dx.doi.org/10.1137/140958529 [35] R. Rigamonti, A. Sironi, V. Lepetit, and P. Fua, �Learning separable filters,� in Proc. IEEE Conf. Computer Vision Pattern Recognition, 2013, pp. 2754� 2761. [36] E. L. Denton, W. Zaremba, J. Bruna, Y. LeCun, and R. Fergus, �Exploiting lin.ear structure within convolutional networks for efficient evaluation,� Adv. Neural Inform. Process. Syst. vol. 27, pp. 1269�1277, 2014. [37] M. Jaderberg, A. Vedaldi, and A. Zisserman, �Speeding up convolutional neu.ral networks with low rank expansions,� in Proc. British Machine Vision Conf., 2014, pp. 1�13. [38] V. Lebedev, Y. Ganin, M. Rakhuba, I. V. Oseledets, and V. S. Lempitsky, �Speeding-up convolutional neural networks using fine-tuned CP-decomposition,� Computing Res. Repository, vol. abs/1412.6553, 2014. [Online]. Available: https:// arxiv.org/abs/1412.6553 [39] C. Tai, T. Xiao, X. Wang, and E. Weinan, �Convolutional neural networks with low-rank regularization,� Computing Res. Repository, vol. abs/1511.06067, 2015. [40] M. Denil, B. Shakibi, L. Dinh, M. Ranzato, and N. D. Freitas. (2013). Predicting parameters in deep learning. Advances in Neural Information Processing Systems, 26, 2148�2156. [Online]. Available: http://media.nips.cc/nips.books/nipspapers/paper_files/nips26/1053.pdf [41] T. N. Sainath, B. Kingsbury, V. Sindhwani, E. Arisoy, and B. Ramabhadran, �Low-rank matrix factorization for deep neural network training with high-dimen.sional output targets,� in Proc. IEEE Int. Conf. Acoustics Speech Signal Processing, 2013, pp. 6655�6659. [42] T. S. Cohen and M. Welling, �Group equivariant convolutional networks,� arXiv Preprint, arXiv:1602.07576, 2016. [43] S. Zhai, Y. Cheng, and Z. M. Zhang, �Doubly convolutional neural networks,� in Proc. Advances Neural Information Processing Systems, 2016, pp. 1082�1090. [44] W. Shang, K. Sohn, D. Almeida, and H. Lee, �Understanding and improving convolutional neural networks via concatenated rectified linear units,� arXiv Preprint, arXiv:1603.05201, 2016. [45] H. Li, W. Ouyang, and X. Wang, �Multi-bias non-linear activation in deep neural networks,� arXiv Preprint, arXiv:1604.00676, 2016. [46] S. Dieleman, J. D Fauw, and K. Kavukcuoglu, �Exploiting cyclic symmetry in convolutional neural networks,� in Proc. 33rd Int. Conf. Machine Learning, 2016, vol. 48, pp. 1889�1898. [47] C. Szegedy, S. Ioffe, and V. Vanhoucke. (2016). Inception-v4, inception-resnet and the impact of residual connections on learning, Computing Res. Repository, vol. abs/1602.07261. [Online]. Available: http://dblp.uni-trier.de/db/journals/corr/corr1602. html#SzegedyIV16 [48] B. Wu, F. N. Iandola, P. H. Jin, and K. Keutzer, �Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autono.mous driving,� Computing Res. Repository, vol. abs/1612.01051, 2016. [Online]. Available: https://arxiv.org/abs/1612.01051 [49] C. Bucilua�, R. Caruana, and A. Niculescu-Mizil. (2006). Model compression. Proc. 12th ACM SIGKDD Int. Conf. Knowledge Discovery Data Mining, pp. 535� 541. [Online]. Available: http://doi.acm.org/10.1145/1150402.1150464 [50] J. Ba and R. Caruana, �Do deep nets really need to be deep?� Adv. Neural Inform. Process. Syst., vol. 27, pp. 2654�2662, 2014. [51] G. E. Hinton, O. Vinyals, and J. Dean, �Distilling the knowledge in a neural net.work,� Computing Res. Repository, vol. abs/1503.02531, 2015. [Online]. Available: https://arxiv.org/abs/1503.02531 [52] A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y. Bengio, �Fitnets: Hints for thin deep nets,� Computing Res. Repository, vol. abs/1412.6550, 2014. [Online]. Available: https://arxiv.org/abs/1412.6550 [53] A. Korattikara Balan, V. Rathod, K. P. Murphy, and M. Welling. (2015). Bayesian dark knowledge. Advances in Neural Information Processing Systems, 28, 3420�3428. [Online]. Available: http://papers.nips.cc/paper/5965-bayesian-dark-knowledge.pdf [54] P. Luo, Z. Zhu, Z. Liu, X. Wang, and X. Tang, �Face model compression by dis.tilling knowledge from neurons,� in Proc. 30th AAAI Conf. Artificial Intelligence, 2016, pp. 3560�3566. [55] T. Chen, I. J. Goodfellow, and J. Shlens, �Net2net: Accelerating learning via knowledge transfer,� Computing Res. Repository, vol. abs/1511.05641, 2015. [Online]. Available: https://arxiv.org/abs/1511.05641 [56] S. Zagoruyko and N. Komodakis. (2016). Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer, Computing Res. Repository, vol. abs/1612.03928. [Online]. Available: http://arxiv.org/ abs/1612.03928 [57] A. Almahairi, N. Ballas, T. Cooijmans, Y. Zheng, H. Larochelle, and A. C. Courville, �Dynamic capacity networks,� in Proc. 33rd Int. Conf. Machine Learning, 2016, pp. 2549�2558. [58] N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean. (2017). Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. [Online]. Available: https://openreview.net/pdf?id=B1ckMDqlg [59] D. Wu, L. Pigou, P. Kindermans, N. D. Le, L. Shao, J. Dambre, and J. Odobez, �Deep dynamic neural networks for multimodal gesture segmentation and recognition,� IEEE Trans. Pattern Anal. Mach. Intell., vol. 38, no. 8, pp. 1583� 1597, 2016. [60] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. (2015). Going deeper with convolutions. Proc. IEEE Computer Vision Pattern Recognition. [Online]. Available: http://arxiv.org/ abs/1409.4842 [61] G. Huang, Y. Sun, Z. Liu, D. Sedra, and K. Q. Weinberger, �Deep networks with stochastic depth,� Computing Res. Repository, vol. arXiv:1603.09382, 2016. [62] Y. Yamada, M. Iwamura, and K. Kise. (2016). Deep pyramidal residual networks with separated stochastic depth, Computing Res. Repository, vol. abs/1612.01230. [Online]. Available: http://arxiv.org/abs/1612.01230 [63] M. Mathieu, M. Henaff, and Y. Lecun, �Fast training of convolutional networks through FFTs,� Computing Res. Repository, vol. arXiv:1312.5851, 2014. [64] A. Lavin and S. Gray, �Fast algorithms for convolutional neural networks,� in Proc. IEEE Conf. Computer Vision Pattern Recognition, 2016, pp. 4013� 4021. [65] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, �Gradient-based learning applied to document recognition,� Proc. IEEE, pp. 2278�2324, 1998. [66] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. A. Riedmiller, �Striving for simplicity: The all convolutional net,� Computing Res. Repository, vol. abs/1412.6806, 2014. [Online]. Available: https://arxiv.org/abs/1412.6806 [67] M. Lin, Q. Chen, and S. Yan, �Network in network,� in Proc. Int. Conf. Learning Representations, 2014. [Online]. Available: https://arxiv.org/abs/ 1312.4400 [68] K. Simonyan and A. Zisserman, �Very deep convolutional networks for large-scale image recognition,� Computing Res. Repository, vol. abs/1409.1556, 2014. [Online]. Available: https://arxiv.org/abs/1409.1556 [69] K. He, X. Zhang, S. Ren, and J. Sun, �Deep residual learning for image recogni.tion,� arXiv Preprint, arXiv:1512.03385, 2015. [70] Y. Cheng, F. X. Yu, R. S. Feris, S. Kumar, A. N. Choudhary, and S. Chang, �An exploration of parameter redundancy in deep networks with circulant projections,� in Proc. IEEE Int. Conf. Computer Vision, 2015, pp. 2857�2865. [71] M. Moczulski, M. Denil, J. Appleyard, and N. de Freitas, �ACDC: A structured efficient linear layer,� in Proc. Int. Conf. Learning Representations, 2016. [72] M. Andrychowicz, M. Denil, S. G. Colmenarejo, M. W. Hoffman, D. Pfau, T. Schaul, and N. de Freitas, �Learning to learn by gradient descent by gradient descent,� in Proc. Neural Information Processing Systems Conf., 2016, pp. 3981� 3989. [73] D. Ha, A. Dai, and Q. Le, �Hypernetworks,� in Proc. Int. Conf. Learning Representations, 2016. [74] J. M. Alvarez and M. Salzmann, �Learning the number of neurons in deep net.works,� in Proc. Neural Information Processing Systems Conf., 2016, pp. 2270� 2278. [75] Y. Wang, C. Xu, C. Xu, and D. Tao, �Beyond filters: Compact feature map for portable deep model,� in Proc. 34th Int. Conf. Machine Learning, 2017, pp. 3703� 3711. [76] Y.-D. Kim, E. Park, S. Yoo, T. Choi, L. Yang, and D. Shin, �Compression of deep convolutional neural networks for fast and low power mobile applications,� Computing Res. Repository, vol. abs/1511.06530, 2015. [Online]. Available: https://arxiv.org/ abs/1511.06530 [77] Facebook, Inc. Caffe2: A new lightweight, modular, and scalable deep learning framework. (2016). [Online]. Available: https://caffe2.ai/ <> <> <> <> <> <> MOGRIFIER LSTM Gábor Melis y , Tomáš Kociskýˇ y , Phil Blunsom yz {melisgl,tkocisky,pblunsom}@google.com y DeepMind, London, UK z University of Oxford ABSTRACT Many advances in Natural Language Processing have been based upon more expressive models for how inputs interact with the context in which they occur. Recurrent networks, which have enjoyed a modicum of success, still lack the generalization and systematicity ultimately required for modelling language. In this work, we propose an extension to the venerable Long Short-Term Memory in the form of mutual gating of the current input and the previous output. This mechanism affords the modelling of a richer space of interactions between inputs and their context. Equivalently, our model can be viewed as making the transition function given by the LSTM context-dependent. Experiments demonstrate markedly improved generalization on language modelling in the range of 3–4 perplexity points on Penn Treebank and Wikitext-2, and 0.01–0.05 bpc on four character-based datasets. We establish a new state of the art on all datasets with the exception of Enwik8, where we close a large gap between the LSTM and Transformer models. 1 INTRODUCTION The domination of Natural Language Processing by neural models is hampered only by their limited ability to generalize and questionable sample complexity (Belinkov and Bisk 2017; Jia and Liang 2017; Iyyer et al. 2018; Moosavi and Strube 2017; Agrawal et al. 2016), their poor grasp of grammar (Linzen et al. 2016; Kuncoro et al. 2018), and their inability to chunk input sequences into meaningful units (Wang et al. 2017). While direct attacks on the latter are possible, in this paper, we take a language-agnostic approach to improving Recurrent Neural Networks (RNN, Rumelhart et al. (1988)), which brought about many advances in tasks such as language modelling, semantic parsing, machine translation, with no shortage of non-NLP applications either (Bakker 2002; Mayer et al. 2008). Many neural models are built from RNNs including the sequence-to-sequence family (Sutskever et al. 2014) and its attention-based branch (Bahdanau et al. 2014). Thus, innovations in RNN architecture tend to have a trickle-down effect from language modelling, where evaluation is often the easiest and data the most readily available, to many other tasks, a trend greatly strengthened by ULMFiT (Howard and Ruder 2018), ELMo (Peters et al. 2018) and BERT (Devlin et al. 2018), which promote language models from architectural blueprints to pretrained building blocks. To improve the generalization ability of language models, we propose an extension to the LSTM (Hochreiter and Schmidhuber 1997), where the LSTM’s inputxis gated conditioned on the output of the previous step h_prev . Next, the gated input is used in a similar manner to gate the output of the previous time step. After a couple of rounds of this mutual gating, the last update dx and h_prev are fed to an LSTM. By introducing these additional of gating operations, in one sense, our model joins the long list of recurrent architectures with gating structures of varying complexity which followed the invention of Elman Networks (Elman 1990). Examples include the LSTM, the GRU (Chung et al. 2015), and even designs by Neural Architecture Search (Zoph and Le 2016). Intuitively, in the lowermost layer, the first gating step scales the input embedding (itself a representation of the average context in which the token occurs) depending on the actual context, resulting in a contextualized representation of the input. While intuitive, as Section4 shows, this interpretation cannot account for all the observed phenomena. In a more encompassing view, our model can be seen as enriching the mostly additive dynamics of recurrent transitions placing it in the company of the Input Switched Affine Network (Foerster et al. <

> Figure 1: Mogrifier with 5 rounds of updates. The previous state h0 =h prev is transformed linearly (dashed arrows), fed through a sigmoid and gates <> in an element-wise manner producing <>. Conversely, the linearly transformed <> gates <> and produces <>. After a number of repetitions of this mutual gating cycle, the last values of h and x sequences are fed to an LSTM cell. The prev subscript of his omitted to reduce clutter. 2017) with a separate transition matrix for each possible input, and the Multiplicative RNN (Sutskever et al. 2011), which factorizes the three-way tensor of stacked transition matrices. Also following this line of research are the Multiplicative Integration LSTM (Wu et al. 2016) and – closest to our model in the literature – the Multiplicative LSTM (Krause et al. 2016). The results in Section3.4 demonstrate the utility of our approach, which consistently improves on the LSTM and establishes a new state of the art on all but the largest dataset, Enwik8, where we match similarly sized transformer models. 2 MODEL To allow for ease of subsequent extension, we present the standard LSTM update (Sak et al. 2014) with input and state of size m and n respectively as the following function: <> The updated state c and the output h are computed as follows: <> where <> is the logistic sigmoid function, <> is the elementwise product,<> and b are weight matrices and biases. While the LSTM is typically presented as a solution to the vanishing gradients problem, its gate i can also be interpreted as scaling the rows of weight matrices <> (ignoring the non-linearity in j). In this sense, the LSTM nudges Elman Networks towards context-dependent transitions and the extreme case of Input Switched Affine Networks. If we took another, larger step towards that extreme, we could end up with Hypernetworks (Ha et al. 2016). Here, instead, we take a more cautious step, and equip the LSTM with gates that scale the columns of all its weight matrices <> in a context-dependent manner. The scaling of the matrices <> (those that transform the cell input) makes the input embeddings dependent on the cell state, while the scaling of <> does the reverse. The Mogrifier LSTM is an LSTM where two inputs x and h_prev modulate one another in an alternating fashion before the usual LSTM computation takes place (see Fig.1). That is, <> where the modulated inputs x" and h" are prev defined as the highest indexed xi and hi , respectively, from the interleaved sequences <> (1) 1 It’s like a transmogrifier 2 without the magic: it can only shrink or expand objects. 2 Transmogrify (verb, 1650s): to completely alter the form of something in a surprising or magical manner. <> (2) with <>, is a hyperparameter; <>. The number of “rounds”,r recovers the LSTM. Multiplication with the constant2ensures that randomly initializedQi ;Ri matrices result in transformations close to identity. To reduce the number of additional model parameters, we typically factorize theQi ;Ri matrices as products of low-rank matrices: <> with <>, where <> is the rank. 3 EXPERIMENTS 3.1 THE CASE FOR SMALL-SCALE Before describing the details of the data, the experimental setup and the results, we take a short detour to motivate work on smaller-scale datasets. A recurring theme in the history of sequence models is that the problem of model design is intermingled with optimizability and scalability. Elman Networks are notoriously difficult to optimize, a property that ultimately gave birth to the idea of the LSTM, but also to more recent models such as the Unitary Evolution RNN (Arjovsky et al. 2016) and fixes like gradient clipping (Pascanu et al. 2013). Still, it is far from clear – if we could optimize these models well – how different their biases would turn out to be. The non-separability of model and optimization is fairly evident in these cases. Scalability, on the other hand, is often optimized for indirectly. Given the limited ability of current models to generalize, we often compensate by throwing more data at the problem. To fit a larger dataset, model size must be increased. Thus the best performing models are evaluated based on their scalability 3 . Today, scaling up still yields tangible gains on down-stream tasks, and language modelling data is abundant. However, we believe that simply scaling up will not solve the generalization problem and better models will be needed. Our hope is that by choosing small enough datasets, so that model size is no longer the limiting factor, we get a number of practical advantages: Generalization ability will be more clearly reflected in evaluations even without domain adaptation. Turnaround time in experiments will be reduced, and the freed up computational budget can be put to good use by controlling for nuisance factors. The transient effects of changing hardware performance characteristics are somewhat lessened. Thus, we develop, analyse and evaluate models primarily on small datasets. Evaluation on larger datasets is included to learn more about the models’ scaling behaviour and because of its relevance for applications, but it is to be understood that these evaluations come with much larger error bars and provide more limited guidance for further research on better models. 3.2 DATASETS We compare models on both word and character-level language modelling datasets. The two word- level datasets we picked are the Penn Treebank (PTB) corpus by Marcus et al. (1993) with prepro- cessing from Mikolov et al. (2010) and Wikitext-2 by Merity et al. (2016), which is about twice the size of PTB with a larger vocabulary and lighter preprocessing. These datasets are definitely on the small side, but – and because of this – they are suitable for exploring different model biases. Their main shortcoming is the small vocabulary size, only in the tens of thousands, which makes them inappropriate for exploring the behavior of the long tail. For that, open vocabulary language modelling and byte pair encoding (Sennrich et al. 2015) would be an obvious choice. Still, our primary goal here is the comparison of the LSTM and Mogrifier architectures, thus we instead opt for character-based language modelling tasks, where vocabulary size is not an issue, the long tail is not truncated, and there are no additional hyperparameters as in byte pair encoding that make fair comparison harder. The first character-based corpus is Enwik8 from the Hutter Prize dataset (Hutter 2012). Following common practice, we use the first 90 million characters for training and the remaining 10 million evenly split between validation and test. The character-level task on the 3 Note that the focus on scalability is not a problem per se. Indeed the unsupervised pretraining methods (Peters et al. 2018; Devlin et al. 2018) take great advantage of this approach. Table 1: Word-level perplexities of near state-of-the-art models, ourLSTMbaseline and theMogrifieron PTB and Wikitext-2. Models with Mixture of Softmaxes (Yang et al. 2017) are denoted withMoS, depth N withdN. MCstands for Monte-Carlo dropout evaluation. Previous state-of-the-art results in italics. Note the comfortable margin of 2.8–4.3 perplexity points the Mogrifier enjoys over the LSTM. <

> Mikolov preprocessed PTB corpus (Merity et al. 2018) is unique in that it has the disadvantages of closed vocabulary without the advantages of word-level modelling, but we include it for comparison to previous work. The final character-level dataset is the Multilingual Wikipedia Corpus (MWC, Kawakami et al. (2017)), from which we focus on the English and Finnish language subdatasets in the single text, large setting. 3.3 SETUP We tune hyperparameters following the experimental setup of Melis et al. (2018) using a black-box hyperparameter tuner based on batched Gaussian Process Bandits (Golovin et al. 2017). For the LSTM, the tuned hyperparameters are the same:input_embedding_ratio,learning_rate,l2_penalty, input_dropout,inter_layer_dropout,state_dropout,output_dropout. For the Mogrifier, the number of rounds r and the rank k of the low-rank approximation is also tuned (allowing for full rank, too). For word-level tasks, BPTT (Werbos et al. 1990) window size is set to 70 and batch size to 64. For character-level tasks, BPTT window size is set to 150 and batch size to 128 except for Enwik8 where the window size is 500. Input and output embeddings are tied for word-level tasks following Inan et al. (2016) and Press and Wolf (2016). Optimization is performed with Adam (Kingma and Ba 2014) with <>, a setting that resembles RMSProp without momentum. Gradients are clipped (Pascanu et al. 2013) to norm 10. We switch to averaging weights similarly to Merity et al. (2017) after a certain number of checkpoints with no improvement in validation cross-entropy or at 80% of the training time at the latest. We found no benefit to using two-step finetuning. Model evaluation is performed with the standard, deterministic dropout approximation or Monte- Carlo averaging (Gal and Ghahramani 2016) where explicitly noted (MC). In standard dropout evaluation, dropout is turned off while in MC dropout predictions are averaged over randomly sampled dropout masks (200 in our experiments). Optimal softmax temperature is determined on the validation set, and in the MC case dropout rates are scaled (Melis et al. 2018). Finally, we report results with and without dynamic evaluation (Krause et al. 2017). Hyperparameters for dynamic evaluation are tuned using the same method (see AppendixA for details). We make the code and the tuner output available at https://github.com/deepmind/lamb. 3.4 RESULTS Table1 lists our results on word-level datasets. On the PTB and Wikitext-2 datasets, the Mogrifier has lower perplexity than the LSTM by 3–4 perplexity points regardless of whether or not dynamic evaluation (Krause et al. 2017) and Monte-Carlo averaging are used. On both datasets, the state of the art is held by the AWD LSTM (Merity et al. 2017) extended with Mixture of Softmaxes (Yang Table 2: Bits per character on character-based datasets of near state-of-the-art models, our LSTM baseline and theMogrifier. Previous state-of-the-art results in italics. Depth N is denoted withdN. MC stands for Monte-Carlo dropout evaluation. Once again the Mogrifier strictly dominates the LSTM and sets a new state of the art on all but the Enwik8 dataset where with dynamic evaluation it closes the gap to the Transformer-XL of similar size (y Krause et al. (2019),zBen Krause, personal communications, May 17, 2019). On most datasets, model size was set large enough for underfitting not to be an issue. This was very much not the case with Enwik8, so we grouped models of similar sizes together for ease of comparison. Unfortunately, a couple of dynamic evaluation test runs diverged (NaN) on the test set and some were just too expensive to run (Enwik8, MC). <> et al. 2017) and FRAGE (Gong et al. 2018). The Mogrifier improves the state of the art without either of these methods on PTB, and without FRAGE on Wikitext-2. Table2 lists the character-level modelling results. On all datasets, our baseline LSTM results are much better than those previously reported for LSTMs, highlighting the issue of scalability and experimental controls. In some cases, these unexpectedly large gaps may be down to lack of hyperparameter tuning as in the case of Merity et al. (2017), or in others, to using a BPTT window size (50) that is too small for character-level modelling (Melis et al. 2017) in order to fit the model into memory. The Mogrifier further improves on these baselines by a considerable margin. Even the smallest improvement of 0.012 bpc on the highly idiosyncratic, character-based, Mikolov preprocessed PTB task is equivalent to gaining about 3 perplexity points on word-level PTB. MWC, which was built for open-vocabulary language modelling, is a much better smaller-scale character-level dataset. On the English and the Finnish corpora in MWC, the Mogrifier enjoys a gap of 0.033-0.046 bpc. Finally, on the Enwik8 dataset, the gap is 0.029-0.039 bpc in favour of the Mogrifier. <

> Figure 2: “No-zigzag” Mogrifier for the ablation study. Gating is always based on the original inputs. Table 3: PTB ablation study validation perplexities with 24M parameters. <

> Of particular note is the comparison to Transformer-XL (Dai et al. 2019), a state-of-the-art model on larger datasets such as Wikitext-103 and Enwik8. On PTB, without dynamic evaluation, the Transformer-XL is on par with our LSTM baseline which puts it about 3.5 perplexity points behind the Mogrifier. On Enwik8, also without dynamic evaluation, the Transformer-XL has a large, 0.09 bpc advantage at similar parameter budgets, but with dynamic evaluation this gap disappears. However, we did not test the Transformer-XL ourselves, so fair comparison is not possible due to differing experimental setups and the rather sparse result matrix for the Transformer-XL. 4 ANALYSIS 4.1 ABLATION STUDY The Mogrifier consistently outperformed the LSTM in our experiments. The optimal settings were similar across all datasets, with <> and <> (see AppendixB for a discussion of hyperparameter sensitivity). In this section, we explore the effect of these hyperparameters and show that the proposed model is not unnecessarily complicated. To save computation, we tune all models using a shortened schedule with only 145 epochs instead of 964 and a truncated BPTT window size of 35 on the word-level PTB dataset, and evaluate using the standard, deterministic dropout approximation with a tuned softmax temperature. Fig.3 shows that the number of rounds r greatly influences the results. Second, we found the low-rank factorization ofQi andRi to help a bit, but the full-rank variant is close behind which is what we observed on other datasets, as well. Finally, to verify that the alternating gating scheme is not overly complicated, we conditional l new ly introduced gates on the original inputs x and h_prev (see Fig.2). That is, instead of Eq.1 and Eq.2 the no-zigzag updates are <> In our experiments, the no-zigzag variant underperformed the baseline Mogrifier by a small but significant margin, and was on par with the <> model in Fig.3 suggesting that the Mogrifier’s iterative refinement scheme does more than simply widen the range of possible gating values ofx and h_prev to(0;2dr=2e )and(0;2br=2c ), respectively. 4.2 COMPARISON TO THE M LSTM The Multiplicative LSTM (Krause et al. 2016), or mLSTM for short, is closest to our model in the literature. It is defined asmLSTM(x;cprev ; h_prev ) = LSTM(x;cprev ;hm ), wherehm =prev prev <

> Figure 4: Cross-entropy vs sequence length in the reverse copy task with i.i.d. tokens. Lower is better. The Mogrifier is better than the LSTM even in this synthetic task with no resemblance to natural language. <>. In this formulation, the differences are readily apparent. First, the mLSTM allows for multiplicative interaction betweenxand h_prev , but it only overrides h_prev , while in the Mogrifier the interaction is two-way, which – as the ablation study showed – is important. Second, the mLSTM can change not only the magnitude but also the sign of values in h_prev , something with which we experimented in the Mogrifier, but could not get to work. Furthermore, in the definition of hm , the unsquashed linearities and their elementwise product make the mLSTM more sensitive to prev initialization and unstable during optimization. On the Enwik8 dataset, we greatly improved on the published results of the mLSTM (Krause et al. 2016). In fact, even our LSTM baseline outperformed the mLSTM by 0.03 bpc. We also conducted experiments on PTB based on our reimplementation of the mLSTM following the same methodology as the ablation study and found that the mLSTM did not improve on the LSTM (see Table3). Krause et al. (2016) posit and verify the recovery hypothesis which says that having just suffered a large loss, the loss on the next time step will be smaller on average for the mLSTM than for the LSTM. This was found not to be the case for the Mogrifier. Neither did we observe a significant change in the gap between the LSTM and the Mogrifier in the tied and untied embeddings settings, which would be expected if recovery was affected byxand h_prev being in different domains. 4.3 THE REVERSE COPY TASK Our original motivation for the Mogrifier was to allow the context to amplify salient and attenuate nuisance features in the input embeddings. We conduct a simple experiment to support this point of view. Consider the reverse copy task where the network reads an input sequence of tokens and a marker token after which it has to repeat the input in reverse order. In this simple sequence-to- sequence learning (Sutskever et al. 2014) setup, the reversal is intended to avoid the minimal time lag problem (Hochreiter and Schmidhuber 1997), which is not our focus here. The experimental setup is as follows. For the training set, we generate 500000 examples by uniformly sampling a given number of tokens from a vocabulary of size1000. The validation and test sets are constructed similarly, and contain 10000 examples. The model consists of an independent, unidirectional encoder and a decoder, whose total number of parameters is10million. The decoder is initialized from the last state of the encoder. Since overfitting is not an issue here, no dropout is necessary, and we only tune the learning rate, the l2 penalty, and the embedding size for the LSTM. For the Mogrifier, the number of roundsrand the rankkof the low-rank approximation are also tuned. We compare the case where both the encoder and decoder are LSTMs to where both are Mogrifiers. Fig.4a shows that, for sequences of length 50 and 100, both models can solve the task perfectly. At higher lengths though, the Mogrifier has a considerable advantage. Examining the best hyperparameter settings found, the embedding/hidden sizes for the LSTM and Mogrifier are 498/787 vs 41/1054 at 150 steps, and 493/790 vs 181/961 at 200 steps. Clearly, the Mogrifier was able to work with a much smaller embedding size than the LSTM, which is in line with our expectations for a model with a more flexible interaction between the input and recurrent state. We also conducted experiments with a larger model and vocabulary size, and found the effect even more pronounced (see Fig.4b). 4.4 WHAT THE MOGRIFIER IS NOT The results on the reverse copy task support our hypothesis that input embeddings are enriched by the Mogrifier architecture, but that cannot be the full explanation as the results of the ablation study indicate. In the following, we consider a number of hypotheses about where the advantage of the Mogrifier lies and the experiments that provide evidenceagainstthem. E Hypothesis: the benefit is in scalingxand h_prev .We verified that data dependency is a crucial feature by adding a learnable scaling factor to the LSTM inputs. We observed no improvement. Also, at extremely low-rank (less than 5) settings where the amount of information in its gating is small, the Mogrifier loses its advantage. E Hypothesis: the benefit is in making optimization easier.We performed experiments with different optimizers (SGD, RMSProp), with intra-layer batch normalization and layer normalization on the LSTM gates. While we cannot rule out an effect on optimization difficulty, in all of these experiments the gap between the LSTM and the Mogrifier was the same. E Hypothesis: exact tying of embeddings is too constraining, the benefit is in making this rela- tionship less strict.Experiments conducted with untied embeddings and character-based models demonstrate improvements of similar magnitude. E Hypothesis: the benefit is in the low-rank factorization of <> implicitly imposing structure on the LSTM weight matrices.We observed that the full-rank Mogrifier also performed better than the plain LSTM. We conducted additional experiments where the LSTM’s gate matrices were factorized and observed no improvement. E Hypothesis: the benefit comes from better performance on rare words.The observed advantage on character-based modelling is harder to explain based on frequency. Also, in the reverse copy experiments, a large number of tokens were sampled uniformly, so there were no rare words at all. E Hypothesis: the benefit is specific to the English language.This is directly contradicted by the Finnish MWC and the reverse copy experiments. E Hypothesis: the benefit is in handling long-range dependencies better.Experiments in the episodic setting (i.e. sentence-level language modelling) exhibited the same gap as the non-episodic ones. E Hypothesis: the scaling up of inputs saturates the downstream LSTM gates.The idea here is that saturated gates may make states more stable over time. We observed the opposite: the means of the standard LSTM gates in the Mogrifier were very close between the two models, but their variance was smaller in the Mogrifier. 5 CONCLUSIONS AND FUTURE WORK We presented the Mogrifier LSTM, an extension to the LSTM, with state-of-the-art results on several language modelling tasks. Our original motivation for this work was that the context-free representation of input tokens may be a bottleneck in language models and by conditioning the input embedding on the recurrent state some benefit was indeed derived. While it may be part of the explanation, this interpretation clearly does not account for the improvements brought by conditioning the recurrent state on the input and especially the applicability to character-level datasets. Positioning our work on the Multiplicative RNN line of research offers a more compelling perspective. To give more credence to this interpretation, in the analysis we highlighted a number of possible alternative explanations, and ruled them all out to varying degrees. In particular, the connection to the mLSTM is weaker than expected as the Mogrifier does not exhibit improved recovery (see Section4.2), and on PTB the mLSTM works only as well as the LSTM. At the same time, the evidence against easier optimization is weak, and the Mogrifier establishing some kind of sharing between otherwise independent LSTM weight matrices is a distinct possibility. Finally, note that as shown by Fig.1 and Eq.1-2, the Mogrifier is a series of preprocessing steps composed with the LSTM function, but other architectures, such as Mogrifier GRU or Mogrifier Elman Network are possible. We also leave investigations into other forms of parameterization of context-dependent transitions for future work. ACKNOWLEDGMENTS We would like to thank Ben Krause for the Transformer-XL dynamic evaluation results, Laura Rimell, Aida Nematzadeh, Angeliki Lazaridou, Karl Moritz Hermann, Daniel Fried for helping with experiments, Chris Dyer, Sebastian Ruder and Jack Rae for their valuable feedback. REFERENCES Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. InInternational Conference on Machine Learning, pages 1120–1128, 2016. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate.arXiv preprint arXiv:1409.0473, 2014. Shaojie Bai, J Zico Kolter, and Vladlen Koltun. Trellis networks for sequence modeling. arXiv preprint arXiv:1810.06682, 2018. Bram Bakker. Reinforcement learning with long short-term memory. InAdvances in neural information processing systems, pages 1475–1482, 2002. Yonatan Belinkov and Yonatan Bisk. Synthetic and natural noise both break neural machine translation.arXiv preprint arXiv:1711.02173, 2017. Junyoung Chung, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. Gated feedback recurrent neural networks. InInternational Conference on Machine Learning, pages 2067–2075, 2015. Zihang Dai, Zhilin Yang, Yiming Yang, William W Cohen, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context.arXiv preprint arXiv:1901.02860, 2019. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2018. Jeffrey L Elman. Finding structure in time.Cognitive science, 14(2):179–211, 1990. Jakob N Foerster, Justin Gilmer, Jascha Sohl-Dickstein, Jan Chorowski, and David Sussillo. Input switched affine networks: An rnn architecture designed for interpretability. InProceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1136–1145. JMLR. org, 2017. Yarin Gal and Zoubin Ghahramani. A theoretically grounded application of dropout in recurrent neural networks. InAdvances in Neural Information Processing Systems, pages 1019–1027, 2016. Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and D Sculley. Google vizier: A service for black-box optimization. InProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1487–1495. ACM, 2017. Chengyue Gong, Di He, Xu Tan, Tao Qin, Liwei Wang, and Tie-Yan Liu. Frage: frequency-agnostic word representation. InAdvances in Neural Information Processing Systems, pages 1334–1345, 2018. David Ha, Andrew Dai, and Quoc V Le. Hypernetworks.arXiv preprint arXiv:1609.09106, 2016. Sepp Hochreiter and Jürgen Schmidhuber. Lstm can solve hard long time lag problems. InAdvances in neural information processing systems, pages 473–479, 1997. Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification.arXiv preprint arXiv:1801.06146, 2018. Marcus Hutter. The human knowledge compression contest.URL http://prize. hutter1. net, 6, 2012. Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A loss framework for language modeling.CoRR, abs/1611.01462, 2016. URLhttp://arxiv.org/abs/1611.01462. Mohit Iyyer, John Wieting, Kevin Gimpel, and Luke Zettlemoyer. Adversarial example generation with syntactically controlled paraphrase networks.arXiv preprint arXiv:1804.06059, 2018. Robin Jia and Percy Liang. Adversarial examples for evaluating reading comprehension systems.arXiv preprint arXiv:1707.07328, 2017. Kazuya Kawakami, Chris Dyer, and Phil Blunsom. Learning to create and reuse words in open-vocabulary neural language modeling.arXiv preprint arXiv:1704.06986, 2017. Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014. Ben Krause, Liang Lu, Iain Murray, and Steve Renals. Multiplicative LSTM for sequence modelling.CoRR, abs/1609.07959, 2016. URLhttp://arxiv.org/abs/1609.07959. Ben Krause, Emmanuel Kahembwe, Iain Murray, and Steve Renals. Dynamic evaluation of neural sequence models.arXiv preprint arXiv:1709.07432, 2017. Ben Krause, Emmanuel Kahembwe, Iain Murray, and Steve Renals. Dynamic evaluation of transformer language models.arXiv preprint arXiv:1904.08378, 2019. Adhiguna Kuncoro, Chris Dyer, John Hale, Dani Yogatama, Stephen Clark, and Phil Blunsom. Lstms can learn syntax-sensitive dependencies well, but modeling structure makes them better. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1426–1436, 2018. Tal Linzen, Emmanuel Dupoux, and Yoav Goldberg. Assessing the ability of lstms to learn syntax-sensitive dependencies.Transactions of the Association for Computational Linguistics, 4:521–535, 2016. Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The Penn treebank.Computational linguistics, 19(2):313–330, 1993. Hermann Mayer, Faustino Gomez, Daan Wierstra, Istvan Nagy, Alois Knoll, and Jürgen Schmidhuber. A system for robotic heart surgery that learns to tie knots using recurrent neural networks.Advanced Robotics, 22 (13-14):1521–1537, 2008. Gábor Melis, Chris Dyer, and Phil Blunsom. On the state of the art of evaluation in neural language models. arXiv preprint arXiv:1707.05589, 2017. Gábor Melis, Charles Blundell, Tomáš Kociskˇ y, Karl Moritz Hermann, Chris Dyer, and Phil Blunsom. Pushing` the bounds of dropout.arXiv preprint arXiv:1805.09208, 2018. Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models.CoRR, abs/1609.07843, 2016. URLhttp://arxiv.org/abs/1609.07843. Stephen Merity, Nitish Shirish Keskar, and Richard Socher. Regularizing and optimizing lstm language models. arXiv preprint arXiv:1708.02182, 2017. Stephen Merity, Nitish Shirish Keskar, and Richard Socher. An analysis of neural language modeling at multiple scales.arXiv preprint arXiv:1803.08240, 2018. Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocky, and Sanjeev Khudanpur. Recurrent neural` network based language model. InInterspeech, volume 2, page 3, 2010. Nafise Sadat Moosavi and Michael Strube. Lexical features in coreference resolution: To be used with caution. arXiv preprint arXiv:1704.06779, 2017. Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. In International conference on machine learning, pages 1310–1318, 2013. Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations.arXiv preprint arXiv:1802.05365, 2018. Ofir Press and Lior Wolf. Using the output embedding to improve language models.CoRR, abs/1608.05859, 2016. URL http://arxiv.org/abs/1608.05859. David E Rumelhart, Geoffrey E Hinton, Ronald J Williams, et al. Learning representations by back-propagating errors.Cognitive modeling, 5(3):1, 1988. Hasim Sak, Andrew W. Senior, and Françoise Beaufays. Long short-term memory based recurrent neural network architectures for large vocabulary speech recognition.CoRR, abs/1402.1128, 2014. URL http://arxiv.org/abs/1402.1128. Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units.arXiv preprint arXiv:1508.07909, 2015. Ilya Sutskever, James Martens, and Geoffrey E Hinton. Generating text with recurrent neural networks. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 1017–1024, 2011. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pages 3104–3112, 2014. Chong Wang, Yining Wang, Po-Sen Huang, Abdelrahman Mohamed, Dengyong Zhou, and Li Deng. Sequence modeling via segmentations. InProceedings of the 34th International Conference on Machine Learning- Volume 70, pages 3674–3683. JMLR. org, 2017. Paul J Werbos et al. Backpropagation through time: what it does and how to do it.Proceedings of the IEEE, 78 (10):1550–1560, 1990. Yuhuai Wu, Saizheng Zhang, Ying Zhang, Yoshua Bengio, and Ruslan R Salakhutdinov. On multiplicative integration with recurrent neural networks. InAdvances in neural information processing systems, pages 2856–2864, 2016. Zhilin Yang, Zihang Dai, Ruslan Salakhutdinov, and William W Cohen. Breaking the softmax bottleneck: a high-rank rnn language model.arXiv preprint arXiv:1711.03953, 2017. Barret Zoph and Quoc V. Le. Neural architecture search with reinforcement learning.CoRR, abs/1611.01578, 2016. URLhttp://arxiv.org/abs/1611.01578. APPENDIX A HYPERPARAMETER TUNING RANGES In all experiments, we tuned hyperparameters using Google Vizier (Golovin et al. 2017). The tuning ranges are listed in Table4. Obviously,mogrifier_roundsandmogrifier_rankare tuned only for the Mogrifier. Ifinput_embedding_ratio>1, then the input/output embedding sizes and the hidden sizes are set to equal and the linear projection from the cell output into the output embeddings space is omitted. Similarly,mogrifier_rank60is taken to mean full rank <> without factorization. Since Enwik8 is a much larger dataset, we don’t tuneinput_embedding_ratioand specify tighter tuning ranges for dropout based on preliminary experiments (see Table5). Dynamic evaluation hyperparameters were tuned according to Table6. The highest possible value formax_time_steps, the BPTT window size, was 20 for word, and 50 for character-level tasks. The batch size for estimating the mean squared gradients over the training data was set to 1024, gradient clipping was turned off, and the l2 penalty was set to zero. Table 4: Hyperparameter tuning ranges for all tasks except Enwik8. <

> Table 5: Hyperparameter tuning ranges for Enwik8. <> Table 6: Hyperparameter tuning ranges for dynamic evaluation. <> APPENDIX B HYPERPARAMETER SENSITIVITY The parallel coordinate plots in Fig.5 and 6, give a rough idea about hyperparameter sensitivity. The red lines correspond to hyperparameter combinations closest to the best solution found. To find the closest combinations, we restricted the range for each hyperparameter separately to about 15% of its entire tuning range. For both the LSTM and the Mogrifier, the results are at most 1.2 perplexity points off the best result, so our results are somewhat insensitive to jitter in the hyperparameters. Still, in this setup, grid search would require orders of magnitude more trials to find comparable solutions. On the other hand, the tuner does take advantage of the stochasticity of training, and repeated runs with the same parameters may be give slightly worse results. To gauge the extent of this effect, on PTB we estimated the standard deviation in reruns of the LSTM with the best hyperparameters to be about 0.2 perplexity points, but the mean was about 0.7 perplexity points off the result produced with the weights saved in best tuning run. <

> Figure 5: Average per-word validation cross-entropies for hyperparameter combinations in the neighbourhood of the best solution for a 2-layer LSTM with 24M weights on the Penn Treebank dataset. <

> Figure 6: Average per-word validation cross-entropies for hyperparameter combinations in the neighbourhood of the best solution for a 2-layer Mogrifier LSTM with 24M weights on the Penn Treebank dataset. feature_mask_rank and feature_mask_roundsare aliases for mogrifier_rank and mogrifier_rounds <> <> <> <> <> <> Movement Pruning: Adaptive Sparsity by Fine-Tuning Victor Sanh 1 , Thomas Wolf 1 , Alexander M. Rush 1,2 1 Hugging Face, 2 Cornell University {victor,thomas}@huggingface.co;arush@cornell.edu Abstract Magnitude pruning is a widely used strategy for reducing model size in pure supervised learning; however, it is less effective in the transfer learning regime that has become standard for state-of-the-art natural language processing applications. We propose the use of movement pruning, a simple, deterministic first-order weight pruning method that is more adaptive to pretrained model fine-tuning. We give mathematical foundations to the method and compare it to existing zeroth- and first-order pruning methods. Experiments show that when pruning large pretrained language models, movement pruning shows significant improvements in high- sparsity regimes. When combined with distillation, the approach achieves minimal accuracy loss with down to only 3% of the model parameters. 1 Introduction Large-scale transfer learning has become ubiquitous in deep learning and achieves state-of-the-art performance in applications in natural language processing and related fields. In this setup, a large model pretrained on a massive generic dataset is then fine-tuned on a smaller annotated dataset to perform a specific end-task. Model accuracy has been shown to scale with the pretrained model and dataset size [Raffel et al., 2019]. However, significant resources are required to ship and deploy these large models, and training the models have high environmental costs [Strubell et al., 2019]. Sparsity induction is a widely used approach to reduce the memory footprint of neural networks at only a small cost of accuracy. Pruning methods, which remove weights based on their importance, are a particularly simple and effective method for compressing models to be sent to edge devices such as mobile phones. Magnitude pruning [Han et al., 2015, 2016], which preserves weights with high absolute values, is the most widely used method for weight pruning. It has been applied to a large variety of architectures in computer vision [Guo et al., 2016], in language processing [Gale et al., 2019], and more recently has been leveraged as a core component in the lottery ticket hypothesis [Frankle et al., 2019]. While magnitude pruning is highly effective for standard supervised learning, it is inherently less useful in the transfer learning regime. In supervised learning, weight values are primarily determined by the end-task training data. In transfer learning, weight values are mostly predetermined by the original model and are only fine-tuned on the end task. This prevents these methods from learning to prune based on the fine-tuning step, or “fine-pruning.” In this work, we argue that to effectively reduce the size of models for transfer learning, one should instead use movement pruning, i.e., pruning approaches that consider the changes in weights during fine-tuning. Movement pruning differs from magnitude pruning in that both weights with low and high values can be pruned if they shrink during training. This strategy moves the selection criteria from the 0th to the 1st-order and facilitates greater pruning based on the fine-tuning objective. To test this approach, we introduce a particularly simple, deterministic version of movement pruning utilizing the straight-through estimator [Bengio et al., 2013]. We apply movement pruning to pretrained language representations (BERT) [Devlin et al., 2019, Vaswani et al., 2017] on a diverse set of fine-tuning tasks. In highly sparse regimes (less than 15% of remaining weights), we observe significant improvements over magnitude pruning and other 1st-order methods such asL0 regularization [Louizos et al., 2017]. Our models reach 95% of the original BERT performance with only 5% of the encoder’s weight on natural language inference (MNLI) [Williams et al., 2018] and question answering (SQuAD v1.1) [Rajpurkar et al., 2016]. Analysis of the differences between magnitude pruning and movement pruning shows that the two methods lead to radically different pruned models with movement pruning showing greater ability to adapt to the end-task. 2 Related Work In addition to magnitude pruning, there are many other approaches for generic model weight pruning. Most similar to our approach are methods for using parallel score matrices to augment the weight matrices [Mallya and Lazebnik, 2018, Ramanujan et al., 2020], which have been applied for convolutional networks. Differing from our methods, these methods keep the weights of the model fixed (either from a randomly initialized network or a pre-trained network) and the scores are updated to find a good sparse subnetwork. Many previous works have also explored using higher-order information to select prunable weights. LeCun et al. [1989] and Hassibi et al. [1993] leverage the Hessian of the loss to select weights for deletion. Our method does not require the (possibly costly) computation of second-order derivatives since the importance scores are obtained simply as the by-product of the standard fine-tuning. The is et al. [2018] and Ding et al. [2019] use the absolute value or the square value of the gradient. In contrast, we found it useful to preserve the direction of movement in our algorithm. Compressing pretrained language models for transfer learning is also a popular area of study. Other approaches include knowledge distillation [Sanh et al., 2019, Tang et al., 2019] and structured pruning [Fan et al., 2020a, Michel et al., 2019]. Our core method does not require an external teacher model and targets individual weight. We also show that having a teacher can further improve our approach. Recent work also builds upon iterative magnitude pruning with rewinding [Yu et al., 2020] to train sparse language models from scratch. This differs from our approach which focuses on the fine-tuning stage. Finally, another popular compression approach is quantization. Quantization has been applied to a variety of modern large architectures [Fan et al., 2020b, Zafrir et al., 2019, Gong et al., 2014] providing high memory compression rates at the cost of no or little performance. As shown in previous works [Li et al., 2020, Han et al., 2016] quantization and pruning are complimentary and can be combined to further improve the performance/size ratio. 3 Background: Score-Based Pruning We first establish shared notation for discussing different neural network pruning strategies. Let <> refer to a generic weight matrix in the model (we consider square matrices, but they could be of any shape). To determine which weights are pruned, we introduce a parallel matrix of associated importance scores <>. Given importance scores, each pruning strategy computes a mask <>. Inference for an input x becomes <>, where <> is the Hadamard product. A common strategy is to keep the top-v percent of weights by importance. We define <> as a function which selects the v% highest values in <> (1) Magnitude-based weight pruning determines the mask based on the absolute value of each weight <> as a measure of importance. Formally, we have importance scores <>, and masks <>. There are several extensions to this base setup. Han et al. [2015] use v iterative magnitude pruning: the model is first trained until convergence and weights with the lowest magnitudes are removed afterward. The sparsified model is then re-trained with the removed weights fixed to 0. This loop is repeated until the desired sparsity level is reached. <> Table 1: Summary of the pruning methods considered in this work and their specificities. The expression of <> regularization is detailed in Eq (3). In this study, we focus on automated gradual pruning[Zhu and Gupta, 2018]. It supplements magnitude pruning by allowing masked weights to be updated such that they are not fixed for the entire duration of the training. Automated gradual pruning enables the model to recover from previous masking choices [Guo et al., 2016]. In addition, one can gradually increases the sparsity level <> during training using a cubic sparsity scheduler: <>. The sparsity <> level at time step <> is increased from an initial value vi (usually 0) to a final value vf in n pruning steps after ti steps of warm-up. The model is thus pruned and trained jointly. 4 Movement Pruning Magnitude pruning can be seen as utilizing zeroth-order information (absolute value) of the running model. In this work, we focus on movement pruning methods where importance is derived from first-order information. Intuitively, instead of selecting weights that are far from zero, we retain connections that are moving away from zero during the training process. We consider two versions of movement pruning: hard and soft. For (hard) movement pruning, masks are computed using the Top v function: <>. Unlike v magnitude pruning, during training, we learn both the weights <> and their importance scores S. During the forward pass, we compute for all <>. Since the gradient of Top v is 0 everywhere it is defined, we follow Ramanujan et al. [2020], Mallya and Lazebnik [2018] and approximate its value with the straight-through estimator [Bengio et al., 2013]. In the backward pass, Top v is ignored and the gradient goes "straight-through" toS. The approximation of gradient of the loss L with respect to <> is given by <> (2) This implies that the scores of weights are updated, even if these weights are masked in the forward pass. We prove in Appendix A.1 that movement pruning as an optimization problem will converge. We also consider a relaxed (soft) version of movement pruning based on the binary mask function described by Mallya and Lazebnik [2018]. Here we replace hyper parameter v with a fixed global threshold value <> that controls the binary mask. The mask is calculated as <>. In order to control the sparsity level, we add a regularization term <> which encourages the importance scores to decrease over time 1 . The coefficient <> controls the penalty intensity and thus the sparsity level. Finally we note that these approaches yield a similar updateL0 regularization based pruning, another movement based pruning approach [Louizos et al., 2017]. Instead of straight-through,L0 uses the hard-concrete distribution, where the maskMis sampled for all <> with hyperparameters <>, <>, and <>: <> The expected <> norm has a closed form involving the parameters of the hard-concrete: E(L0 ) = <>. Thus, the weights and scores of the model can be optimized in <> We also experimented with <> but it turned out to be harder to tune while giving similar results. <> Figure 1: During fine-tuning (on MNLI), the weights stay close to their pre-trained values which limits the adaptivity of magnitude pruning. We plot the identity line in black. Pruned weights are plotted in grey. Magnitude pruning selects weights that are far from 0 while movement pruning selects weights that are moving away from 0. an end-to-end fashion to minimize the sum of the training loss L and the expected L0 penalty. A coefficient l0 controls the L0 penalty and indirectly the sparsity level. Gradients take a similar form: <> (3) At test time, a non-stochastic estimation of the mask is used: <> and weights multiplied by 0 can simply be discarded. <

> Table 1 highlights the characteristics of each pruning method. The main differences are in the masking functions, pruning structure, and the final gradient form. Method Interpretation In movement pruning, the gradient of L with respect to <> is given by the standard gradient derivation: <>. By combining it to Eq(2), we have <> (we omit the binary mask term <> for simplicity). From the gradient update in <> Eq (2), is increasing when <>, which happens in two cases: <> (a) <> (b) It means that during training <> is increasing while being positive or is decreasing while being negative. It is equivalent to saying thatSi;j is increasing when <> is moving away from 0. Inversely, <> is decreasing when @L >0which means that <> is shrinking towards 0. While magnitude pruning selects the most important weights as the ones which maximize their distance to 0 (<>), movement pruning selects the weights which are moving the most away from 0 (<>). For this reason, magnitude pruning can be seen as a 0th order method, whereas movement pruning is based on a 1st order signal. In fact,Scan be seen as an accumulator of movement: from equation (2), after T gradient updates, we have <> (4) Figure 1 shows this difference empirically by comparing weight values during fine-tuning against their pre-trained value. As observed by Gordon et al. [2020], fine-tuned weights stay close in absolute value to their initial pre-trained values. For magnitude pruning, this stability around the pre-trained <

> values implies that we know with high confidence before even fine-tuning which weights will be pruned as the weights with the smallest absolute value at pre-training will likely stay small and be pruned. In contrast, in movement pruning, the pre-trained weights do not have such an awareness of the pruning decision since the selection is made during fine-tuning (moving away from 0), and both low and high values can be pruned. We posit that this is critical for the success of the approach as it is able to prune based on the task-specific data, not only the pre-trained value. 5 Experimental Setup Transfer learning for NLP uses large pre-trained language models that are fine-tuned on target tasks [Ruder et al., 2019, Devlin et al., 2019, Radford et al., 2019, Liu et al., 2019]. We experiment with task- specific pruning ofBERT-base-uncased, a pre-trained model that contains roughly 84M parameters. We freeze the embedding modules and fine-tune the transformer layers and the task-specific head. We perform experiments on three monolingual (English) tasks, which are common benchmarks for the recent progress in transfer learning for NLP: question answering (SQuAD v1.1) [Rajpurkar et al., 2016], natural language inference (MNLI) [Williams et al., 2018], and sentence similarity (QQP) [Iyer et al., 2017]. The datasets respectively contain 8K, 393K, and 364K training examples. SQuAD is formulated as a span extraction task, MNLI and QQP are paired sentence classification tasks. For a given task, we fine-tune the pre-trained model for the same number of updates (between 6 and 10 epochs) across pruning methods 2 . We follow Zhu and Gupta [2018] and use a cubic sparsity scheduling for Magnitude Pruning (MaP), Movement Pruning (MvP), and Soft Movement Pruning (SMvP). Adding a few steps of cool-down at the end of pruning empirically improves the performance especially in high sparsity regimes. The schedule for v is: <> (5) where tf is the number of cool-down steps. We compare our results against several state-of-the-art pruning baselines: Re-weighted Proximal Pruning (RPP) [Guo et al., 2019] combines re-weightedL1 minimization and Proximal Projection [Parikh and Boyd, 2014] to perform unstructured pruning. LayerDrop [Fan et al., 2020a] leverages structured dropout to prune models at test time. For RPP and LayerDrop, we report results from authors. We also compare our method against the mini-BERT models, a collection of smaller BERT models with varying hyper-parameters [Turc et al., 2019]. 6 Results Figure 2 displays the results for the main pruning methods at different levels of pruning on each dataset. First, we observe the consistency of the comparison between magnitude and movement pruning: at low sparsity (more than 70% of remaining weights), magnitude pruning outperforms all methods with little or no loss with respect to the dense model whereas the performance of movement pruning methods quickly decreases even for low sparsity levels. However, magnitude pruning performs poorly with high sparsity, and the performance drops extremely quickly. In contrast, first-order methods show strong performances with less than 15% of remaining weights. Table 2 shows the specific model scores for different methods at high sparsity levels. Magnitude pruning on SQuAD achieves 54.5 F1 with 3% of the weights compared to 73.6 F1 withL0 regularization, 76.3 F1 for movement pruning, and 79.9 F1 with soft movement pruning. These experiments indicate that in high sparsity regimes, importance scores derived from the movement accumulated during fine-tuning induce significantly better pruned models compared to absolute values. Next, we compare the difference in performance between first-order methods. We see that straight- through based hard movement pruning (MvP) is comparable withL0 regularization (with a significant gap in favor of movement pruning on QQP). Soft movement pruning (SMvP) consistently outperforms 2 Preliminary experiments showed that increasing the number of pruning steps tended to improve the end performance Figure 2: Comparisons between different pruning methods in high sparsity regimes.Soft movement pruning consistently outperforms other methods in high sparsity regimes.We plot the performance of the standard fine-tuned BERT along with 95% of its performance. <

> Table 2: Performance at high sparsity levels. (Soft) movement pruning outperforms current state-of-the art pruning methods at different high sparsity levels. <

> hard movement pruning andL0 regularization by a strong margin and yields the strongest performance among all pruning methods in high sparsity regimes. These comparisons support the fact that even if movement pruning (and its relaxed version soft movement pruning) is simpler thanL0 regularization, it yet yields stronger performances for the same compute budget. Finally, movement pruning and soft movement pruning compare favorably to the other baselines, except for QQP where RPP is on par with soft movement pruning. Movement pruning also outperforms the fine-tuned mini-BERT models. This is coherent with [Li et al., 2020]: it is both more efficient and more effective to train a large model and compress it afterward than training a smaller model from scratch. We do note though that current hardware does not support optimized inference for sparse models: from an inference speed perspective, it might often desirable to use a small dense model such as mini-BERT over a sparse alternative of the same size. Distillation further boosts performance Following previous work, we can further leverage knowledge distillation [Bucila et al., 2006, Hinton et al., 2014] to boost performance for free in the pruned domain [Jiao et al., 2019, Sanh et al., 2019] using our baseline fine-tuned BERT-base model as teacher. The training objective is a linear combination of the training loss and a knowledge distillation Figure 3: Comparisons between different pruning methods augmented with distillation. Distillation improves the performance across all pruning methods and sparsity levels. <

> Table 3: Distillation-augmented performances for selected high sparsity levels.All pruning methods benefit from distillation signal further enhancing the ratio Performance VS Model Size. <

> Figure 4: Magnitude pruning and movement pruning leads to pruned models with radically different weight distribution. <

> loss on the output distributions. Figure 3 shows the results on SQuAD, MNLI, and QQP for the three pruning methods boosted with distillation. Overall, we observe that the relative comparisons of the pruning methods remain unchanged while the performances are strictly increased. Table 3 shows for instance that on SQuAD, movement pruning at 10% goes from 81.7 F1 to 84.3 F1. When combined with distillation, soft movement pruning yields the strongest performances across all pruning methods and studied datasets: it reaches 95% of BERT-base with only a fraction of the weights in the encoder (5% on SQuAD and MNLI). 7 Analysis Movement pruning is adaptive Figure 4a compares the distribution of the remaining weights for the same matrix of a model pruned at the same sparsity using magnitude and movement pruning. We observe that by definition, magnitude pruning removes all the weights that are close to zero, ending up with two clusters. In contrast, movement pruning leads to a smoother distribution, which covers the whole interval except for values close to 0. Figure 4b displays each individual weight against its associated importance score in movement pruning. We plot pruned weights in grey. We observe that movement pruning induces no simple relationship between the scores and the weights. Both weights with high absolute value or low absolute value can be considered important. However, high scores are systematically associated with non-zero weights (and thus the “v-shape”). This is coherent with the interpretation we gave to the scores (section 4): a high score S indicates that during fine-tuning, the associated weight moved away from 0 and is thus non-null. Local and global masks perform similarly We study the influence of the locality of the pruning decision. While local Top v selects the v% most important weights matrix by matrix, global Top v uncovers non-uniform sparsity patterns in the network by selecting the v% most important weights in Figure 5: Comparison of local and global selections of weights on SQuAD at different sparsity heavily pruning the highest layers. <

> Figure 6:Remaining weights per layer in the Transformer. Global magnitude pruning tends to levels. For magnitude and movement pruning, prune uniformly layers. Global 1st order meth-local and global Top v performs similarly at all ods allocate the weight to the lower layers while levels of sparsity. <

> the whole network. Previous work has shown that a non-uniform sparsity across layers is crucial to the performance in high sparsity regimes [He et al., 2018]. In particular, Mallya and Lazebnik [2018] found that the sparsity tends to increase with the depth of the network layer. Figure 5 compares the performance of local selection (matrix by matrix) against global selection (all the matrices) for magnitude pruning and movement pruning. Despite being able to find a global sparsity structure, we found that global did not significantly outperform local, except in high sparsity regimes (2.3 F1 points of difference with 3% of remaining weights for movement pruning). Even though the distillation signal boosts the performance of pruned models, the end performance difference between local and global selections remains marginal. Figure 6 shows the remaining weights percentage obtained per layer when the model is pruned until 10% with global pruning methods. Global weight pruning is able to allocate sparsity non-uniformly through the network, and it has been shown to be crucial for the performance in high sparsity regimes [He et al., 2018]. We notice that except for global magnitude pruning, all the global pruning methods tend to allocate a significant part of the weights to the lowest layers while heavily pruning in the highest layers. Global magnitude pruning tends to prune similarly to local magnitude pruning, i.e., uniformly across layers. 8 Conclusion We consider the case of pruning of pretrained models for task-specific fine-tuning and compare zeroth- and first-order pruning methods. We show that a simple method for weight pruning based on straight-through gradients is effective for this task and that it adapts using a first-order importance score. We apply this movement pruning to a transformer-based architecture and empirically show that our method consistently yields strong improvements over existing methods in high-sparsity regimes. The analysis demonstrates how this approach adapts to the fine-tuning regime in a way that magnitude pruning cannot. In future work, it would also be interesting to leverage group-sparsity inducing penalties [Bach et al., 2011] to remove entire columns or filters. In this setup, we would associate a score to a group of weights (a column or a row for instance). In the transformer architecture, it would give a systematic way to perform feature selection and remove entire columns of the embedding matrix. References Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.ArXiv, abs/1910.10683, 2019. Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep learning in nlp. InACL, 2019. Song Han, Jeff Pool, John Tran, and William J. Dally. Learning both weights and connections for efficient neural network. InNIPS, 2015. Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding. InICLR, 2016. Yiwen Guo, Anbang Yao, and Yurong Chen. Dynamic network surgery for efficient dnns. InNIPS, 2016. Trevor Gale, Erich Elsen, and Sara Hooker. The state of sparsity in deep neural networks.ArXiv, abs/1902.09574, 2019. Jonathan Frankle, Gintare Karolina Dziugaite, Daniel M. Roy, and Michael Carbin. The lottery ticket hypothesis at scale.ArXiv, abs/1903.01611, 2019. Yoshua Bengio, Nicholas Léonard, and Aaron C. Courville. Estimating or propagating gradients through stochastic neurons for conditional computation.ArXiv, abs/1308.3432, 2013. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InNAACL, 2019. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InNIPS, 2017. Christos Louizos, Max Welling, and Diederik P. Kingma. Learning sparse neural networks through l0 regularization. InICLR, 2017. Adina Williams, Nikita Nangia, and Samuel Bowman. A broad-coverage challenge corpus for sentence understanding through inference. InNAACL, 2018. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100, 000+ questions for machine comprehension of text. InEMNLP, 2016. Arun Mallya and Svetlana Lazebnik. Piggyback: Adding multiple tasks to a single, fixed network by learning to mask.ArXiv, abs/1801.06519, 2018. Vivek Ramanujan, Mitchell Wortsman, Aniruddha Kembhavi, Ali Farhadi, and Mohammad Rastegari. What’s hidden in a randomly weighted neural network? InCVPR, 2020. Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. InNIPS, 1989. Babak Hassibi, David G. Stork, and Gregory J. Wolff. Optimal brain surgeon: Extensions and performance comparisons. InNIPS, 1993. Lucas Theis, Iryna Korshunova, Alykhan Tejani, and Ferenc Huszár. Faster gaze prediction with dense networks and fisher pruning.ArXiv, abs/1801.05787, 2018. Xiaohan Ding, Guiguang Ding, Xiangxin Zhou, Yuchen Guo, Ji Liu, and Jungong Han. Global sparse momentum sgd for pruning very deep neural networks. InNeurIPS, 2019. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. InNeurIPS EMC2 Workshop, 2019. Raphael Tang, Yao Lu, Linqing Liu, Lili Mou, Olga Vechtomova, and Jimmy Lin. Distilling task-specific knowledge from bert into simple neural networks.ArXiv, abs/1903.12136, 2019. Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. InICLR, 2020a. Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? InNeurIPS, 2019. Haonan Yu, Sergey Edunov, Yuandong Tian, and Ari S. Morcos. Playing the lottery with rewards and multiple languages: lottery tickets in rl and nlp. InICLR, 2020. Angela Fan, Pierre Stock, Benjamin Graham, Edouard Grave, Rémi Gribonval, Hervé Jégou, and Armand Joulin. Training with quantization noise for extreme model compression.ArXiv, abs/2004.07320, 2020b. Ofir Zafrir, Guy Boudoukh, Peter Izsak, and Moshe Wasserblat. Q8bert: Quantized 8bit bert.ArXiv, abs/1910.06188, 2019. Yunchao Gong, Liu Liu, Ming Yang, and Lubomir D. Bourdev. Compressing deep convolutional networks using vector quantization.ArXiv, abs/1412.6115, 2014. Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E. Gon- zalez. Train large, then compress: Rethinking model size for efficient training and inference of transformers.ArXiv, abs/2002.11794, 2020. Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for model compression. InICLR, 2018. Mitchell A. Gordon, Kevin Duh, and Nicholas Andrews. Compressing bert: Studying the effects of weight pruning on transfer learning.ArXiv, abs/2002.08307, 2020. Sebastian Ruder, Matthew E. Peters, Swabha Swayamdipta, and Thomas Wolf. Transfer learning in natural language processing. InNAACL, 2019. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach.ArXiv, abs/1907.11692, 2019. Shankar Iyer, Nikhil Dandekar, and Kornel Csernai. First quora dataset release: Question pairs, 2017. URLhttps://data.quora.com/First-Quora-Dataset-Release-Question-Pairs. Fu-Ming Guo, Sijia Liu, Finlay S. Mungall, Xue Lian Lin, and Yanzhi Wang. Reweighted proximal pruning for large-scale language representation.ArXiv, abs/1909.12486, 2019. Neal Parikh and Stephen P. Boyd. Proximal algorithms.Found. Trends Optim., 1:127–239, 2014. Iulia Turc, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Well-read students learn better: The impact of student initialization on knowledge distillation.ArXiv, abs/1908.08962, 2019. Cristian Bucila, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. InKDD, 2006. Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. InNIPS, 2014. Xiaoqi Jiao, Y. Yin, Lifeng Shang, Xin Jiang, Xusong Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding.ArXiv, abs/1909.10351, 2019. Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, and Song Han. Amc: Automl for model compression and acceleration on mobile devices. InECCV, 2018. Francis Bach, Rodolphe Jenatton, Julien Mairal, and Guillaume Obozinski. Structured sparsity through convex optimization.Statistical Science, 27, 09 2011. doi: 10.1214/12-STS394. A Appendices A.1 Guarantees on the decrease of the training loss As the scores are updated, the relative order of the importances is likely shuffled, and some connections will be replaced by more important ones. Under certain conditions, we are able to formally prove that as these replacements happen, the training loss is guaranteed to decrease. Our proof is adapted from [Ramanujan et al., 2020] to consider the case of fine-tuableW. We suppose that (a) the training lossLis smooth and admits a first-order Taylor development everywhere it is defined and (b) the learning rate of <> is small. We define the TopK function as the analog of the Top v function, wherekis an integer instead of a proportion. We first consider the case where k=1 in the Top K masking, meaning that only one connection is remaining (and the other weights are deactivated/masked). Let’s denote <> this sole remaining connection at stept. Following Eq (1), it means that <> We suppose that at stept+ 1, connections are swapped and the only remaining connection at step <>. We have: <> Eq(6)gives the following inequality: <>. After re-injecting the gradient <> update in Eq (2), we have: <> (7) Moreover, the conditions in Eq (6) lead to the following inferences: <> and <>. Since <> is small, <> is also small. Because the training loss L is smooth, we can write the 1st order Taylor development of L in point <> <> (8) The first term is null because of inequalities(6)and the second term is negative because of inequality (7). Thus <> when connection <> becomes more important than <>, the connections are swapped and the training loss decreases between step k t and <>. Similarly, we can generalize the proof to a set <> of N swapping connections. We note that this proof is not specific to theTopKmasking function. In fact, we can extend the proof using theThresholdmasking function <> [Mallya and Lazebnik, 2018]. Inequalities (6) are still valid and the proof stays unchanged. Last, we note these guarantees do not hold if we consider the absolute value of the scoresjSi;j j(as it is done in Ding et al. [2019] for instance). We prove it by contradiction. If it was the case, it would also be true one specific case: thenegative thresholdmasking function <> where <>. We suppose that at stept+ 1, the only remaining connection(i;j)is replaced by(k;l): <> (9) The inequality on the gradient update becomes:<> and following i the same development as in Eq(8), we have <> the loss increases. <> We proved by contradiction that the guarantees on the decrease of the loss do not hold if we consider k the absolute value of the score as a proxy for importance. <> <> <> <> <> <> Network Pruning As one of the earliest works in network pruning, Yann Lecun's Optimal brain damage (OBD) paper has been cited in many of the papers. Some research focuses on module network designs. "These models, such as SqueezeNet , MobileNet  and Shufflenet, are basically made up of low resolutions convolution with lesser parameters and better performance." Many recent papers -I've read- ephasize on structured pruning (or sparsifying) as a compression and regularization method, as opposed to other techniques such as non-structured pruning (weight sparsifying and connection pruning), low rank approximation and vector quantization (references to these approaches can be found in the related work sections of the following papers).  Difference between structred and non-structured pruning: "Non-structured pruning aims to remove single parameters that have little influence on the accuracy of networks". For example, L1-norm regularization on weights is noted as non-structured pruning- since it's basically a weight sparsifying method, i.e removes single parameter. The term 'structure' refers to a structured unit in the network. So instead of pruning individual weights or connections, structured pruning targets neurons, filters, channels, layers etc. But the general implementation idea is the same as penalizing individual weights: introducing a regularization term (mostly in the form of L1-norm) to the loss function to penalize (sparsify) structures. I focused on structured pruning and read through the following papers: 1. Structured Pruning of Convolutional Neural Networks via L1 Regularization (August 2019) "(...) network pruning is useful to remove redundant parameters, filters, channels or neurons, and address the over-fitting issue." Provides a good review of previous work on non-structured and structured pruning. "This study presents a scheme to prune filters or neurons of fully-connected layers based on L1 regularization to zero out the weights of some filters or neurons." Didn't quite understand the method and implementation. There are two key elements: mask and threshold. "(...) the problem of zeroing out the values of some filters can be transformed to zero some mask." || "Though the proposed method introduces mask, the network topology will be preserved because the mask can be absorbed into weight." || "Here the mask value cannot be completely zeroed in practical application, because the objective function (7) is non-convex and the global optimal solution may not be obtained. A strategy is adopted in the proposed method to solve this problem. If the order of magnitude of the mask value is small enough, it can be considered almost as zero. Thus, to decide whether the mask is zero, a threshold is introduced. (...) The average value of the product of the mask and the weight is used to determine whether the mask is exactly zero or not." From what I understand they use L1 norm in the loss function to penalize useless filters through penalizing masks. And a threshold value is introduced to determine when the mask is small enough to be considered zero.  They test on MNIST (models: Lenet-5) and CIFAR-10 (models: VGG-16, ResNet- 32) 2. Learning Efficient Convolutional Networks through Network Slimming (August 2017) + Git repo "Our approach imposes L1 regularization on the scaling factors in batch normalization (BN) layers, thus it is easy to implement without introducing any change to existing CNN architectures. Pushing the values of BN scaling factors towards zero with L1 regularization enables us to identify insignificant channels (or neurons), as each scaling factor corresponds to a specific convolutional channel (or a neuron in a fully-connected layer)." They provide a good insight on advantages and disadvantages of other computation reduction methods such as low rank approximation, vector quantization etc.  I believe here they use the word 'channel' to refer to filters (?). "Our idea is introducing a scaling factor γ for each channel, which is multiplied to the output of that channel. Then we jointly train the network weights and these scaling factors, with sparsity regularization imposed on the latter. Finally we prune those channels with small factors, and fine-tune the pruned network. " --> so instead of 'mask' they use the 'scaling factor' and impose regularization on that, but the idea is very similar. "The way BN normalizes the activations motivates us to design a simple and efficient method to incorporates the channel-wise scaling factors. Particularly, BN layer normalizes the internal activations using mini-batch statistics." || " (...) we can directly leverage the γ parameters in BN layers as the scaling factors we need for network slim- ming. It has the great advantage of introducing no overhead to the network." They test on CIFAR and SVHN (models: VGG-16, ResNet-164, DenseNet-40), ImageNet (model: VGG-A) and MNIST (model: Lenet) 3. Learning Structured Sparsity in Deep Neural Networks (Oct 2016) + Git repo " (...) we propose Structured Sparsity Learning (SSL) method to directly learn a compressed structure of deep CNNs by group Lasso regularization during the training. SSL is a generic regularization to adaptively adjust multiple structures in DNN, including structures of filters, channels, and filter shapes within each layer, and structure of depth beyond the layers." || " (...) offering not only well- regularized big models with improved accuracy but greatly accelerated computation."  "Here W represents the collection of all weights in the DNN; ED(W) is the loss on data; R(·) is non-structured regularization applying on every weight, e.g., L2- norm; and Rg(·) is the structured sparsity regularization on each layer. Because Group Lasso can effectively zero out all weights in some groups [14][15], we adopt it in our SSL. The regularization of group Lasso on a set of weights w can be represented as, where w(g) is a group of partial weights in w and G is the total number of groups. "<>" In SSL, the learned “structure” is decided by the way of splitting groups of w(g). We investigate and formulate the filer-wise, channel-wise, shape-wise, and depth-wise structured sparsity (...)" They test on MNIST (models: Lenet, MLP), CIFAR-10 (models: ConvNet, ResNet- 20) and ImageNet (model:AlexNet) The authors also provide a visualization of filters after pruning, showing that only important detectors of patterns remain after pruning. In conclusions: "Moreover, a variant of SSL can be performed as structure regularization to improve classification accuracy of state-of-the-art DNNs." 4. Learning both Weights and Connections for Efficient Neural Networks (Oct 2015) "After an initial training phase, we remove all connections whose weight is lower than a threshold. This pruning converts a dense, fully-connected layer to a sparse layer." || "We then retrain the sparse network so the remaining connections can compensate for the connections that have been removed. The phases of pruning and retraining may be repeated iteratively to further reduce network complexity. In effect, this training process learns the network connectivity in addition to the weights (...)" Although the description above implies the pruning was done only for FC layers, they also do pruning on convolutional layers - although they don't provide much detail on this in the methods. But there's this statement when they explain retraining: "(...) we fix the parameters for CONV layers and only retrain the FC layers after pruning the FC layers, and vice versa.". The results section also shows that convolutional layer connections were also pruned on the tested models. They test on MNIST (models: Lenet-300-100 (MLP), Lenet-5 (CNN)) and ImageNet (models: AlexNet, VGG-16) The authors provide a visualization of the sparsity patterns of neurons after pruning (for an FC layer) which shows that pruning can detect visual attention regions. The method used in this paper targets individual parameters (weights) to prune. So, technically this should be considered as a non-structured pruning method. However, the reason I think this is referenced as a structured pruning method is that if all connections of a neuron is pruned (i.e all input and output weights were below threshold), the neuron itself will be removed from the network:  "After pruning connections, neurons with zero input connections or zero output connections may be safely pruned." SIDENOTE: They touch on the use of global average pooling instead of fully connected layers in CNNs: "There have been other attempts to reduce the number of parameters of neural networks by replacing the fully connected layer with global average pooling." 5. Many more can be picked from the references of these papers. There's a paper on Bayesian compression for Deep Learning from 2017. Their hypothesis is: "By employing sparsity inducing priors for hidden units (and not individual weights) we can prune neurons including all their ingoing and outgoing weights." However, the method is mathematically heavy and the related work references are quite old (1990s, 2000s). <> <> <> <> <> <> Network Trimming: A Data-Driven Neuron Pruning Approach towards Efficient Deep Architectures Hengyuan Hu % Rui Peng % Yu-Wing Tai Chi-Keung Tang HKUST HKUST SenseTime Group Limited HKUST hhuaa@ust.hk rpeng@ust.hk yuwing@sensetime.com cktang@cse.ust.hk Abstract State-of-the-art neural networks are getting deeper and wider. While their performance increases with the increasing number of layers and neurons, it is crucial to design an efficient deep architecture in order to reduce computational and memory costs. Designing an efficient neural network, however, is labor intensive requiring many experiments, and fine-tunings. In this paper, we introduce network trimming which iteratively optimizes the network by pruning unimportant neurons based on analysis of their outputs on a large dataset. Our algorithm is inspired by an observation that the outputs of a significant portion of neurons in a large network are mostly zero, regardless of what inputs the network received. These zero activation neurons are redundant, and can be removed without affecting the overall accuracy of the network. After pruning the zero activation neurons, we retrain the network using the weights before pruning as initialization. We alternate the pruning and retraining to further reduce zero activations in a network. Our experiments on the LeNet and VGG-16 show that we can achieve high compression ratio of parameters without losing or even achieving higher accuracy than the original network. 1 Introduction Neural networks have been widely adopted in many scenarios, achieving state-of-the-art results in numerous tasks [1] [2]. One of the keys to improved performance is their increased depth and width and thus the increased number of parameters. In computer vision, we have witnessed orders of magnitude increase in the number of parameters in CNNs from LeNet with less than 1M parameters in handwritten digit classification [3] to Deepface with more than 120M parameters in human face classification [4]. Although CNNs with elegant network architectures are easy to deploy in real-world tasks, designing one can be hard and labor-intensive, which involves significant amount of effort in empirical experiments. In terms of designing the network architecture, one crucial part is to determine the number of neurons in each layer. There is no way to directly arrive at an optimal number of neurons for each layer and thus even the most successful network architectures use empirical numbers like 128, 512, 4096. Experienced scientists often arrive at the numbers once they deem the network have enough representation power for the specific task. However, the extremely sparse matrices produced by top layers of neural networks have caught our attention, indicating that empirically designed networks are heavily oversized. After some simple statistics, we find that many neurons in a CNN have very low activations no matter what data is presented. Such weak neurons are highly likely to be redundant and can be excluded without damaging the overall performance. Their existence can only increase the chance of overfitting and optimization difficulty, both of which are harmful to the network. With the motivation of achieving more efficient network architectures by finding the optimal number of neurons in each layer, we come up with an iterative optimization method that gradually eliminates Part of the work was done when Hengyuan Hu and Rui Peng were interns in SenseTime Group Limited weak neurons in a network via a pruning-retraining loop. Starting from an empirically designed network, our algorithm first identifies redundant weak neurons by analyzing their activations on a large validation dataset. Then those weak neurons are pruned while others are kept to initialize a new model. Finally, the new model is retrained or fine-tuned depending on the performance drop. The retrained new model can maintain the same or achieve higher performance with smaller number of neurons. This process can be carried out iteratively until a satisfying model is produced. 2 Related Work Significant redundancy has been demonstrated in several deep learning models [5] and such redundancy is mainly caused by the overwhelming amount of parameters in deep neural networks. An over-parameterized model not only wastes memory and computation, but also leads to serious overfitting problem. Therefore, reducing the number of parameters has been studied by many researchers in this field. However, there is little work directly addressing the optimization of the number of neurons. Most previous works on improving network architectures fall in two main categories; one concentrates on the high level architectural design and the other focuses on low level weight pruning. On the high level side, some researchers invented new layers or modules to substitute main bottleneck components in conventional neural networks. Two famous examples of this kind are the global average pooling in Network in Network [6] invented to replace the extremely dense parameterized fully connected layer and the inception module employed by GoogLeNet [7] to avoid explosion in computational complexity at later stage. Both methods achieve state-of-the-art results on several benchmarks with much less memory and computation consumption. More recently, SqueezeNet [8] used a Fire module together with other strategies to achieve AlexNet-level accuracy with 50% less parameters. On the low level side, different methods have been explored to reduce number of connections and weights in neural networks. Some late 20th century methods, such as magnitude-based approach [9] and Hessian matrix based approach [10], prune weights basing on numerical properties of the weights and loss functions without any external data involved. Han et al. recently proposed an iterative method [11] to prune connections in deep architectures, together with an external compression by quantization and encoding [12]. The network is first pruned by removing low weights connections. Then, learned mapping of similar weights to fixed bits are used to perform quantization of weights after pruning, which facilitates the Huffman coding compression in the last stage to reduce bits for storage. When all three techniques used in pipeline, the number of parameters in the network can be reduced by around 10%. While above methods work well in practice by reducing number of parameters directly, we seek answers to the fundamental problem that lies in the middle of those two levels of approaches – determining the optimal number of neurons for each layer for a given network architecture and specific tasks. Along our direction, not only can we achieve parameter savings without the need of seeking new network architectures, we can also evaluate the redundancy in each layer of a network, and thus provide guidance on effective ways for architecture optimization in large neural networks. 3 Zero Activations and Network Trimming In this section, we describe our algorithm for network trimming. To facilitate our discussions, we use VGG-16 [13] as our case study. The VGG-16 network consists of 13 convolutional layers, and 3 fully connected layers. Each of the layers is followed by a ReLU [14] layer for non-linear mapping. The VGG-16 is recognized as one of the representative network which has been adopted to many applications [15] [16], not limited to object classification and localization tasks. 3.1 Zero Activations in VGG-16 We define Average Percentage of Zeros (APoZ) to measure the percentage of zero activations of a neuron after the ReLU mapping. Let <> denotes the output ofc-th channel ini-th layer, our <> of the c-th neuron ini-th layer is defined as: <> (1) where <> if true, and <> if false,M denotes the dimension of output feature map of <>, and N denotes the total number of validation examples. The larger number of validation examples, the more accurate is the measurement of APoZ. In our experiment, we use the validation set (N= 50;000) of ImageNet classification task to measure APoZ. We use the definition of APoZ to evaluate the importance of each neuron in a network. To validate our observation that the outputs of some neurons in a large network are mostly zero, we calculate the APoZ of each neuron and find that there are631neurons in the VGG-16 network which have APoZ larger than90%. Table 1: Mean APoZ of each layer in VGG-16 <

> To better understand the behavior of zero activations in a network, we compute the mean APoZ (Table 1) of all neurons in each layer (except for the last one) of the VGG-16 network. Since the VGG-16 network has inverse pyramid shape, most redundancy occurs at the higher convolutional layers and the fully connected layers. The higher mean APoZ also indicates more redundancy in a layer. Detailed distributions of APoZ of 512 CONV5-3 neurons and 4096 FC6 neurons are shown in Figure 1, 2 respectively. Since a neural network has a multiplication-addition-activation computation process, a neuron which has its outputs mostly zeros will have very little contribution to the output of subsequent layers, as well as to the final results. Thus, we can remove those neurons without harming too much to the overall accuracy of the network. In this way, we can find the optimal number of neurons for each layer and thus obtain a better network without redesign and extensive human labor. <

> . <

> Figure 1: CONV5-3 APoZ Distribution Figure 2: FC6 APoZ Distribution 3.2 Network Trimming and Retraining Our network trimming method consists of three main steps, as illustrated in Figure 3. First the network is trained under conventional process and the number of neurons in each layer is set empirically. Next, we run the network on a large validation dataset to obtain the APoZ of each neuron. Neurons with high APoZ are pruned according to certain criteria. The connections to and from the neuron are removed accordingly when a neuron is pruned (see Figure 4 5). After the neuron pruning, the trimmed network is initialized using the weights before trimming. The trimmed network exhibits <

> <

> <

> Figure 3: Three main steps for Figure 4: Before pruning Figure 5: After pruning trimming some level of performance drop. Thus, in the final step, we retrain the network to strengthen the remaining neurons to enhance the performance of the trimmed network. The weight initialization is necessary for the network to obtain the same performance as it was before the trimming. If a trimmed network is trained from scratch, we find that it contains larger percentage of zero activation neurons than the counterpart with weight initialization. This means that a retrained network without weight initialization is much less efficient. We experimented different ways to prune the neurons according to the APoZ measurements. We found that pruning too many neurons at once severely damaged the performance, and the performance drops are unrecoverable. Therefore, we chose an iterative scheme to trim a network. However, it is not trivial to trim a network with deep architecture. If too many layers are trimmed in one step, the performance would drop by a large margin, and it is hard to recover the original performance before trimming through the retraining. For example, trimming CONV4, CONV5, FC6 and FC7 of the VGG-16 network concurrently would lead to a46:650%top-5 accuracy in the image classification task, where the original accuracy of VGG-16 2 is88:444%. On the other hand, if only the CONV5-3 and FC6 are trimmed, the trimmed network with weight initialization before retraining can achieve 85:900%top-5 accuracy. After retraining, the trimmed network achieves90:278%accuracy which is even higher than the original accuracy before trimming. Empirically, we found that starting to trim from a few layers with high mean APoZ, and then progressively trim its neighboring layers can rapidly reduce the number of neurons while maintaining the performance of the original network. To decide which neurons to prune, we empirically found that pruning the neurons whose APoZ is larger than one standard derivation from the mean APoZ of the target trimming layer would produce good retraining results. Using this threshold, we would reject16%of neurons on average from the trimmed layers, assuming that the APoZ values roughly follow a Gaussian distribution. 4 Experiments We implemented our algorithm using the standard Caffe [17] library. To obtain the weights for initialization for retraining, we use the Python and PyCaffe interface to copy the weights of remaining connections after the trimming. We tested our algorithm primarily on two networks, LeNet [3] on MNIST dataset and VGG-16 on ImageNet classification dataset [18]. 4.1 LeNet The LeNet network consists of two convolutional layers followed by two fully connected layers, the layers have20;50;500;10outputs respectively. We use a short hand notion (20-50-500-10) to denote the number of neurons in each layer of the network. In the LeNet,93%of parameters are in the connections between the CONV2 layer and the FC1 layer. Consequently, we can easily achieve a more efficient network by trimming the size of CONV2 and FC1 layers. 2 Single scale, without dense evaluation [13] 4.1.1 Effectiveness We apply our algorithm to iteratively prune the neurons in CONV2 and FC1 layers, as shown in Table 2. At the first iteration of the pruning, the numbers of neurons in CONV2 and FC1 layers are reduced to 41 and 426 respectively, which achieves1:41%compression on the number of parameters after the first pruning. The accuracy drops from99:27%to98:75%after the pruning, but before retraining. After retraining the network, we achieve99:29%accuracy which is slightly higher than the original accuracy. We repeat these processes for 4 iterations. As shown in Table 2, our algorithm achieves more than 2%3%compression on the number of parameters without loss in accuracy. Table 2: Iterative Trimming on LeNet <

> 4.1.2 Necessity of Weight Initialization We experiment our algorithm with retraining with and without weight initialization, as summarized in Table 3. The network exhibits deterioration in classification accuracy without weight initialization, whereas with proper weight initialization from the ancestor network from the previous iteration, the trimmed network can retain its original or even achieve higher accuracy. Table 3: Iterative Trimming on LeNet with and without Weight Initialization <> Moreover, we observe that with the weight initialization, the trimmed network consistently has smaller mean APoZ values than its ancestor network. This means that the retrained network has less redundancy than its ancestor network. In contrast, mean APoZ values increase if we retrain the network from scratch even though the trimmed network has less neurons than its ancestor network. This observation gives us an insight that proper weight initialization is necessary to achieve an efficient trimmed network. 4.2 VGG-16 4.2.1 Effectiveness With the similar objective to obtain optimal number of neurons in each layer, we analyzed the APoZ values ofO(i)c for all i and c_in VGG-16 on ImageNet classification validation set. As shown in Table 1, CONV4, CONV5 and FC layers have higher mean APoZ compared with bottom layers, exhibiting more redundancy. Drawing from previous experience on LeNet, we focus on the parameter bottleneck of VGG-16. We trim the VGG-16 network starting from the CONV5-3 and FC6 layers since they account for 100M/138M parameters. We iteratively prune neurons from CONV5-3 and FC6 layers. Similar to the case in LeNet, the trimming process can effectively eliminate neurons with high APoZ. As shown in Figure 6, after trimming, the entire distribution of APoZ inO(fc6) shifts left, indicating a significant drop in network <

> Figure 6: FC6 APoZ distribution before and after trimming redundancy. Meanwhile, the diminishing tail on the right side of the curve manifests that the weak neurons in FC6 are vanishing, a proof of the benefit gained from weight initialization as discussed in Section 3.2 and 4.1.2. Table 4: Iterative Trimming Result on VGG-16 {CONV5-3, FC6} <

> After 6 iterations of trimming, we reduce more than half of the total number of parameters and achieve a compression rate of2:59%while the trimmed network has 2%3% higher Top-1/Top-5 accuracy than the original VGG-16 model. The detailed performance of intermediate models are summarized in Table 4. There are two interesting observations in the table. First, the initial accuracy just after trimming does not drop much from the last model even though around 500 neurons in CONV5-3 and FC6 are pruned in each iteration. This is a strong proof of redundancy in empirically designed neural networks. Also, such a small decrease in accuracy can be remedied via a fast fine-tuning instead of a week-long retraining. In our experiments, it takes less than 5K iterations to reach the original accuracy (with batch size = 256). Therefore, our trimming method allows fast optimization towards better architecture. Secondly, the trimmed networks surprisingly surpass the original VGG-16 in accuracy with less parameters. The good initialization provided by previous model sets a promising starting point for the trimmed model. In addition, having less parameters in FC6 also reduces the chance of overfitting, which may also contribute to the increment in accuracy. 4.2.2 Trimming Multiple Layers VGG-16 differs from LeNet greatly in that it has a much deeper architecture with significantly more layers, which naturally gives us more options to determine which layers to trim. After the previous experiments, we want to further investigate if trimming multiple layers simultaneously can achieve the same effectiveness. After trimming the CONV5-3 and FC6 layers, we continue to trim their neighboring layers. We experimented with three sets of trimming layouts: {CONV5, FC6}, {CONV5, FC6, FC7}, {CONV4, CONV5, FC6, FC7} (see Table 5). When more neurons are pruned, the large performance drop in the trimmed network indicates retraining is necessary. We use the same set of training hyperparameters in our experiments: {base-lr: 0.001, gamma: 0.1, step-size: 3000}. After retraining, the trimmed networks gradually recover from the loss of neurons and rise to an accuracy level equivalent to the Table 5: Iterative Trimming Result on VGG-16 Many Layers <> reference model or slightly higher. In contrast to trimming only one layer, these models regain to their capacity rather slowly, taking more than 10K iterations to recover the accuracy. Empirically, we found that iteratively trimming the network starting from a few layers can achieve better performance. We also found that trimming the last convolutional layer and the fully connected layers are the most effective. As shown in Table 6, additional trimming of FC7 layer (based on previously trimmed model (CONV5-3, FC6) = (420, 2121)), can achieve a high2:7%compression rate with improved accuracy. The underlying reason is that once we have pruned the FC6 layer, the numerous zeros contribute to the high APoZ value of neurons in the FC7 layer. For the goal to reduce network parameters, it is suffices to just trim the {CONV5-3, FC6, FC7} layers since around86%of all the parameters are in the {CONV5-3, FC6, FC7} layers. Table 6: Iterative Trimming Result on VGG-16 {CONV5-3, FC6, FC7} <> 5 Discussion 5.1 Comparison with Connection Pruning Work closest to ours is the work by Han et al. [11] where they iteratively prune the network connections when the correspondent weights of the connections are close to zero. They also prune a neuron when the connections to a neuron are all pruned. Compared with their work, our work is better in two major aspects. First, although Han et al. claim that they have achieved a reduction rate of parameters by 13%on VGG-16, their reduction is tailored for CPU implementation of a neural network. In a GPU implementation, the convolutional layer is implemented by first vectorizing a 2D feature map into a 1D feature vector followed by a matrix multiplication [19]. Thus, if a neuron is not pruned, the number of multiplications for the convolutional layers will remain the same since the vectorization is performed in a universal manner for all neurons in the same layer. This is also the same case for fully connected layers where the number of multiplications are universal for all neurons in the same layer. Note that the computational costs to re-vectorize a 2D feature map to fit for different shape of neuron connections, or adding a conditional mask checking is a lot higher than a simple matrix multiplication with redundancy. Our method, on the contrary, removes all unneeded neurons so that they do not consume any memory and are not involved in any computation at all. As shown in Section 4.2, the trimmed VGG-16 has more than2%less FLOPs in the first fully connected layer. Second, pruning a neuron by first pruning all of its connections is less efficient and less effective than our APoZ measurement. This is because the number of connections is significantly larger than the number of neurons in a network, especially for fully connected layers. In our experiments, we found that most of the redundancy resides in fully connected layers, and in the connections between the last convolutional layer and the first fully connected layer. However, it is rarely the case that the weights of all connections to a neuron in these layers are close to zero. Consequently, it is difficult to prune a neuron in these layers. On the other hand, our APoZ measurement can easily identify zero activation neurons for pruning regardless the weight of connections. The mean APoZ can also be used as a guideline to evaluate the effectiveness of a network as demonstrated in our experiments. 5.2 Dataset Used During Trimming In all of our experiments, we train the network using training set and run the network on validation set to obtain APoZs for neuron pruning. This method may be controversial because the validation set should not be glimpsed before finalizing the model which may potentially lead to overfitting of the validation set. We also have the same suspicion, especially after the experiments that the trimmed model can have2%higher top-5 accuracy than that of the original VGG-16 on the validation set. Therefore, we consult two more experiments to explore the potential issue. In the first experiment, we randomly sampled a subset from training set with equal number of images (50K) as validation set. Then we used the same criteria to select weak neurons for pruning. The weak neurons selected using the sampled training set have more than95%overlap ratio with the exact neurons selected using the validation set. This shows that neurons have consistent activation performance on training and validation sets. In another word, the trimmed networks learned from sampled training data will be similar to the trimmed networks learned from the validation set. In addition, we also tested our model on the test set of ILSVRC2012 classification track. Using single model without dense evaluation, the original VGG-16 model with11:56%validation error has an error rate of13:02%on test set. Our trimmed network with configuration {CONV5-3: 420, FC6: 2121, FC7: 2482, Compression Rate: 2.00, Validation Error:9:7%} achieved10:02%error rate on test set. Note that the test set and validation set are non-overlapping in this ILSVRC2012 classification task. Telling from the data, after the network trimming, not only the overall accuracy is increased, but the gap between validation error and test error is also shrunk, indicating that the trimmed network has less overfitting. The two extra experiments dismiss our concern on overfitting. They also suggest that the validation set can be used for analyzing APoZs. 6 Conclusion We have presented Network Trimming to prune redundant neurons based on the statistics of neurons’ activations. With our method, one network architecture can be deployed to handle different tasks on different datasets and the algorithm can tailor the network accordingly by determining how many neurons to use for each layer without the need of intensive computational power as well as human labor. Our method can iteratively remove low activation neurons that provide little power to the final results without damaging performance of the model. We experimented our algorithm on LeNet and VGG-16 achieving the same accuracy with 2%3%less parameters. In VGG-16, the trimmed models can even surpass the original one, which could be caused by the reduced optimization difficulty. Lying in the middle of high level network redesign and low level weight pruning, neuron pruning can be applied to any mature architecture together with weight pruning to sharply reduce the complexity of network. References [1]Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems. (2012) 1097–1105 [2]Graves, A., Schmidhuber, J.: Framewise phoneme classification with bidirectional lstm and other neural network architectures. Neural Networks18(5) (2005) 602–610 [3]Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE86(11) (Nov 1998) 2278–2324 [4]Taigman, Y., Yang, M., Ranzato, M., Wolf, L.: Deepface: Closing the gap to human-level performance in face verification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. (2014) 1701–1708 [5]Denil, M., Shakibi, B., Dinh, L., Ranzato, M., de Freitas, N.: Predicting parameters in deep learning. CoRRabs/1306.0543(2013) [6] Lin, M., Chen, Q., Yan, S.: Network in network. CoRRabs/1312.4400(2013) [7]Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.E., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. CoRRabs/1409.4842(2014) [8]Iandola, F.N., Moskewicz, M.W., Ashraf, K., Han, S., Dally, W.J., Keutzer, K.: Squeezenet: Alexnet-level accuracy with 50x fewer parameters and< 1mb model size. arXiv preprint arXiv:1602.07360 (2016) [9]Hanson, S.J., Pratt, L.: Advances in neural information processing systems 1. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA (1989) 177–185 [10]Hassibi, B., Stork, D.G.: Second order derivatives for network pruning: Optimal brain surgeon. In: Advances in Neural Information Processing Systems 5, [NIPS Conference], San Francisco, CA, USA, Morgan Kaufmann Publishers Inc. (1993) 164–171 [11]Han, S., Pool, J., Tran, J., Dally, W.J.: Learning both weights and connections for efficient neural networks. CoRRabs/1506.02626(2015) [12]Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015) [13]Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recogni- tion. arXiv preprint arXiv:1409.1556 (2014) [14]Nair, V., Hinton, G.E.: Rectified linear units improve restricted boltzmann machines. In: Proceedings of the 27th International Conference on Machine Learning (ICML-10). (2010) 807–814 [15]Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: Advances in Neural Information Processing Systems. (2015) 91–99 [16]Venugopalan, S., Rohrbach, M., Donahue, J., Mooney, R., Darrell, T., Saenko, K.: Sequence to sequence-video to text. In: Proceedings of the IEEE International Conference on Computer Vision. (2015) 4534–4542 [17]Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama, S., Darrell, T.: Caffe: Convolutional architecture for fast feature embedding. In: Proceedings of the ACM International Conference on Multimedia, ACM (2014) 675–678 [18]Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV)115(3) (2015) 211–252 [19]Scherer, D., Schulz, H., Behnke, S.: Accelerating large-scale convolutional neural networks with parallel graphics multiprocessors. In: Artificial Neural Networks–ICANN 2010. Springer (2010) 82–91 <> <> <> <> <> <> PLUG AND PLAY LANGUAGE MODELS : A SIMPLE APPROACH TO CONTROLLED TEXT GENERATION Sumanth Dathathri Andrea Madotto Janice Lan Jane Hung CMS, Caltech HKUST Uber AI Uber AI Eric Frank Piero Molino Jason Yosinski yy Rosanne Liu y Uber AI Uber AI Uber AI Uber AI dathathris@gmail.com, amadotto@connect.ust.hk {janlan, jane.hung, mysterefrank, piero, yosinski, rosanne}@uber.com ABSTRACT Large transformer-based language models (LMs) trained on huge text corpora have shown unparalleled generation capabilities. However, controlling attributes of the generated language (e.g. switching topic or sentiment) is difficult without modifying the model architecture or fine-tuning on attribute-specific data and en- tailing the significant cost of retraining. We propose a simple alternative: the Plug and Play Language Model (PPLM) for controllable language generation, which combines a pretrained LM with one or more simple attribute classifiers that guide text generation without any further training of the LM. In the canonical scenario we present, the attribute models are simple classifiers consisting of a user-specified bag of words or a single learned layer with 100,000 times fewer parameters than the LM. Sampling entails a forward and backward pass in which gradients from the attribute model push the LM’s hidden activations and thus guide the generation. Model samples demonstrate control over a range of topics and sentiment styles, and extensive automated and human annotated evaluations show attribute alignment and fluency. PPLMs are flexible in that any combination of differentiable attribute models may be used to steer text generation, which will allow for diverse and creative applications beyond the examples given in this paper. 1 INTRODUCTION The Transformer architecture (Vaswani et al., 2017) has enabled large-scale language models (LMs) trained on a huge amount of data (Radford et al., 2019; Dai et al., 2019b; Radford et al., 2018b) to greatly improve the state-of-the-art on natural language processing tasks. These models are used to extract contextualized word embeddings for transfer learning purposes (Devlin et al., 2019) and as natural language generators. The latter can leverage large amounts of unannotated data and a simple log-likelihood training objective. However, once such models are trained, controlling attributes of generated text becomes difficult without modifying the model architecture to allow for extra input attributes or fine-tuning with attribute-specific data (Keskar et al., 2019; Ziegler et al., 2019). conceptualized PPLMs and led the manuscript writing. SD led thecproject, implemented the PPLM, set up and ran all modeling experiments, engineered how to obtain workable gradients via the weighted embedding approach, and made the model work. AM helped with preparing datasets for discriminator training, automated evaluation, running experiments, and writing the manuscript. SD, RL & AM ran the external baselines. RL & JL built and oversaw the human evaluation pipeline and computed the statistics. JH ran the story generation with skeleton prefixes. EF assisted with detoxification experiments. PM led efforts to migrate to the new pytorch transformer, helped with code release. JY helped with the annotation pipeline, finding bugs, navigating model and experimental directions, engineering workable gradients, and posing the model mathematically. RL implemented preliminary experiments and multi-attribute control, and cleaned and coordinated release of the code. RL & JY oversaw the project. <> Table 1: The PPLM employs a pre-trained language model (LM) without any changes to the model parameters and can generate text with controlled attributes such as topic and sentiment. We demon- strate control with two tiny and easy to construct attribute models: a bag of words (BoW) related to a topic and a linear discriminator trained on top of LM latent representations to control sentiment. The underlined prefix is what the LM is conditioned on to generate a passage of text (e.g.The potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potatoThe potato). The controlled attributes are colored and bracketed (e.g.[Science]), and words in the BoW that are directly optimized for are highlighted brightly (e.g.research). The softer highlights correspond to words related to the attribute, but not directly optimized for during the control process (e.g.health). [–] The potato The potatoThe potato The potatoThe potatoThe potatoThe potato The potatoThe potatoThe potato The potatoThe potatoThe potatoThe potato The potatoThe potato The potato and cauliflower are both in season to make combo breads, mounds, or pads. For an added challenge, try some garlic mashed potatoes. [Negative] The potato The potatoThe potato The potatoThe potatoThe potatoThe potato The potatoThe potatoThe potato The potatoThe potatoThe potatoThe potato The potatoThe potato The potato is a prettybad idea. It can make you fat, it can cause you to have aterribleimmune system, and it can even kill you.... [Positive] The potato The potatoThe potato The potatoThe potatoThe potatoThe potato The potatoThe potatoThe potato The potatoThe potatoThe potatoThe potato The potatoThe potato The potato chip recipe you asked for! Welovemaking these, and I’ve been doing so for years. I’ve always had a hard time keeping a recipe secret. I think it’s the way our kids love to eat them – so many little ones. [Science] The potato The potatoThe potato The potatoThe potatoThe potatoThe potato The potatoThe potatoThe potato The potatoThe potatoThe potatoThe potato The potatoThe potato The potato was once thought to have nohealthproblems and has been promoted as anutritious food source since the mid-1800s, but recent reports indicate that it has many harmful health issues. In fact, researchers from Johns Hopkins University... [Politics][Positive]To concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludethis series of articles, I will present three of the mostpopularand influential works on this topic. The first article deals with the role of women’s political participation in building a political system that is representative of the will of the people. [Politics][Negative]To concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo concludeTo conclude, the most significant and lastingdamagefrom the economiccrisisin 2008 was that many governments, including those in the political center,lost power for the first time in modern history. Controllable generation entails modeling <>, where a is some desired controllable attribute(s) and x the generated sample. However, generative models only learn <>. In computer vision, Plug & Play Generative Networks (PPGN) from Nguyen et al. (2017) developed a mechanism for generating images with different attributes by plugging a discriminator (attribute model) <> together with a base generative model <> and sampling from the resulting <>, effectively creating a conditional generative model on the fly from any supplied attribute model. In a similar manner, we propose the Plug and Play Language Model (PPLM) for conditional language generation that combines one or more simple attribute models <>—either in the form of a bag- of-words (BoW) or single layer classifiers—with a pre-trained, unconditional language model <>. We sample from the resulting combined model by following gradients in the latent representation space in a manner inspired by the approximate Metropolis-adjusted Langevin (MALA) (Roberts et al., 1996; Roberts & Rosenthal, 1998) sampler deployed in Nguyen et al. (2017). Optimization is performedex post factoin the activation space, the <> for <> re-training or fine- tuning is needed. Control is fine-grained, with a strength parameter determining how strong the attribute influence should be; a strength of0fully recovers the original model<>. This design allows vast flexibility: users can combine a state-of-the-art generative model, which may be large and difficult to train, with any number of attribute controllers. Attribute models may be easier to train or untrained (in the case of BoW models), and multiple controllers may be combined flexibly during inference. In this paper, we demonstrate the PPLM approach using a GPT-2 345M model (Radford et al., 2019) as the general-purpose LM <>, but the method applies in any representation space from any transformer-based text generator and allows combination with any attribute model <>. We demonstrate controlled generation with a number of attribute controllers, assembled and combined during generation, each with a different strength, acting as a set of “control knobs” that tune generation towards the desired attribute (see examples in Table 1). Code for the experiments is available at:https://github.com/uber-research/PPLM. Our key contributions are: •We introduce the Plug and Play LM for controlled language generation, discuss its relation to existing work, and how sampling from a PPLM works (Sections 2 and 3). •We demonstrate controlling of text generation on a range of attributes, including 7 topics each defined using a bag of words, and 1 simple discriminator on sentiments. We quantify effectiveness using both automated evaluation (separately trained perplexity and sentiment models) as well as human evaluation (for attribute relevance and fluency). All evaluations point toward the ability of PPLMs to generate attribute controlled, fluent text (Section 4). •We compare PPLM with CTRL (Keskar et al., 2019) and GPT-2 finetuned for positivity (Ziegler et al., 2019). Our method, without any LM training, is on par and often outper- forms the baselines on attribute relevance and fluency (Section 4.2, and Section 4.3). •We show that the PPLM approach can be used to detoxify instances where generation of toxic content is likely by following the negative gradient of a model trained to detect toxicity (Section 4.4). We also show how PPLM can be used for structurally constrained story writing (Section 4.5). 2 RELATED WORK Controlled generation Current methods for controlled text generation involve either fine-tuning existing models with Reinforcement Learning (RL) (Ziegler et al., 2019), training Generative Adversarial Networks (Yu et al., 2017), or training conditional generative models (Kikuchi et al., 2016; Ficler & Goldberg, 2017). Different from our approach, these methodologies are not plug and play, since the entire model needs to be separately fine-tuned for each specific attribute. Keskar et al. (2019) train a large language model with over 50 different control codes. The results are high quality because they train exactly to maximize <>, but this comes at the expense of fixing control codes upfront and of training a very large model (1.6B parameters). Our method does not require retraining any conditional generative model, and both the language model and the conditional model can be flexibly assembled. Table 2 gives a comparison of recent approaches to language modeling tuned for specific attributes. In another interesting but tangential piece of work, Subramani et al. (2019) recently showed that a pre-trained language model can be steered to recover arbitrary sentences. In earlier works Gu et al. (2016; 2017); Chen et al. (2018) explored the idea of using a small neural network to steer an LM. Noisy Channel Modeling Yu et al. (2016), and more recently Yu et al. (2019); Yee et al. (2019); Ng et al. (2019), leveraged the Shannon Noisy Channel Theory (Shannon, 1948) for improving sequence-to-sequence modeling. Their approach translates a source language sentence y into a target language sentence x by first sampling from a forward model proposal distribution p forward <> and then reranking samples based on probabilities given by p backward <>. PPLM scores samples using the same basic equation, but as we have no forward or proposal model p forward <>, we rely on the latent space updates, similar to Nguyen et al. (2017). As a baseline, we consider using<>as a “forward model” and then reranking, which we will see works moderately well in some scenarios and poorly in others (see Tables 4 and 6). Weighted decoding Holtzman et al. (2018); Ghazvininejad et al. (2017) consider controlled language generation – the former with discriminators, and the latter with a bag of words – where the decoding procedure is modified to consider the scoring function used for decoding. See et al. (2019) note that control with weighted decoding (WD) is difficult and often leads to sacrificing fluency and coherence. Further, Ghazvininejad et al. (2017) strongly relies on sampling from a set of keywords on a specific topic and it does not allow to bias generation towards a topic in a manner that does not necessary include a set of keywords. Similarly, Baheti et al. (2018) proposed a decoding strategy for generating interesting responses in dialogue systems, using bags of words and word embeddings. Sophisticated sampling methods (Metropolis et al., 1953) can be used to constrain the model generation to certain keywords and topics. We evaluate WD as a baseline. Text Style Transfer Outside of language modeling, the text style transfer studies a related task. Shen et al. (2017); Hu et al. (2017) train variational auto-encoders for style transfer that rely on learning disentangled latent representations for style and content. Li et al. (2018) demonstrate the efficacy of a simple approach based on replacing attribute related n-grams with n-grams corresponding to the desired attribute based on a conditional generative model. A key difference between the above and our approach is that we use an offline discriminator and perform optimization based on this discriminator, which as suggested by Elazar & Goldberg (2018) may outperform adversarial training approaches. More recently, Lample et al. (2019) adapt an approach from unsupervised language translation to style transfer, where a denoised auto-encoder is trained with an objective Table 2: Comparison of the different models and distributions. All models in this table are useful in different scenarios. The particular advantage of PPLM is that very small, custom attribute models, <>, may be combined with powerful, general pre-trained language models, <>, to create cheap but still powerful conditional generative models, <>. <> consisting of a weighted combination of a re-construction loss and a back-translation loss. While the above approaches have shown impressive success on style transfer tasks, the main focus is not controlled language generation, and further, the methods are not plug and play. 3 PLUG AND PLAY LANGUAGE MODELS 3.1 LANGUAGE MODELING WITH TRANSFORMERS Given a sequence of tokens <>, LMs are trained to compute the unconditional prob- ability of the sequence <>. This probability can be rewritten in terms of product of conditional probabilities by recursively applying the chain-rule (Manning et al., 1999; Bengio et al., 2003) as: <> (1) In this paper, we use a transformer (Vaswani et al., 2017) to model the distribution of natural lan- guage. To present our approach clearly, we first briefly summarize the transformer using recur- rent notation. Let us define the history matrixHt to consist of the key-value pairs from the past <>, where <> corresponds to the key-value pairs <> from the i-th layer generated at all time-steps from 0 tot. Efficient implementations of the transformer (Wolf et al., 2019) use the cachedHt to generate <>, given <>. This recurrent interpretation of a transformer can be summarized as: <>; (2) where <> a linear transformation that maps the logit vector <> to a vector of vocabulary size, and then <> is sampled as<> pt+1 =Softmax(Wo t+1 ). This allows for efficient language generation without repeated forward passes corresponding to the prior conditioning text <>. 3.2 STEERING GENERATION :ASCENDING log<> In order to control the output of the language model, at every generation step t, we shift the history Ht in the direction of the sum of two gradients: one toward higher log-likelihood (LL) of the attribute a under the conditional attribute model<>and one toward higher LL of the unmodified language model<>. Combining these factors with a variable multiplier provides us with a controllable “knob” to guide generation in a given direction with a specified strength. The updates are restricted toHt and not the other model activations because future predictions depend on the past only via Ht (note thatHt is composed of all transformer key and value pairs generated up to time t). Taking steps inHt space leads to gradual changes to model activations — which may be thought of as gradual reinterpretations of the past — that guide future generation in the desired direction. Let<> be the update toHt , such that generation with (<>) shifts the distribution of the generated text such that it is more likely to possess the desired attribute. Ht is initialized <

> Figure 1: Simplified illustration of the proposed approach in three phases. In Step 1, a forward pass is performed through the language model to compute the likelihood of a desired attribute using an attribute model that predicts<>. In Step 2, a backward pass updates the internal latent representations of the LM, using gradients from the attribute model, to increase the likelihood of the passage having the desired attribute. In Step 3, a new distribution over the vocabulary (<>) is generated from the updated latents(Het )and the current token <>. The next token is then sampled from the updated distribution. This process of updating the latents is repeated at each time-step, leading to a gradual transition towards the desired attribute. For computational efficiency, one may choose to modify only the latents within some window of the recent past, depicted as the dotted-red region. at zero and updated with gradients from an attribute model that measures the extent to which the generated text possesses the desired attribute (e.g. positivity). We rewrite the attribute model<> FORMULA>> and then make gradient based updates to <> as follows: <> (3) where <> is the step size, <> is the scaling coefficient for the normalization term. 1 This update step can be repeated m times; in practice we use3to10. Subsequently, a forward pass through the LM with the updated key-value pairs is performed to obtain the updated <>, where <>. The perturbed oet+1 is then used to generate a new distribution <> as in Equation 2. 3.3 ENSURING FLUENCY :ASCENDING log<> The approach described in the previous section is able to generate text tuned for a particular discriminator, but left unchecked it will quickly result in unrealistic adversarial or fooling examples (Szegedy et al., 2013; Nguyen et al., 2015) as the text moves into low probability regions. To com- bat this, we use the unconditional language model in two ways that ensure the fluency is maintained at or near the level of the unconditional language model (here GPT-2). Kullback–Leibler (KL) Divergence We update<> to minimize the KL divergence between the output distribution of the modified and unmodified language models in addition to the step above. In practice, this is accomplished by adding the quantities together before taking a gradient, though it can be visualized as two separate steps as in Figure 2. We scale the KL coefficient by a scalarKL , and in practice, setting this hyperparameter to 0.01 works well in general across tasks. Post-norm Geometric Mean Fusion In addition to minimizing KL divergence, which affects the past via<> , we perform post-norm fusion similarly to Stahlberg et al. (2018). This does not directly affect<> ; rather, it just serves to constantly tie the generated text to the unconditional <>LM distribution. We accomplish this by sampling from <>, where <> and <> are the unmodified and modified output distributions, respectively, and <> is a normalizing factor such that it forms a valid distribution. As <> this converges to the distribution from the updated LM, and as <> converges to the unconditional LM distribution. We find that in practice values for <> in the range 0.8-0.95 work well. 1 One normalization term is computed for each layer of the transformer. Figure 2: An oversimplified view into why steps that maximize both log<>and log<> are needed. The sentence under consideration is shown as a black dot, which is first pushed in the direction of maximizing <> and then in the ascend <> direction of maximizing <>. In practice we ascend <> higher use a single step and simply add the log probabilities; <

> we take steps in continuous space of hid- lower den representations H rather than in the discrete x higher <> (byte pair) space, and rather than resampling the entire sentence each step, we take one step inH space per byte-pair sample. 3.4 SAMPLING AND RANKING The attribute model<>in PPLM provides two functionalities: first, a score that can be used to rank samples based on the LL of the desired attribute (forward pass only; Step 1, Figure 1), and second, a gradient ascent direction to perform an update in the latent space (Step 2 & 3; Figure 1). The former can be used to generate r samples and rank them to choose the best one. This can serve as an additional method for attribute control in addition to sampling with updated latents. Further, to avoid the problem of repetitive, low quality text (Holtzman et al., 2018), we compute the mean over the Dist-1, Dist-2 and Dist-3 scores (for the generated passage), which is an indicator of repetitiveness (Li et al., 2015), and then discard samples with a mean score below a threshold. 4 EXPERIMENTS, RESULTS, AND EVALUATION In this section, we describe our evaluation methodology and then show controlled generation results under various attribute models. We also show use cases of PPLM in language detoxification and in controlled story telling. For all results reported in this section, we use top-k sampling (Fan et al., 2018) with k=10 to draw from the softmax distribution over the vocabulary. 4.1 EVALUATION METHODS AND ABLATION STUDY We evaluate to assess two properties: whether PPLM generates text that satisfies the desired attribute (topic or sentiment) and whether the quality of its text deteriorates as we intensify control of the attribute. Note we can always turn the control knob down to zero to disable control of attributes and reach the fluency of the original model. If desired, a user can tune the knobs at inference until a chosen tradeoff between attribute strength and fluency is reached. We evaluate using both automated methods and human annotators: Automated Eval.Perplexity is an automated measure of fluency, though its effectiveness has been questioned in open-domain text generation (Liu et al., 2016). We measure perplexity using a differ- ent pre-trained language model, GPT (Radford et al., 2018b). The diversity of text in the passages is measured using the number of distinct n-grams (normalized by the length of text) as in Li et al. (2015). We report Dist-1, Dist-2, and Dist-3 scores for the distinct 1-2-3-grams (measured across all samples generated for a given attribute control task, e.g. a specific topic for topic control). Such scores are an indicator of the diversity of the samples generated (Li et al., 2015). We also use external sentiment classifiers for sentiment evaluation. Human Eval.We consider two types of human annotation: fluency and A/B testing on attribute relevance. Annotators are asked to evaluate the fluency of each individual sample on a scale of 1-5, with 1 being “not fluent at all” and 5 being “very fluent,” as done in Lample et al. (2019). In the A/B testing for attribute relevance, we consider all combinatorial pairs of all four variants: B, BR, BC, and BCR (6 combinations). We then ask annotators to rank the pair on the desired attribute (e.g. topic relevance, sentiment strength), while allowing “neither” and “both” options to account for equally good/bad generations (Lample et al., 2019). We obtain annotations from nine external occupational annotators. Each pair of samples is evaluated by three individuals and we use majority-voting to Table 3: Comparison of different samples generated by (top row) baseline GPT-2 and (other rows) PPLM with different BoW corresponding to different topics (e.g.[Military]), all conditioned on a single prefix: "The issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focusedThe issue focused". Both directly optimized (inred) and related words (insoft red) are highlighted, showing how the optimization takes effect. <

> compute attribute relevance. For fluency, we use average of the three annotations. The method of generation is completely hidden and the order of samples in A/B testing is randomized. Ablation study and baselines.We conduct an ablation study with four variants:B: the baseline, unchanged GPT-2 LM, sampled once;BR: B but sampled r times, with best sample chosen based on the LL ranking and filtering based on Dist score;BC: update the latent representations(Het )and then sample once; and lastlyBCR: update the latent representations(Het )and generate r samples, choose the best sample based on the LL score (after filtering out samples with low Dist scores). As baseline approaches we considerCTRL: (Keskar et al., 2019), a recent language model;GPT2-FT- RL: a GPT-2 LM fine-tuned for human evaluated positivity with RL (Ziegler et al., 2019); andWD: a weighted decoding baseline in which the B LM’s outputs are weighted directly toward maximizing <> (Ghazvininejad et al., 2017); see Section S7 for details, and Section S11 for hyperparameters. 4.2 BOW ATTRIBUTE MODELS The simplest attribute model we use gives the log of the sum of likelihoods of each word in some predefined Bag of Words (BoW). Given a set of keywords <> that specify a topic of interest and the output distribution of the language model <>, the log likelihood is: <> (4) We construct BoWs that represent seven distinct topics: SCIENCE, MILITARY, LEGAL, COMPUTERS, SPACE, POLITICS and RELIGION (see Section S17 for complete word lists). Samples are shown in Table 3, generated from a single prefix, while being controlled towards each topic. Interestingly, we find that increasing the probability of generating the words in the bag also increases the probability of generating related topical words not in the BoW (e.g. in the[Science] sample shown in Table 3, note that question and philosophers are sampled before the first BoW word,laws). Table S17 shows the gradual change of topic intensity under fine-grained control. We found that the optimization procedure works better with updating representations from the past over a finite window and using an adaptive normalization scheme (see Section S11.3). For automatic and human evaluation, we generate 420 samples evenly distributed among seven BoW attribute models and 20 prefixes (see the full list in Section S15), for each of the four variants de- scribed in the ablation study. See Section S8 for further details on evaluation and results. Table 4 shows that human annotators find text from BCR (51.7%) and BC (46.9%) to be significantly more Table 4: For each treatment in the ablation study, we report mean std-dev across (human and automated) fluency metrics. The topic (%) reports the fraction of samples matching the target topic, as evaluated by human annotators. Table S8 provides per-topic results. Approaches BC and BCR demonstrate significant control over the topic of the generated text, while retaining similar diversity (Dist-1, Dist-2, Dist-3) scores and minimal degradation in Perplexity and Fluency evaluations vs the baseline LM (B). The gain from ranking and choosing from multiple samples BR over B is limited (4.7%). The gain in topic-accuracy from latent (Het ) manipulation (from B to BC) is significantly higher (35.8%). Perplexity is computed using the GPT LM (Radford et al., 2018a), which differs from the LM generating text (GPT-2). For CTRL and WD, since human evaluation is performed in comparison with BCR via A/B testing, we report the numbers for BCR as well from these comparisons, for the human evaluated metrics. Further, we consider one sample per prefix for CTRL, resulting in fewer samples and higher Dist-1, 2, 3 scores as a consequence. PPLM outperforms CTRL and WD on topic-relevance, while being comparable on fluency scores. <> on topic than B (15.8%) and BR (11.1%). With only a slight degradation in fluency scores, passages generated with manipulated latents (BCR and BR) are significantly on topic, demonstrating the de- sired attribute control on this task. The Dist-1, Dist-2 and Dist-3 scores, which accounts for diversity of text across the generated passages, are similar across all four ablation approaches. Further, BCR slightly outperforms CTRL (51.7% & 50.0%), and significantly outperforms WD (36 %). BC itself outperforms WD (36 %). BCR, CTRL and WD all score similarly on the fluency metric. We note that gradient-based latent updates have significantly greater influence on topic relevance (R with or without C) than reranking based on the score (C with or without R), showing that shifting meaning in latent space is more effective than shifting the output distribution directly through reweighting. The effectiveness of shifting latents is further corroborated by the WD’s relatively worse performance. WD directly controls the output distribution, which will not lead to increased probability of sampling words from outside the bag that are related to the topic. Finally, there is a large variance in the extent of controllability across topics (Table S8). We find that some topics (religion, science, politics) are easier to control for compared to others (computers, space). Section S9 considers unusual or nonsensical combinations of prefixes and attributes (e.g. prefix ‘potato’ and topic ’religion’), and we find that even for these settings PPLM is able to successfully control for the desired attribute, often with hilarious twists! 4.3 DISCRIMINATOR ATTRIBUTE MODELS While BoW models have been demonstrated to be able to control text attributes such as sentiment (e.g., Li et al. (2018) rely on extracting a set of attribute-based phrases to control the sentiment during style transfer), being able to control attributes using more sophisticated discriminators is desirable when it is difficult to express the attribute with a simple bag of words. We train a discriminator on a dataset with input sentences x and corresponding labels yx . For an input xof length t, we compute ox and train fon the mean (<>) of the embeddings across time. All :t discriminators in this work consist of a single layer classifier that predicts the target label from <> The number of parameters in this layer is (embedding-dimension (e) number of attributes (a) + number of attributes (a)), which is negligible compared to the number of parameters in the LM model itself (Table 2). Although the loss is a function of the entire sequence, here we adopt a greedy approach, similar to Ebrahimi et al. (2018); Wallace et al. (2019), in which we optimize for Table 5: Sentence samples in triplets, generated by {baseline GPT-2, PPLM-Discrim POSITIVE , PPLM-Discrim NEGATIVE }, conditioned on prefixes:The chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chickenThe chicken&The countryThe countryThe countryThe countryThe countryThe countryThe countryThe countryThe countryThe countryThe countryThe countryThe countryThe countryThe countryThe countryThe country. Words related to the sentiment are highlighted (in soft red). Each triplet is generated from the same random seed. <> a higher-probability of the sequence having a specific attribute by considering changes only to the next token to be generated. This objective can be described as follows, where f is the discriminator: <> (5) Note that <> is a function of <> . Further, <>, which depends on <>. In the limit, minimizing the objective in Equation 5 corresponds to choosing <> that produces the optimal <> that maximizes <>. However, this limits the diversity of the generated text and could potentially lead to language degeneration (Holtzman et al., 2019). Alternatively, we focus on a softer optimization approach where we aim to shift the distribution <> towards one that in expectation has a higher likelihood of having the desired attribute a. Possible approaches to accomplishing this are using REINFORCE (Williams, 1992) and the Gumbel-Softmax trick (Jang et al., 2016). However, both of these would slow down convergence. Instead, as in Dai et al. (2019a), we use the distribution <> (instead of a hard sample<> ), and feed it forward to obtain (a biased) estimate of the next token’s embedding and then update<> . The sentiment discriminator here distinguishes sentiment between POSITIVE and NEGATIVE and is trained on the SST-5 dataset (Socher et al., 2013). Table 5 shows PPLM-Discrim generated samples in triplets: uncontrolled, controlled for POSITIVE sentiment, controlled for NEGATIVE sentiment. For automatic and human evaluation, we use 15 prefixes (see the full list in Section S15) to generate 45 samples for each of two sentiment classes:very positive and very negative. Note that even though the sentiment discriminator is trained with movie review data, the prefixes (e.g. “The painting”, “The potato”, “The country”) we used are not necessarily associated with movie reviews. This supports the generality of our approach: an attribute model trained with data from a different domain can still provide meaningful gradients. Table 6 shows evaluation results. For human evaluation, we obtain 1620 annotations for the ablation study and 495 for baseline comparisons from the annotators distributed across the samples and sentiments. Unlike the topic control setting, sampling and ranking results in a considerable increase in attribute accuracy (19:3%!41:5%), because the prior probability of sampling, say, a negative sentence, is relatively high. BC results in a decrease in fluency when compared to B, while being significantly more consistent with the desired attribute (19:3%!39:6%). With latent manipulation and ranking (BCR), we see a significant increase in attribute control accuracy (73:7%) while retain- ing fluency similar to B and BR. Further, the gain in sentiment accuracy from re-sampling is larger in the case of manipulated latents vs non-manipulated (34:1%increase from BC to BCR>22:2% increase from B to BR), indicating that these two approaches may be profitably combined. We also evaluate attribute control with an external sentiment classifier trained on IMDB movie reviews (Maas et al., 2011), which is a different dataset from the one used to train the attribute model (Socher et al., 2013), and the same rough story holds, albeit with smaller gaps between approaches. We compare to baselines CTRL, GPT2-FT-RL, and WD. BCR performs comparably to CTRL (73.7% and 80.0%), and BR, BC and BCR all outperform GPT2-FT-RL, the GPT-2 LM fine tuned for positivity, and WD. Table 6: Evaluation of models/ variants on the sentiment control task, with meanstd-dev reported across fluency metrics. Sentiment accuracy reports the fraction of samples with an accurate tar- get sentiment. Approach BCR provides significant control over sentiment while showing minimal degradation in fluency. See Table S9 for full results on individual sentiments. *GPT2-FT-RL is only evaluated for the positivity half of the task, as it is fine-tuned only for positivity (Ziegler et al., 2019). For human evaluation metrics, we compare the baselines CTRL, GPT2-FT-RL and WD with BCR and perform A/B style testing. We include both numbers for comparison. <> 4.4 LANGUAGE DETOXIFICATION Language models trained with large corpora of Internet data reflect biases and discrimination existing in the data. A recent paper by Wallace et al. (2019) conducted adversarial attacks that make GPT-2 produce racist output when given a carefully optimized trigger string as prefix. They also find that when simply using “Blacks” as prefix, 2% of GPT-2 samples contain explicit racism. Other prefixes (e.g., “Asians” or “Jews”) are mentioned but no percentage is reported. We conduct experiments and report the baseline toxicity percentages to be 10% (“Asians”), 12% (“Jews”) and 8% (“Blacks”). With adversarial triggers generated from the released codebase by Wallace et al. (2019) the average toxicity percentage is 63.6%. Further details can be found in Section S13. PPLMs can be easily adapted for language detoxification by plugging in a toxicity classifier as the attribute control model and update latents with the negative gradient. We train a single layer classifier on the toxicity data from the Toxic Comment Classification Challenge (Jigsaw) and show that with a similar hyper-parameter setting as other PPLM-Discrim methods, it works well on both natural prompts and adversarial triggers. For natural prompts percentages of toxicity are 6%, 4% and 10%, respectively, and for adversarial triggers it drastically dropped to 4.6% on average, with statistical significance. Details on the annotation procedure and full table of percentage and p-values can be found in Table S23 and Section S13. Note that a model for detoxifying language can also potentially be maliciously used for generating toxic language, a topic we briefly discuss in Section S6. 4.5 CONTROLLED STORY WRITING We explore controlled generation for assistive story writing (Peng et al., 2018; Luo et al., 2019; Yao et al., 2019; Fan et al., 2018). Using uncontrolled LMs for assistive art creation can be difficult. To help with the structure, we use predefined story skeletons often used in improvisation (Adams). We fill in the blank between these prefixes with a PPLM. See examples in Table S20 and Table S21. 5 CONCLUSION We have presented PPLM, a plug and play method for controlled language generation that flexibly combines a large, pre-trained LM and a BoW or a small, easy-to-train discriminator. In Section S6 we discuss the ethics of controlled LMs. PPLM achieves fine-grained control of attributes via a simple gradient-based sampling mechanism. Because PPLMs can flexibly control generation while maintaining fluency, they hold great promise for enabling the next generation of language models. ACKNOWLEDGEMENTS The authors are grateful to Bryan McCann for providing samples for the CTRL baseline, Joel Lehman for discussion regarding the ethical implications for this work, Jiale Zhi for help with the computational framework, Colan Chen for creating associated artwork for the blog, Avishek Joey Bose for helpful discussions, Julien Chaumond, Lysandre Debut, Thomas Wolf, and the Hugging Face team for co-producing the PPLM demo and helping integrate the code into their transformers repository, all the annotators at Uber, HKUST and Caltech for their labeling, and members of the Deep Collective research group for helpful discussion, ideas, and feedback on experiments. REFERENCES Kenn Adams. Improv encyclopedia story spine. http://improvencyclopedia.org/ games/Story_Spine.html. (accessed September 20, 2019). Ashutosh Baheti, Alan Ritter, Jiwei Li, and Bill Dolan. Generating more interesting responses in neural conversation models with distributional constraints. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 3970–3980, 2018. Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin. A neural probabilistic language model.Journal of machine learning research, 3(Feb):1137–1155, 2003. Yun Chen, Victor OK Li, Kyunghyun Cho, and Samuel R Bowman. A stable and effective learning strategy for trainable greedy decoding.arXiv preprint arXiv:1804.07915, 2018. Ning Dai, Jianze Liang, Xipeng Qiu, and Xuanjing Huang. Style transformer: Unpaired text style transfer without disentangled latent representation.arXiv preprint arXiv:1905.05621, 2019a. Zihang Dai, Zhilin Yang, Yiming Yang, William W Cohen, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context.arXiv preprint arXiv:1901.02860, 2019b. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp. 4171–4186, 2019. Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. HotFlip: White-box adversarial ex- amples for text classification. InProceedings of the 56th Annual Meeting of the Associa- tion for Computational Linguistics (Volume 2: Short Papers), pp. 31–36, Melbourne, Aus- tralia, July 2018. Association for Computational Linguistics. doi: 10.18653/v1/P18-2006. URL https://www.aclweb.org/anthology/P18-2006. Yanai Elazar and Yoav Goldberg. Adversarial removal of demographic attributes from text data. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Process- ing, pp. 11–21, Brussels, Belgium, October-November 2018. Association for Computational Lin- guistics. doi: 10.18653/v1/D18-1002. URLhttps://www.aclweb.org/anthology/ D18-1002. Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical neural story generation.arXiv preprint arXiv:1805.04833, 2018. Jessica Ficler and Yoav Goldberg. Controlling linguistic style aspects in neural language generation. InProceedings of the Workshop on Stylistic Variation, pp. 94–104, 2017. Marjan Ghazvininejad, Xing Shi, Jay Priyadarshi, and Kevin Knight. Hafez: an interactive poetry generation system. InProceedings of ACL 2017, System Demonstrations, pp. 43–48, Vancouver, Canada, July 2017. Association for Computational Linguistics. URLhttps://www.aclweb. org/anthology/P17-4008. Jiatao Gu, Graham Neubig, Kyunghyun Cho, and Victor OK Li. Learning to translate in real-time with neural machine translation.arXiv preprint arXiv:1610.00388, 2016. Jiatao Gu, Kyunghyun Cho, and Victor OK Li. Trainable greedy decoding for neural machine translation.arXiv preprint arXiv:1702.02429, 2017. Ari Holtzman, Jan Buys, Maxwell Forbes, Antoine Bosselut, David Golub, and Yejin Choi. Learning to write with cooperative discriminators.CoRR, abs/1805.06087, 2018. URLhttp://arxiv. org/abs/1805.06087. Ari Holtzman, Jan Buys, Maxwell Forbes, and Yejin Choi. The curious case of neural text degener- ation.arXiv preprint arXiv:1904.09751, 2019. Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P. Xing. Controllable text generation.CoRR, abs/1703.00955, 2017. URLhttp://arxiv.org/abs/1703.00955. Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. 2016. Jigsaw. Toxic comment classification challenge. https://www.kaggle.com/c/ jigsaw-toxic-comment-classification-challenge/. Accessed: 2019-11-13. Nitish Shirish Keskar, Bryan McCann, Lav Varshney, Caiming Xiong, and Richard Socher. CTRL - A Conditional Transformer Language Model for Controllable Generation. arXiv preprint arXiv:1909, 2019. Yuta Kikuchi, Graham Neubig, Ryohei Sasano, Hiroya Takamura, and Manabu Okumura. Con- trolling output length in neural encoder-decoders. InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 1328–1338, Austin, Texas, Novem- ber 2016. Association for Computational Linguistics. doi: 10.18653/v1/D16-1140. URL https://www.aclweb.org/anthology/D16-1140. Guillaume Lample, Sandeep Subramanian, Eric Smith, Ludovic Denoyer, Marc’Aurelio Ranzato, and Y-Lan Boureau. Multiple-attribute text rewriting. InInternational Conference on Learning Representations, 2019. URLhttps://openreview.net/forum?id=H1g2NhC5KQ. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. A Diversity-Promoting Objective Function for Neural Conversation Models.arXiv e-prints, art. arXiv:1510.03055, Oct 2015. Juncen Li, Robin Jia, He He, and Percy Liang. Delete, retrieve, generate: A simple approach to sentiment and style transfer.CoRR, abs/1804.06437, 2018. URLhttp://arxiv.org/abs/ 1804.06437. Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau. How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 2122–2132, 2016. Fuli Luo, Damai Dai, Pengcheng Yang, Tianyu Liu, Baobao Chang, Zhifang Sui, and Xu Sun. Learning to control the fine-grained sentiment for story ending generation. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 6020–6026, 2019. Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. Learning word vectors for sentiment analysis. InProceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pp. 142–150, Portland, Oregon, USA, June 2011. Association for Computational Linguistics. URLhttp: //www.aclweb.org/anthology/P11-1015. Christopher D Manning, Christopher D Manning, and Hinrich Schütze.Foundations of statistical natural language processing. MIT press, 1999. Nicholas Metropolis, Arianna W Rosenbluth, Marshall N Rosenbluth, Augusta H Teller, and Edward Teller. Equation of state calculations by fast computing machines. The journal of chemical physics, 21(6):1087–1092, 1953. Nathan Ng, Kyra Yee, Alexei Baevski, Myle Ott, Michael Auli, and Sergey Edunov. Facebook fair’s wmt19 news translation task submission.arXiv preprint arXiv:1907.06616, 2019. Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High con- fidence predictions for unrecognizable images.The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015. Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space. InThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. Nanyun Peng, Marjan Ghazvininejad, Jonathan May, and Kevin Knight. Towards controllable story generation. InProceedings of the First Workshop on Storytelling, pp. 43–49, 2018. Martin Potthast, Tim Gollub, Kristof Komlossy, Sebastian Schuster, Matti Wiegmann, Erika Pa- tricia Garces Fernandez, Matthias Hagen, and Benno Stein. Crowdsourcing a large corpus of clickbait on twitter. InProceedings of the 27th International Conference on Computational Lin- guistics, pp. 1498–1507, 2018. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language un- derstanding by generative pre-training. URL https://s3-us-west-2. amazonaws. com/openai- assets/researchcovers/languageunsupervised/language understanding paper. pdf, 2018a. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language un- derstanding by generative pre-training. URL https://s3-us-west-2. amazonaws. com/openai- assets/researchcovers/languageunsupervised/language understanding paper. pdf, 2018b. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners.OpenAI Blog, 1(8), 2019. Gareth O Roberts and Jeffrey S Rosenthal. Optimal scaling of discrete approximations to langevin diffusions. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 60(1): 255–268, 1998. Gareth O Roberts, Richard L Tweedie, et al. Exponential convergence of langevin distributions and their discrete approximations.Bernoulli, 2(4):341–363, 1996. Abigail See, Stephen Roller, Douwe Kiela, and Jason Weston. What makes a good conversation? How controllable attributes affect human judgments.arXiv e-prints, art. arXiv:1902.08654, Feb 2019. Claude Elwood Shannon. A mathematical theory of communication.Bell system technical journal, 27(3):379–423, 1948. Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi S. Jaakkola. Style transfer from non-parallel text by cross-alignment. CoRR, abs/1705.09655, 2017. URLhttp://arxiv.org/abs/ 1705.09655. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. InProceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pp. 1631–1642, Seattle, Washington, USA, October 2013. Association for Computa- tional Linguistics. URLhttps://www.aclweb.org/anthology/D13-1170. Felix Stahlberg, James Cross, and Veselin Stoyanov. Simple fusion: Return of the language model. arXiv preprint arXiv:1809.00125, 2018. Nishant Subramani, Sam Bowman, and Kyunghyun Cho. Can unconditional language models re- cover arbitrary sentences?arXiv preprint arXiv:1907.04944, 2019. Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfel- low, and Rob Fergus. Intriguing properties of neural networks.CoRR, abs/1312.6199, 2013. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Infor- mation Processing Systems, pp. 6000–6010, 2017. Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for nlp.arXiv preprint arXiv:1908.07125, 2019. Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning.Machine learning, 8(3-4):229–256, 1992. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie Brew. Transformers: State- of-the-art natural language processing, 2019. Lili Yao, Nanyun Peng, Ralph Weischedel, Kevin Knight, Dongyan Zhao, and Rui Yan. Plan-and- write: Towards better automatic storytelling. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pp. 7378–7385, 2019. Kyra Yee, Nathan Ng, Yann N Dauphin, and Michael Auli. Simple and effective noisy channel modeling for neural machine translation.arXiv preprint arXiv:1908.05731, 2019. Lantao Yu, Weinan Zhang, Jun Wang, and Yong Yu. Seqgan: Sequence generative adversarial nets with policy gradient. InThirty-First AAAI Conference on Artificial Intelligence, 2017. Lei Yu, Phil Blunsom, Chris Dyer, Edward Grefenstette, and Tomas Kocisky. The neural noisy channel.arXiv preprint arXiv:1611.02554, 2016. Lei Yu, Laurent Sartran, Wojciech Stokowiec, Wang Ling, Lingpeng Kong, Phil Blunsom, and Chris Dyer. Putting machine translation in context with the noisy channel model.arXiv preprint arXiv:1910.00553, 2019. Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences.arXiv preprint arXiv:1909.08593, 2019. URLhttps://arxiv.org/abs/1909.08593. SUPPLEMENTARY INFORMATION FOR : PLUG AND PLAY LANGUAGE MODELS : A SIMPLE APPROACH TO CONTROLLED TEXT GENERATION S6 ETHICS OF CONTROLLED LANGUAGE MODELS There has recently been a substantial discussion around the ethics of capable language models (Rad- ford et al., 2019; Keskar et al., 2019), both in their potential to recapitulate problematic social biases and for them to be directly abused for societal harm (e.g. to generate disinformation). While one aim of this paper is to suggest a mechanism to detoxify language models (Section 4.4), we also acknowl- edge that nearly the same mechanism could be exploited to instead create more toxic language. Such possibilities are inherent to general-purpose technologies such as machine learning, and we believe that on balance this work creates more value than risks. S7 DETAILS ON BASELINE METHODS We consider three baselines: CTRL, GPT2-FT-RL, and WD. The first two are strong baselines where large language models are trained (or fine-tuned) specifically to generate texts conditioned on certain attributes, while WD is considered a weak baseline based on a direct integration of the conditioning into the decoding. For each baseline, we generate data from their method, and conduct the same human and automated evaluations. For human evaluation of attribute relevance, we match baseline data with our method (BCR in the ablation study), and pass to human annotators for an A/B testing style annotation. As in the ablation study, human annotators are given a pair of texts, one from baseline, one from ours, with orders randomized and source hidden, and asked to rank which one is more topic or sentiment relevant, while having the options of “both” and “neither”. On top of that, we have human annotators to give the fluency score of each text sample under each method individually. And automated evaluations of perplexity, sentiment, etc. are also done individually. S7.1 CTRL The recent conditional language model, CTRL, from Keskar et al. (2019), trains a 1.6B LM condi- tioned on around 50 control codes. We use the official released codebase 2 and their open-sourced model to generate samples for the CTRL baseline. Out of the 7 topics considered in PPLM-BoW, we found that 5 can be matched with a specific control code in CTRL. We append a secondary code "Text:" to each primary control code, per the author’s suggestion, to encourage more fluent and longer passages. The 2 topics missing a match with CTRL are: Military, Space. For positive and negative sentiments in PPLM-Discrim, we match with the Reviews control code and append a high and low rating score. The matched attributes and control codes are listed in Table S7. Under this setting, for each control code we generate texts prompted by the same prefixes used for corresponding PPLM attribute model (20 for PPLM-BoW, 15 for PPLM-Discrim). For example, “In summary” and “To review,” for PPLM-BoW, and “The chicken”, “The lake” for PPLM-Discrim. Due to the near-greedy sampling method CTRL uses, for each prefix it generates one sample. Hence we have 20 samples for each matching topic with PPLM-BoW, and 15 samples for positive and 15 for negative. S7.2 GPT2-FT-RL A recently released GPT-2 model fine-tuned using human feedback, from Ziegler et al. (2019), showed success in summarization and text continuation in desired styles. To compare with PPLM, 2 CTRL codebase:https://github.com/salesforce/ctrl Table S7: Control codes used for the model from Keskar et al. (2019) for experiments in Section 4. <> we run GPT2-FT-RL 3 to generate positive texts on the same prefixes used in our PPLM-Discrim experiment. For each prefix, we generate three GPT2-FT-RL samples, and pair them with those generated from PPLM (BCR in the ablation study) randomly. S7.3 WEIGHTED DECODING (WD) We consider a simple baseline based on a direct integration of the conditioning into the decoding procedure, similar to the approach from Ghazvininejad et al. (2017). Topic Control with Bag of Words In Ghazvininejad et al. (2017), the authors consider increasing the likelihood of sampling from a bag of key-words by performing beam-search with a modified scoring function. <>; where 1 BoW (<>) is an indicator function indicating if the tokenwi is present in the bag BoW. Since, it has been shown that beam-search results in degradation of language for GPT-2 (Holtzman et al., 2019), we consider top-5 sampling from a distribution <> defined such that: <> where <> and FORMULA is the distribution over the vocabulary as predicted by the GPT-2 LM . For the experiments in Section 4, we set <>. Sentiment Control with Discriminator Here, we implemented weighted decoding similarly for sentiment control. Here we wish to incorporate the score from the attribute model into decoding. To control for stylea^, instead of sampling from the distributionpt+1 , we sample from <> defined as: <> <> is the probabilty of the sequence <> possessing attribute <> assigned by the attribute model. By Bayes’ rule, <>, and we do top-5 sampling from this distribution. Recall that <> under the language model. S8 FURTHER DETAILS ON HUMAN AND AUTOMATED EVALUATION We conduct evaluations on attribute relevance and language fluency, both including human and automated evaluation. For topic relevance (a.k.a attribute relevance where the attribute is a topic, in our case represented by a BoW), we rely entirely on human annotation. For sentiment relevance, we rely on human annotation as well as a separately trained sentiment classifier. We also performed a “clickbait” style control, for which the effectiveness relies on human annotation. 3 GPT2-FT-RL codebase:https://github.com/openai/lm-human-preferences For fluency, we use human annotations (between 1 to 5) and automated methods: perplexity, Dist-1, Dist-2, and Dist-3 scores. The number of human evaluations are as below: <> In ablation studies, the generation procedure for BCR, BR and BC is always initiated from the same random seeds. The same set of random seeds that lead to the samples chosen with BCR are stored and used to generate the samples with B. The full table of all these measures, human and automated, on PPLM-BoW, seperated by sentiment and style, is in Table S8. Included also are strong baselines (CTRL and WD) for each sentiment. The human annotated topic relevance is further visualized in Figure S3. The fluency scores, while being across {B, BC,BR, BCR,} methods in the table, when shown in distribution are very similar, as seen in Figure S5. The full table of all these measures, human and automated, on PPLM-discrm sentiments, is in Ta- ble S9. Included also are strong baselines (CTRL, WD and GPT2-FT-RL) for each topic. The human annotated sentiment and style (e.g. “Clickbait”) relevance is further visualized in Figure S4, along with congregated measures: all sentiments, all discriminators, all topics. The fluency scores again have similar distributions across {B, BC,BR, BCR,} methods, as seen in Figure S6. <

> Figure S3: Topic relevance by human evaluation. We can see that taking a PPLM gradient step (B!BC) makes a big difference. Reranking is mostly helpful (B!BR; BC!BCR). We can also see a rough distribution of various topics in unperturbed, GPT-2 generation (B), which possibly mirrors the distribution of topis in its training data. Some topics, like science, naturally appear rather frequently. S9 ODD COMBINATION OF TOPICS AND PREFIXES It is interesting to see how PPLM can steer the text generation when the topic and prefix combination appears odd or illogical. For example, will “The potato” still prompt sensible text generation under the topic R ELIGION ? In this study we design a set of odd combinations, as bellow. Table S8: Full result of human and automated evaluation of PPLM-BoW, attribute relevance and language fluency. This is a detailed version of Table 4, where results were averaged over all topics. Results here correspond to the average over all samples in each topic, for each method in the ablation study (B, BC, BR, BCR), and in baselines (CTRL, WD). Perplexity is computed based on an external LM (Radford et al., 2018a), that is different from the LM generating text. <

> Figure S4: Bar charts of discriminator relevance by human evaluation, together with different ver- sions of combined results. <

> Table S9: Full result of human and automated evaluation of PPLM-Discrim, attribute relevance and language fluency. The top two rows are a detailed version of Table 6, where results were averaged over both sentiments (except for GPT2-FT-RL, where there is only positive sentiment). The last row is the additional C LICKBAIT style control, where there is only ablation study and no baseline comparison. Results here correspond to the average over all samples in each sentiment and style, for each method in the ablation study (B, BC, BR, BCR), and in baselines (CTRL, GPT-2-FT-RL, WD). Perplexity is computed based on an external LM (Radford et al., 2018a), that is different from the LM generating text. <

> We found that PPLM control is easy even under those scenarios. We had to increase the strength two or three fold (to0:02or0:03as opposed to0:01in most studies) to allow for a stronger influence of attribute, but this is as expected: the strength parameter is a knob that user can tune to reach fine-grained control. The resulting generation is included in Table S10 - Table S16. Table S10: Examples generated from a designed odd combination of topic and prefix pairs. The topic here is[Military]. We show that PPLM is still able to generate fluent, sensible and interesting samples, respecting both the topic and the prefix. <> Table S11: Examples generated from a designed odd combination of topic and prefix pairs. The topic here is[Legal]. We show that PPLM is still able to generate fluent, sensible and interesting samples, respecting both the topic and the prefix. <> Table S12: Examples generated from a designed odd combination of topic and prefix pairs. The topic here is[Computers]. We show that PPLM is still able to generate fluent, sensible and inter- esting samples, respecting both the topic and the prefix. <> Table S13: Examples generated from a designed odd combination of topic and prefix pairs. The topic here is[Politics]. We show that PPLM is still able to generate fluent, sensible and interesting samples, respecting both the topic and the prefix. <> Table S14: Examples generated from a designed odd combination of topic and prefix pairs. The topic here is[Religion]. We show that PPLM is still able to generate fluent, sensible and interesting samples, respecting both the topic and the prefix. <> Table S15: Examples generated from a designed odd combination of topic and prefix pairs. The topic here is[Space]. We show that PPLM is still able to generate fluent, sensible and interesting samples, respecting both the topic and the prefix. <> Table S16: Examples generated from a designed odd combination of topic and prefix pairs. The sentiment here is[Positive] and[Negative]. We show that PPLM is still able to generate fluent, sensible and interesting samples, respecting both the topic and the prefix. <> S10 FINE-GRAINED CONTROL WITH PPLM-BOW Table S17 shows the subtle effect when you turn the step sizeup, while keeping everything else (hyperparameters, text prefix) the same. S11 HYPERPARAMETERS We list, in Table S18, the full set of hyperparameters used in each task in the experiments section, corresponding to results in Table 4 and Table 6, as well as in Section 4.4. In addition, we explain in details three hyperparameters and their effect, below. S11.1 E RLY STOPPING OF LATENT UPDATES Degeneration (the occurrence of repetitive words) is a known issue with language generation (Holtz- man et al., 2019), and we found it to be a case in PPLM-BoW when the update step size t is too large. The model tends to degenerate towards repeating certain keywords targeted in the optimiza- tion (e.g. words in the BoW). In this case, we can either reduce, or use the trick of early stopping latent updates. Examples shown in Table S19. With the exact same setting, but just stopping latent updates after 20 time steps, the samples show much less degeneration. S11.2 FINITE HORIZON UPDATE As opposed to updating the entire vectorHt , which consists of key-value pairs corresponding to every token in the prefix, we consider modifying the key-value pairs corresponding to the most recentwtokens. At each time-stept, we only modify <>. This means that we modify Hi at mostwtimes, and requires lesser computation that updating the whole past. We find that w= 5produces more fluent passages for control with the bag of words. For control with the neural attribute model, we update the entire latent history. S11.3 ADAPTIVE GRADIENT NORMALIZATION For the bag-of-words based attribute model, what we wish to enforce is that a word from the bag appears at least once in the generated passage and not at every time-step. To account for this, instead of normalizing directly by the gradient norm as in Equation 3, we normalize by the maximum gradient norm over time. This implies that we make smaller updates when it is less likely for 3 We choose top 3 samples from a single batch of 10 here S17: Comparison of different sentences generated by (top row) unconditioned GPT-2 and <> a word from the bag of words to appear. Formally, the normalization constant at time-step t is: <> Table S18: The full set of hyperparameters used in each task in the experiments section. Note that for PPLM-BoW, we select three of the highest scoring samples from a single batch ofr= 10. For PPLM-Discrim, we get 1 sample per batch, across 3 batches ofr= 10. <> Table S19: The effect of using early stopping of latent updates to prevent sample degeneration. Before (Latent updates at every generation step) After (Latent updates for only the first 20 generation steps) <> Table S20: Skeleton story generation with different attribute models. Each story is generated within <> S12 ADDITIONAL DETAILS AND SAMPLES OF CONTROLLED STORY TELLING Table S20 and Table S21 show examples of the controlled story telling with skeleton. Table S22 shows samples of combinging multiple (three to four) attribut models, across different types (PPLM- BoW and PPLM-Discrim). S13 FURTHER DETAILS ON LANGUAGE DETOXIFICATION We ask a group of four volunteers to assign binary toxicity labels to texts generated from both the original GPT-2 and toxicity-controlled GPT-2 and to texts prompted with natural words as well as adversarial triggers. Natural prompts include “Blacks”, “Asians”, “Jews” as suggested in Wallace et al. (2019). 9 Adversarial triggers are generated using the released codebase from the same work. The percentages of toxicity, before (with uncontrolled GPT-2) and after (with toxicity-controlled GPT-2), are reported in Table S23. S14 MORE EXAMPLES We include more PPLM controlled generation examples in Table S24 – Table S27. S15 PREFIXES USED IN PPLM EVALUATION We consider 20 prefixes as sentence starters for evaluating PPLM-BoW generation, chosen randomly from www2.eit.ac.nz/library/ls_guides_sentencestarters.html. For PPLM-Discrim, we use 15 prefixes. The entire set is below. PPLM-Bow “In summary”, “This essay discusses”, “Views on”, “The connection”, “Foundational to this is”, “To review,”, “In brief,”, “An illustration of”, “Furthermore,”, “The central theme”, “To conclude,”, “The key aspect”, “Prior to this”, “Emphasised are”, “To summarise”, “The relationship”, “More importantly,”, “It has been shown”, “The issue focused on”, “In this essay”. PPLM-Discrim “Once upon a time”, “The book”, “The chicken”, “The city”, “The country”, “The horse”, “The lake”, “The last time”, Table S21: More examples of skeleton story generation with different attribute models. Each story <> S16 COMBINING MULTIPLE CONTROLLERS FOR INSPIRATION Earlier we demonstrated attribute control using a single attribute model or two attribute models of the same type (e.g. BoW from two separate topics). Here we mix different types of attribute models (BoW and discriminator). For example, we can control the generation toward a mixed topic about WINTER , P OLITICS , K ITCHEN , while turning POSITIVE . See examples in Table S22. <

> Figure S5: Histogram illustrating the distribution of fluency scores based on controlled generated with PPLM-BoW from the four methods considered for ablation study. We find that fluency scores from all four approaches are similarly distributed. <

> Figure S6: Histogram illustrating the distribution of fluency scores based on controlled generated with PPLM-Discrim from the four methods considered for ablation study. We find that fluency scores from all four approaches are similarly distributed. S17 WORD LISTS FOR BAG OF WORDS APPROACHES We curate word lists fromwww.enchantedlearning.com/wordlist. <

> Table S22: Examples of attribute controlled text generation with multiple knobs. We train a clickbait discriminator using the dataset from Potthast et al. (2018) <> Table S23: Language detoxification applied to natural prompts and adversarial triggers. Shown are number of toxic passages / number of samples annotated, and percentage of toxicity. The column p-value shows the statistical significance of "After" lower than "Before". <> Table S24: Comparison of different samples generated with different prefixes using the same PPLM- BoW control under the[Military]topic. All samples are generated using exact same hyperparam- eters. <> Table S25: Comparison of different samples generated with different prefixes using the same PPLM- BoW control under the[Space]topic. All samples are generated using exact same hyperparameters. <> Table S26: Comparison of different samples generated with different prefixes using the same PPLM- BoW control under the[Science]topic. All samples are generated using exact same hyperparame- ters. <> Table S27: Comparison of different samples generated with different prefixes using the same PPLM- BoW control under the[Politics]topic. All samples are generated using exact same hyperparame- ters. <> <> <> <> <> <> <> Predicting Performance for Natural Language Processing Tasks Mengzhou Xia, Antonios Anastasopoulos, Ruochen Xu, Yiming Yang, Graham Neubig Language Technologies Institute, Carnegie Mellon University {mengzhox,aanastas,yiming,gneubig}@cs.cmu.edu ruochenx@gmail.com Abstract Given the complexity of combinations of tasks, languages, and domains in natural language processing (NLP) research, it is computationally prohibitive to exhaustively test newly proposed models on each possible experimental setting. In this work, we attempt to explore the possibility of gaining plausible judgments of how well an NLP model can perform under an experimental setting, with.out actually training or testing the model. To do so, we build regression models to predict the evaluation score of an NLP experiment given the experimental settings as input. Experimenting on 9 different NLP tasks, we find that our predictors can produce meaningful predictions over unseen languages and different modeling architectures, outperforming reasonable baselines as well as human experts. Going further, we outline how our predictor can be used to find a small subset of representative experiments that should be run in order to obtain plausible predictions for all other experimental settings. 1 Introduction Natural language processing (NLP) is an extraordinarily vast field, with a wide variety of models being applied to a multitude of tasks across a plenitude of domains and languages. In order to mea.sure progress in all these scenarios, it is necessary to compare performance on test datasets represent.ing each scenario. However, the cross-product of tasks, languages, and domains creates an explosion of potential application scenarios, and it is infeasible to collect high-quality test sets for each. In addition, even for tasks where we do have a wide variety of test data, e.g. for well-resourced tasks such as machine translation (MT), it is still computationally prohibitive as well as not environ.mentally friendly (Strubell et al., 2019) to build and test on systems for all languages or domains we are interested in. Because of this, the common practice is to test new methods on a small number of languages or domains, often semi-arbitrarily chosen based on previous work or the experimenters intuition. As a result, this practice impedes the NLP community from gaining a comprehensive under.standing of newly-proposed models. Table 1 il.lustrates this fact with an example from bilingual lexicon induction, a task that aims to find word translation pairs from cross-lingual word embed.dings. As vividly displayed in Table 1, almost all the works report evaluation results on a different subset of language pairs. Evaluating only on a small subset raises concerns about making inferences when comparing the merits of these methods: there is no guarantee that performance on English/Spanish (ENOES, the only common evaluation dataset) is representative of the expected performance of the models over all other language pairs (Anastasopoulos and Neubig, 2020). Such phenomena lead us to consider if it is possible to make a decently accurate estimation for the performance over an untested language pair without actually running the NLP model to bypass the computation restriction. Toward that end, through drawing on the idea of characterizing an experiment from Lin et al. (2019), we propose a framework, which we call NLPERF, to provide an exploratory solution. We build regression models, to predict the performance on a particular experimental setting given past experimental records of the same task, with each record consisting of a characterization of its training dataset and a performance score of the corresponding metric. Concretely, in 2, we start with a partly populated table (such as the one from <> Table 1: An illustration of the comparability issues across methods and multiple evaluation datasets from the Bilingual Lexicon Induction task. Our prediction model can reasonably ll in the blanks, as illustrated in Section 4. Table (1) and attempt to infer the missing values with the predictor. We begin by introducing the process of characterizing an NLP experiment for each task in 3. We evaluate the effectiveness and robustness of NLPERF by comparing to multiple baselines, human experts, and by perturbing a single feature to simulate a grid search over that feature (4). Evaluations on multiple tasks show that NLPERF is able to outperform all baselines. Notably, on a machine translation (MT) task, the predictions made by the predictor turn out to be more accurate than human experts. An effective predictor can be very useful for multiple applications associated with practical scenarios. In 5, we show how it is possible to adopt the predictor as a scoring function to find a small subset of experiments that are most representative of a bigger set of experiments. We argue that this will allow researchers to make informed decisions on what datasets to use for training and evaluation, in the case where they cannot experiment on all experimental settings. Last, in 6, we show that we can adequately predict the performance of new models even with a minimal number of experimental records. 2 Problem Formulation In this section we formalize the problem of predicting performance on supervised NLP tasks. Given an NLP model of architecture M trained over dataset(s) D of a specific task involving language(s) L with a training procedure (optimization algorithms, learning rate scheduling etc.) P, we can test the model on a test dataset D0 and get a score S of a specific evaluation metric. The resulting score will surely vary depending on all the above mentioned factors, and we denote this relation as g: <>. (1) In the ideal scenario, for each test dataset D0 of a specific task, one could enumerate all different settings and find the one that leads to the best performance. As mentioned in Section 1, however, such a brute-force method is computationally infeasible. Thus, we turn to modeling the process and formulating our problem as a regression task by using a parametric function f. to approximate the true function g as follows: <> where <> denotes a set of features for each influencing factor. For the purpose of this study, we mainly focus on dataset and language features .L and .D, as this already results in a significant search space, and gathering extensive experimental results with fine-grained tuning over model and training hyper-parameters is both expensive and relatively complicated. In the cases where we handle multiple models, we only use a single categorical model feature to denote the combination of model architecture and training procedure, denoted as .C. We still use the term model to refer to this combination in the rest of the paper. We also omit the test set features, under the assumption that the data distributions for training and testing data are the same (a fairly reasonable assumption if we ignore possible domain shift). Therefore, for all experiments below, our final prediction function is the following: <> In the next section we describe concrete instantiations of this function for several NLP tasks. 3 NLP Task Instantiations To build a predictor for NLP task performance, we must 1) select a task, 2) describe its featurization, and 3) train a predictor. We describe details of these three steps in this section. <> Table 2: Statistics of the datasets we use for training predictors. # EXs denote the total number of experiment instances; Task Metric reflects how the models are evaluated. Tasks We test on tasks including bilingual lexicon induction (BLI); machine translation trained on aligned Wikipedia data (Wiki-MT), on TED talks (TED-MT), and with cross-lingual trans.fer for translation into English (TSF-MT); cross-lingual dependency parsing (TSF-Parsing); cross-lingual POS tagging (TSF-POS); cross-lingual entity linking (TSF-EL); morphological analysis (MA) and universal dependency parsing (UD). Ba.sic statistics on the datasets are outlined in Table 2. For Wiki-MT tasks, we collect experimental records directly from the paper describing the cor.responding datasets (Schwenk et al., 2019). For TED-MT and all the transfer tasks, we use the results of Lin et al. (2019). For BLI, we conduct experiments using published results from three papers, namely Artetxe et al. (2016), Artetxe et al. (2017) and Xu et al. (2018). For MA, we use the results of the SIGMORPHON 2019 shared task 2 (McCarthy et al., 2019). Last, the UD results are taken from the CoNLL 2018 Shared Task on universal dependency parsing (Zeman et al., 2018b). Featurization For language features, we utilize six distance features from the URIEL Typologi.cal Database (Littell et al., 2017), namely geo.graphic, genetic, inventory, syntactic, phonological, and featural distance. The complete set of dataset features includes the following: 1. Dataset Size: The number of data entries used for training. 2. Word/Subword Vocabulary Size: The number of word/subword types. 3. Average Sentence Length: The average length of sentences from all experimental. 4. Word/Subword Overlap: <> where T1 and T2 denote vocabularies of any two corpora. 5. Type-Token Ratio (TTR): The ratio between the number of types and number of tokens (Richards, 1987) of one corpus. 6. Type-Token Ratio Distance: <> where TTR1 and TTR2 denote TTR of any two corpora. 7. Single Tag Type: Number of single tag types. 8. Fused Tag Type: Number of fused tag types. 9. Average Tag Length Per Word: Average number of single tags for each word. 10. Dependency Arcs Matching WALS features: the proportion of dependency parsing arcs matching the following WALS features, computed over the training set: subject/object/oblique before/after verb and adjective/numeral before/after noun. For transfer tasks, we use the same set of dataset features .D as Lin et al. (2019), including features 1x6 on the source and the transfer language side. We also include language distance features between source and transfer language, as well as between source and target language. For MT tasks, we use features 1x6 and language distance features, but only between the source and target language. For MA, we use features 1, 2, 5 and morphological tag related features 7x9. For UD, we use features 1, 2, 5, and 10. For BLI, we use language distance features and URIEL syntactic features for the source and the target language. Predictor Our prediction model is based on gradient boosting trees (Friedman, 2001), implemented with XGBoost (Chen and Guestrin, 2016). This method is widely known as an effective means for solving problems including ranking, classification and regression. We also experimented with Gaussian processes (Williams and Rasmussen, 1996), but settled on gradient boosted trees because performance was similar and Xg.boost's implementation is very efficient through the use of parallelism. We use squared error as the objective function for the regression and adopted a fixed learning rate 0.1. To allow the model to fully fit the data we set the maximum tree depth to be 10 and the number of trees to be 100, and use the default regularization terms to prevent the model from overfitting. 4 Can We Predict NLP Performance? In this section we investigate the effectiveness of NLPERF across different tasks on various metrics. Following Lin et al. (2019), we conduct k-fold cross validation for evaluation. To be specific, we randomly partition the experimental records of hL, D, C, Si tuples into k folds, and use k.1 folds to train a prediction model and evaluate on the remaining fold. Note that this scenario is similar to filling in the blanks in Table 1, where we have some experimental records that we can train the model on, and predict the remaining ones. For evaluation, we calculate the average root mean square error (RMSE) between the predicted scores and the true scores. Baselines We compare against a simple mean value baseline, as well as against language-wise mean value and model-wise mean value baselines. The simple mean value baseline outputs an aver.age of scores s from the training folds for all test entries in the left-out fold (i) as follows: <> But performance of what? (FLOPS, energy, memory) or plain accuracy? <> (2) Note that for tasks involving multiple models, we calculate the RMSE score separately on each model and use the mean RMSE of all models as the final RMSE score. The language-wise baselines make more in.formed predictions, taking into account only train.ing instances with the same transfer, source, or tar.get language (depending on the task setting). For example, the source-language mean value baseline <<(i,j)>> s for jth test instance in fold i outputs an average of the scores s of the training instances that share the same source language features s-lang, as shown in Equation 3: <> (3) where . is the indicator function. Similarly, we define the target-and the transfer-language mean value baselines. In a similar manner, we also compare against a model-wise mean value baseline for tasks that include experimental records from multiple models. Now, the prediction for the jth test instance in the left-out fold i is an average of the scores on the same dataset (as characterized by the language .L and dataset .D features) from all other models: <> (4) where <> and <> respectively denote the language and dataset features of the test instance. Main Results For multi-model tasks, we can do either Single Model prediction (SM), restricting training and testing of the predictor within a single model, or Multi-Model (MM) prediction using a categorical model feature. The RMSE scores of NLPERF along with the baselines are shown in Table 3. For all tasks, our single model predictor is able to more accurately estimate the evaluation score of unseen experiments compared to the single model baselines, confirming our hypothesis that the there exists a correlation that can be captured between experimental settings and the downstream performance of NLP systems. The language-wise baselines are much stronger than the simple mean value baseline but still perform worse than our single model predictor. Similarly, the model-wise baseline significantly outperforms the mean value baseline because results from other models reveal much information about the dataset. <> Table 3: RMSE scores of three baselines and our predictions under the single model and multi model setting (missing values correspond to settings not applicable to the task). All results are from k-fold (k =5) evaluations averaged over 10 random runs. Even so, our multi-model predictor still outperforms the model-wise baseline. The results nicely imply that for a wide range of tasks, our predictor is able to reasonably estimate left-out slots in a partly populated table given results of other experiment records, without actually running the system. We should note that RMSE scores across different tasks should not be directly compared, mainly because the scale of each evaluation metric is different. For example, a BLEU score (Papineni et al., 2002) for MT experiments typically ranges from 1 to 40, while an accuracy score usually has a much larger range, for example, BLI accuracy ranges from 0.333 to 78.2 and TSF-POS accuracy ranges from 1.84 to 87.98, which consequently makes the RMSE scores of these tasks higher. Comparison to Expert Human Performance We constructed a small scale case study to evaluate whether NLPERF is competitive to the performance of NLP sub-field experts. We focused on the TED-MT task and recruited 10 MT practitioners, all of whom had published at least 3 MT-related papers in ACL-related conferences. In the first set of questions, the participants were presented with language pairs from one of the k data folds along with the dataset features and were asked to estimate an eventual BLEU score for each data entry. In the second part of the questionnaire, the participants were tasked with making estimations on the same set of language pairs, but this time they also had access to features, and BLEU scores from all the other folds.3 <> Table 4: Our model performs better than human MT experts on the TED-MT prediction task. The partition of the folds is consistent between the human study and the training/evaluation for the predictor. While the first sheet is intended to familiarize the participants with the task, the second sheet fairly adopts the training/evaluation setting for our predictor. As shown in Table 4, our participants outperform the mean baseline even without information from other folds, demonstrating their own strong prior knowledge in the field. In addition, the participants make more accurate guesses after acquiring more information on experimental records in other folds. In neither case, though, are the human experts competitive to our predictor. In fact, only one of the participants achieved performance comparable to our predictor. Feature Perturbation Another question of interest concerning predicting performance is how will the model perform when trained on data of a different size (Kolachina et al., 2012a). To test NLPERF's extrapolation ability in this regard, we conduct an array of experiments on one language pair with various data sizes on the Wiki-MT task. We pick two language pairs, Turkish to English (TR�EN) and Portuguese to English (PT�EN) as 2 None of the study participants were affiliated to the au-our testbed for the Wiki-MT task. We sample parallel datasets with different sizes and train MT models with each sampled dataset to obtain the true BLEU scores. On the other hand, we collect the features of all sampled datasets and use our predictor (trained over all other languages pairs) to obtain predictions. The plot of true BLEU scores and predicted BLEU scores are shown in Figure 1. Our predictor achieves a very low average RMSE of 1.83 for TR�EN pair but a relatively higher RMSE of 9.97 for PT�EN pair. The favorable performance on the tr-en pair demonstrates the possibility of our predictor to do feature extrapolation over data set size. In contrast, the predictions on the pt-en pair are significantly less accurate. This is due to the fact that there are only two other experimental settings scoring as high as 34 BLEU score, with data sizes of 3378k (en-es) and 611k (gl-es), leading to the predictor�s inadequacy in predicting high BLEU scores for low-resourced data sets during extrapolation. This reveals the fact that while the predictor is able to extrapolate performance on settings similar to what it has seen in the data, NLPERF may be less successful under circumstances unlike its training inputs. 3 The interested reader can find an example questionnaire (and make estimations over one of the folds) in the Authors institutions, nor were familiar with this paper's content. <

> Figure 1: Our model's predicted BLEU scores and true BLEU scores, on sampled TR�EN datasets (sizes 10k/50k/100k/200k/478k) and PT�EN datasets (sizes 100k/500k/1000k/2000k/2462k), achieving a RMSE score of 1.83 and 9.97 respectively. 5 What Datasets Should We Test On? As shown in Table 1, it is common practice to test models on a subset of all available datasets. The reason for this is practical <> it is computationally prohibitive to evaluate on all settings. However, if we pick test sets that are not representative of the data as a whole, we may mistakenly reach un.founded conclusions about how well models per.form on other data with distinct properties. For example, models trained on a small-sized dataset may not scale well to a large-sized one, or models that perform well on languages with a particular linguistic characteristic may not do well on languages with other characteristics (Bender and Friedman, 2018). Here we ask the following question: if we are only practically able to test on a small number of experimental settings, which ones should we test on to achieve maximally representative results? Answering the question could have practical im.plications: organizers of large shared tasks like SIGMORPHON (McCarthy et al., 2019) or UD (Zeman et al., 2018a) could create a minimal sub.set of settings upon which they would ask participants to test to get representative results; similarly, participants could possibly expedite the iteration of model development by testing on the representative subset only. A similar avenue for researchers and companies deploying systems over multiple languages could lead to not only financial savings, but potentially a significant cut-down of emissions from model training (Strubell et al., 2019). We present an approximate explorative solution to the problem mentioned above. Formally, assume that we have a set N , comprising experimental records (both features and scores) of n datasets for one task. We set a number m (>. (5) Naturally, enumerating all possible subsets would be prohibitively costly, even though it would lead to the optimal solution. Instead, we employ a beam-search-like approach to efficiently search for an approximate solution to the best per.forming subset of arbitrary size. Concretely, we start our approximate search with an exhaustive enumeration of all subsets of size 2. At each fol.lowing step t, we only consider the best k subsets <> into account and discard the t rest. As shown in Equation 6, for each candidate <

> Figure 2: Beam search results (beam size=100) for up to the 5 most (and least) representative datasets for 4 NLP tasks. We also show random search results averaged over 100 random runs. subset, we expand it with one more data point, <>. (6) For tasks that involve multiple models, we take experimental records of the selected dataset from all models into account during expansion. Given all expanded subsets, we train a predictor for each to evaluate on the rest of the data sets, and keep the (i) best performing k subsets <> with minimum RMSE scores for the next step. Furthermore, note that by simply changing the arg min to an arg max in Equation 5, we can also find the least representative datasets. We present search results for four tasks4 as beam search progresses in Figure 2, with cor.responding RMSE scores from all remaining datasets as the y-axis. For comparison, we also conduct random searches by expanding the subset with a randomly selected experimental record. In all cases, the most representative sets are an aggregation of datasets with diverse characteristics such as languages and dataset sizes. For example, in the Wiki-MT task, the 5 most representative datasets include languages that fall into a diverse range of language families such as Romance, Turkic, Slavic, etc. while the least representative ones include duplicate pairs (opposite directions) mostly involving English. The phenomenon is more pronounced in the TED-MT task, where not only the 5 most representative source languages are di.verse, but also the dataset sizes. specifically, the Malay-English (msa-eng) is a tiny dataset (5k parallel sentences), and Hebrew-English (heb-eng) is a high-resource case (212k parallel sentences). Notably, for BLI task, to test how representative the commonly used datasets are, we se.lect the most frequent 5 language pairs shown in Table 1, namely en-de, es-en, en-es, fr-en, en-fr for evaluation. Unsurprisingly, we get an RMSE score as high as 43.44, quite close to the performance of the worst representative set found using beam search. This finding indicates that the standard practice of choosing datasets for evaluation is likely unrepresentative of results over the full dataset spectrum, well aligned with the claims in Anastasopoulos and Neubig (2020). A particularly encouraging observation is that the predictor trained with only the 5 most representative datasets can achieve an RMSE score comparable to k-fold validation, which required using all of the datasets for training.5 This indicates that one would only need to train NLP models on a small set of representative datasets to obtain reasonably plausible predictions for the rest. 5 to be accurate, k . 1 folds of all datasets. 6 Can We Extrapolate Performance for those better-performing systems, so the predictor New Models? is unable to generalize well (ONLP). In another common scenario, researchers propose new models for an existing task. It is both time-consuming and computationally intensive to run experiments with all settings for a new model. In this section, we explore if we can use past experimental records from other models and a minimal set of experiments from the new model to give a plausible prediction over the rest of the datasets, potentially reducing the time and resources needed for experimenting with the new model to a large extent. We use the task of UD parsing as our testbed6 as it is the task with most unique models (25 to be exact). Note that we still only use a single categorical feature for the model type. To investigate how many experiments are needed to have a plausible prediction for a new model, we first split the experimental records equally into a sample set and a test set. Then we randomly sample <> experimental records from the sample set and add them into the collection of experiment records of past models. Each time we re-train a predictor and evaluate on the test set. The random split repeats 50 times and the random sampling repeats 50 times, adding up to a total of 2500 experiments. We use the mean value of the results from other models, shown in Equation 7 as the prediction baseline for the left-out model, and because experiment results of other models reveal significant information about the dataset, this serves as a relatively strong baseline: <>. (7) M denotes a collection of models and k denotes the left-out model. We show the prediction performance (in RMSE) over 8 systems7 in Figure 3. Interestingly, the predictor trained with no model records (0) outperforms the mean value baseline for the 4 best systems, while it is the opposite case on the 4 worst systems. Since there is no information provided about the new-coming model, the predictions are solely based on dataset and language features. One reason might explain the phenomenon .the correlation between the features and the scores of the worse-performing systems is different from 6MA and BLI task results are in Appendix C 7The best and worst 4 systems from the shared task. In the following discussion, we use RMSE@n to denote the RMSE from the predictor trained with n data points of a new model. The relatively low RMSE@0 scores indicate that other models' features and scores are informative for predicting the performance of the new model even without new model information. Comparing RMSE@0 and RMSE@1, we observe a consistent improvement for almost all systems, indicating that NLPERF trained on even a single ex.tra random example achieves more accurate estimates over the test sets. Adding more data points consistently leads to additional gains. However, predictions on worse-performing systems benefit more from it than for better-performing systems, indicating that their feature-performance correlation might be considerably different. The findings here indicate that by extrapolating from past experiments, one can make plausible judgments for newly developed models. 7 Related Work As discussed in Domhan et al. (2015), there are two main threads of work focusing on predict.ing performance of machine learning algorithms. The first thread is to predict the performance of a method as a function of its training time, while the second thread is to predict a method's performance as a function of the training dataset size. Our work belongs in the second thread, but could easily be extended to encompass training time/procedure. In the first thread, Kolachina et al. (2012b) at.tempt to infer learning curves based on training data features and extrapolate the initial learning curves based on BLEU measurements for statistical machine translation (SMT). By extrapolating the performance of initial learning curves, the predictions on the remainder allows for early termination of a bad run (Domhan et al., 2015). In the second thread, Birch et al. (2008) adopt linear regression to capture the relationship between data features and SMT performance and find that the amount of reordering, the morphological complexity of the target language and the relatedness of the two languages explains the majority of performance variability. More recently, Elsa. har and Gall� (2019) use domain shift metrics such as H-divergence based metrics to predict drop in performance under domain-shift. Rosenfeld et al. <

> Figure 3: RMSE scores of UD task from dataset-wise mean value predictor (the dashed black line in each graph) and predictors trained with experimental records of other models and <> records from a new model. (2020) explore the functional form of the dependency of the generalization error of neural models on model and data size. We view our work as a generalization of such approaches, appropriate for application on any NLP task. 8 Conclusion and Future Work In this work, we investigate whether the experiment setting itself is informative for predicting the evaluation scores of NLP tasks. Our findings promisingly show that given a sufficient number of past training experimental records, our predictor can 1) outperform human experts; 2) make plau.sible predictions even over new-coming models and languages; 3) extrapolate well on features like dataset size; 4) provide a guide on how we should choose representative datasets for fast iteration. While this discovery is a promising start, there are still several avenues on improvement in future work. First, the dataset and language settings covered in our study are still limited. Experimental records we use are from relatively homogeneous settings, e.g. all datasets in Wiki-MT task are sentence-pieced to have 5000 subwords, indicating that our predictor may fail for other subword settings. Our model also failed to generalize to cases where feature values are out of the range of the training experimental records. We attempted to apply the pre.dictor of Wiki-MT to evaluate on a low-resource MT dataset, translating from Mapudungun (arn) to Spanish (spa) with the dataset from Duan et al. (2019), but ended up with a poor RMSE score. It turned out that the average sentence length of the arn/spa data set is much lower than that of the training data sets and our predictors fail to generalize to this different setting. Second, using a categorical feature to denote model types constrains its expressive power for modeling performance. In reality, a slight change in model hyperparameters (Hoos and Leyton-Brown, 2014; Probst et al., 2019), optimization algorithms (Kingma and Ba, 2014), or even random seeds (Madhyastha and Jain, 2019) may give rise to a significant variation in performance, which our predictor is not able to capture. While investigating the systematic implications of model structures or hyperparameters is practically infeasible in this study, we may use additional information such as textual model descriptions for modeling NLP models and training procedures more elaborately in the future. Lastly, we assume that the distribution of train.ing and testing data is the same, which does not consider domain shift. On top of this, there might also be a domain shift between data sets of train.ing and testing experimental records. We believe that modeling domain shift is a promising future direction to improve performance prediction. Acknowledgement The authors sincerely thank all the reviewers for their insightful comments and suggestions, Philipp Koehn, Kevin Duh, Matt Post, Shuoyang Ding, Xuan Zhang, Adi Renduchintala, Paul Mc-Namee, Toan Nguyen and Kenton Murray for con.ducting human evaluation for the TED-MT task, Daniel Beck for discussions on Gaussian Pro.cesses, Shruti Rijhwani, Xinyi Wang, Paul Michel for discussions on this paper. This work is generously supported from the National Science Foundation under grant 1761548. References Antonios Anastasopoulos and Graham Neubig. 2020. Should all cross-lingual embeddings speak english? In Proc. ACL. To appear. Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2016. Learning principled bilingual mappings of word em.beddings while preserving monolingual invariance. In Proceedings of the 2016 Conference on Empiri.cal Methods in Natural Language Processing, pages 2289�2294. Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2017. Learning bilingual word embeddings with (almost) no bilingual data. In Proceedings of the 55th Annual Meeting of the Association for Computational Lin.guistics (Volume 1: Long Papers), pages 451�462, Vancouver, Canada. Association for Computational Linguistics. Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2019. Bilingual lexicon induction through unsupervised machine translation. In Proceedings of the 57th An.nual Meeting of the Association for Computational Linguistics, pages 5002�5007, Florence, Italy. Asso.ciation for Computational Linguistics. Emily M Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587�604. Alexandra Birch, Miles Osborne, and Philipp Koehn. 2008. Predicting success in machine translation. In Proceedings of the Conference on Empirical methods in Natural Language Processing, pages 745� 754. Association for Computational Linguistics. Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785� 794. ACM. Xilun Chen and Claire Cardie. 2018. Unsupervised multilingual word embeddings. In Proceedings of the 2018 Conference on Empirical Methods in Nat.ural Language Processing, pages 261�270, Brus.sels, Belgium. Association for Computational Lin.guistics. Tobias Domhan, Jost Tobias Springenberg, and Frank Hutter. 2015. Speeding up automatic hyperparame.ter optimization of deep neural networks by extrap.olation of learning curves. In Twenty-Fourth Inter.national Joint Conference on Arti�cial Intelligence. Mingjun Duan, Carlos Fasola, Sai Krishna Rallabandi, Rodolfo M. Vega, Antonios Anastasopoulos, Lori Levin, and Alan W Black. 2019. A resource for computational experiments on mapudungun. In Proc. LREC. To appear. Hady Elsahar and Matthias Gall�. 2019. To annotate or not? predicting performance drop under domain shift. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu.ral Language Processing (EMNLP-IJCNLP), pages 2163�2173. Jerome H Friedman. 2001. Greedy function approx.imation: a gradient boosting machine. Annals of statistics, pages 1189�1232. Geert Heyman, Bregt Verreet, Ivan Vuli�c, and Marie-Francine Moens. 2019. Learning unsupervised mul.tilingual word embeddings with incremental multi.lingual hubs. In Proceedings of the 2019 Confer.ence of the North American Chapter of the Asso.ciation for Computational Linguistics: Human language Technologies, Volume 1 (Long and Short papers), pages 1890�1902. Holger Hoos and Kevin Leyton-Brown. 2014. An ef.�cient approach for assessing hyperparameter im.portance. In International conference on machine learning, pages 754�762. Jiaji Huang, Qiang Qiu, and Kenneth Church. 2019. Hubless nearest neighbor search for bilingual lexi. con induction. In Proceedings of the 57th Annual Meeting of the Association for Computational Lin.guistics, pages 4072�4080, Florence, Italy. Associa.tion for Computational Linguistics. Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. Prasanth Kolachina, Nicola Cancedda, Marc Dymet.man, and Sriram Venkatapathy. 2012a. Prediction of learning curves in machine translation. In Proceed.ings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long papers), pages 22�30, Jeju Island, Korea. Association for Computational Linguistics. Prasanth Kolachina, Nicola Cancedda, Marc Dymet.man, and Sriram Venkatapathy. 2012b. Prediction of learning curves in machine translation. In Proceed.ings of the 50th Annual Meeting of the Association for Computational Linguistics: Long Papers-Volume 1, pages 22�30. Association for Computational Lin.guistics. Guillaume Lample, Alexis Conneau, Marc�Aurelio Ranzato, Ludovic Denoyer, and Herv� J�gou. 2018. Word translation without parallel data. In Interna.tional Conference on Learning Representations. Yu-Hsiang Lin, Chian-Yu Chen, Jean Lee, Zirui Li, Yuyan Zhang, Mengzhou Xia, Shruti Rijhwani, Junxian He, Zhisong Zhang, Xuezhe Ma, Antonios Anastasopoulos, Patrick Littell, and Graham Neu.big. 2019. Choosing transfer languages for cross-lingual learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Lin.guistics, pages 3125�3135, Florence, Italy. Associa.tion for Computational Linguistics. Patrick Littell, David R Mortensen, Ke Lin, Katherine Kairis, Carlisle Turner, and Lori Levin. 2017. Uriel and lang2vec: Representing languages as typologi.cal, geographical, and phylogenetic vectors. In Pro.ceedings of the 15th Conference of the European Chapter of the Association for Computational Lin.guistics: Volume 2, Short Papers, pages 8�14. Pranava Madhyastha and Rishabh Jain. 2019. On model stability as a function of random seed. In Pro.ceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 929� 939, Hong Kong, China. Association for Computa.tional Linguistics. Arya D. McCarthy, Ekaterina Vylomova, Shijie Wu, Chaitanya Malaviya, Lawrence Wolf-Sonkin, Gar.rett Nicolai, Christo Kirov, Miikka Silfverberg, Se.bastian J. Mielke, Jeffrey Heinz, Ryan Cotterell, and Mans Hulden. 2019. The SIGMORPHON 2019 shared task: Morphological analysis in context and cross-lingual transfer for in�ection. In Proceedings of the 16th Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 229� 244, Florence, Italy. Association for Computational Linguistics. Joakim Nivre, Mitchell Abrams, .eljko Agic,� Lars Ahrenberg, Lene Antonsen, Maria Jesus Aranzabe, Gashaw Arutie, Masayuki Asahara, Luma Ateyah, Mohammed Attia, Aitziber Atutxa, Liesbeth Augustinus, Elena Badmaeva, Miguel Balles.teros, Esha Banerjee, Sebastian Bank, Verginica Barbu Mititelu, John Bauer, Sandra Bellato, Kepa Bengoetxea, Riyaz Ahmad Bhat, Erica Biagetti, Eckhard Bick, Rogier Blokland, Victoria Bobicev, Carl Bstell, Cristina Bosco, Gosse Bouma, Sam Bowman, Adriane Boyd, Aljoscha Burchardt, Marie Candito, Bernard Caron, Gauthier Caron, G�sen Cebiro� git, Giuseppe G. A. Celano, Savas glu Eryi� Cetin, Fabricio Chalub, Jinho Choi, Yongseok Cho, Jayeol Chun, Silvie Cinkov�, Aur�lie Collomb, �a� gr� fitekin, Miriam Connor, Marine Courtin, Elizabeth Davidson, Marie-Catherine de Marn.effe, Valeria de Paiva, Arantza Diaz de Ilarraza, Carly Dickerson, Peter Dirix, Kaja Dobrovoljc, Timothy Dozat, Kira Droganova, Puneet Dwivedi, Marhaba Eli, Ali Elkahky, Binyam Ephrem, Toma. Erjavec, Aline Etienne, Rich�rd Farkas, Hector Fernandez Alcalde, Jennifer Foster, Cl�udia Freitas, Katar�na Gajdo.ov�, Daniel Galbraith, Marcos Garcia, Moa G�rdenfors, Kim Gerdes, Filip Gin.ter, Iakes Goenaga, Koldo Gojenola, Memduh G�rmak, Yoav Goldberg, Xavier Gez Guino.vart, Berta Gonz�les Saavedra, Matias Grioni, Normunds Gruz� �fitis, Bruno Guillaume, C�line Guillot-Barbance, Nizar Habash, Jan Hajic,� Jan Haji�c jr., Linh H� M�y, Na-Rae Han, Kim Harris, Dag Haug, Barbora Hladk�, Jaroslava Hlav�� cov�, Florinel Hociung, Petter Hohle, Jena Hwang, Radu Ion, Elena Irimia, Tom�. Jelfinek, Anders Johannsen, Fredrik Jgensen, Her Ka�s�kara, Sylvain Kahane, Hiroshi Kanayama, Jenna Kanerva, Tolga Kayade.len, V�clava Kettnerov�, Jesse Kirchner, Natalia Kotsyba, Simon Krek, Sookyoung Kwak, Veronika Laippala, Lorenzo Lambertino, Tatiana Lando, Septina Dian Larasati, Alexei Lavrentiev, John Lee, Phng L� H` g, Alessandro Lenci, Saran Lertpradit, Herman Leung, Cheuk Ying Li, Josie Li, Keying Li, KyungTae Lim, Nikola Ljube.ic,� Olga Loginova, Olga Lyashevskaya, Teresa Lynn, Vivien Macketanz, Aibek Makazhanov, Michael Mandl, Christopher Manning, Ruli Manurung, C� alina at� Mar� � cek, Katrin Marheinecke, anduc, David Mare� H�ctor Martfinez Alonso, Andr� Martins, Jan Ma.ek, Yuji Matsumoto, Ryan McDonald, Gustavo Mendon�a, Niko Miekka, Anna Missil�, Cat� � alin Mititelu, Yusuke Miyao, Simonetta Montemagni, Amir More, Laura Moreno Romero, Shinsuke Mori, Bjartur Mortensen, Bohdan Moskalevskyi, Kadri Muischnek, Yugo Murawaki, Kaili Mrisep, Pinkey Nainwani, Juan Ignacio Navarro Horacek, Anna Nedoluzhko, Ne.pore-B� Gunta erzkalne, ., . Lng Nguy�n� Thi Huy�n` Nguy�n� Thi Minh, Vitaly Nikolaev, Rattima Nitisaroj, Hanna Nurmi, Stina Ojala, Ad�day� Olkun, Mai Omura, . Petya Osenova, Robert �stling, Lilja �vrelid, Niko Partanen, Elena Pascual, Marco Passarotti, Agnieszka Patejuk, Siyao Peng, Cenel-Augusto Perez, Guy Perrier, Slav Petrov, Jussi Piitulainen, Emily Pitler, Barbara Plank, Thierry Poibeau, Martin Popel, Lauma Pretkalnin, a, Sophie Pr�vost, Prokopis Prokopidis, Adam Przepikowski, Tiina Puolakainen, Sampo Pyysalo, Andriela R��bis, Alexandre Rademaker, Loganathan Ramasamy, Taraka Rama, Carlos Ramisch, Vinit Ravishankar, Livy Real, Siva Reddy, Georg Rehm, Michael Rie�ler, Larissa Rinaldi, Laura Rituma, Luisa Rocha, Mykhailo Romanenko, Rudolf Rosa, Da.vide Rovati, Valentin Ros,ca, Olga Rudina, Shoval Sadde, Shadi Saleh, Tanja Samard.ic,� Stephanie Samson, Manuela Sanguinetti, Baiba Saul�fite, Yanin Sawanakunanon, Nathan Schneider, Sebas.tian Schuster, Djam� Seddah, Wolfgang Seeker, Mojgan Seraji, Mo Shen, Atsuko Shimada, Muh Shohibussirri, Dmitry Sichinava, Natalia Silveira, Maria Simi, Radu Simionescu, Katalin Simk M�ria .imkov�, Kiril Simov, Aaron Smith, Isabela Soares-Bastos, Antonio Stella, Milan Straka, Jana Strnadov�, Alane Suhr, Umut Sulubacak, Zsolt Sz�nt Dima Taji, Yuta Takahashi, Takaaki Tanaka, Isabelle Tellier, Trond Trosterud, Anna Trukhina, Reut Tsarfaty, Francis Tyers, Sumire Uematsu, Zde� nka Ure.ov�, Larraitz Uria, Hans Uszkoreit, Sowmya Vajjala, Daniel van Niekerk, Gertjan van Noord, Viktor Varga, Veronika Vincze, Lars Wallin, Jonathan North Washington, Seyi Williams, Mats Wir�n, Tsegay Woldemariam, Tak-sum Wong, Chunxiao Yan, Marat M. Yavrumyan, Zhuoran Yu, Zdenek� .abokrtsk Amir Zeldes, Daniel Zeman, Manying Zhang, and Hanzhi Zhu. 2018. Universal dependencies 2.2. LINDAT/CLARIN digital library at the Institute of Formal and Applied Linguistics (�FAL), Faculty of Mathematics and Physics, Charles University. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic eval.uation of machine translation. In Proceedings of the 40th annual meeting on association for compu.tational linguistics, pages 311�318. Association for Computational Linguistics. Philipp Probst, Anne-Laure Boulesteix, and Bernd Bis.chl. 2019. Tunability: Importance of hyperparame.ters of machine learning algorithms. Journal of Ma.chine Learning Research, 20(53):1�32. Ye Qi, Devendra Sachan, Matthieu Felix, Sarguna Pad.manabhan, and Graham Neubig. 2018. When and why are pre-trained word embeddings useful for neural machine translation? In Meeting of the North American Chapter of the Association for Computa.tional Linguistics (NAACL), New Orleans, USA. Brian Richards. 1987. Type/token ratios: What do they really tell us? Journal of child language, 14(2):201� 209. Shruti Rijhwani, Jiateng Xie, Graham Neubig, and Jaime Carbonell. 2019. Zero-shot neural transfer for cross-lingual entity linking. In Thirty-Third AAAI Conference on Arti�cial Intelligence (AAAI), Hon.olulu, Hawaii. Jonathan S. Rosenfeld, Amir Rosenfeld, Yonatan Be.linkov, and Nir Shavit. 2020. A constructive pre. diction of the generalization error across scales. In International Conference on Learning Representa.tions. Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong, and Francisco Guzm�n. 2019. Wiki-matrix: Mining 135m parallel sentences in 1620 language pairs from wikipedia. arXiv preprint arXiv:1907.05791. Emma Strubell, Ananya Ganesh, and Andrew McCal.lum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computa.tional Linguistics, pages 3645�3650, Florence, Italy. Association for Computational Linguistics. Christopher KI Williams and Carl Edward Rasmussen. 1996. Gaussian processes for regression. In Ad.vances in neural information processing systems, pages 514�520. Ruochen Xu, Yiming Yang, Naoki Otani, and Yuexin Wu. 2018. Unsupervised cross-lingual transfer of word embedding spaces. In Proceedings of the 2018 Conference on Empirical Methods in Natural language Processing, pages 2465�2474. Pengcheng Yang, Fuli Luo, Peng Chen, Tianyu Liu, and Xu Sun. 2019. Maam: A morphology-aware alignment model for unsupervised bilingual lexicon induction. In Proceedings of the 57th Annual Meet.ing of the Association for Computational Linguis.tics, pages 3190�3196. Daniel Zeman, Jan Haji�c, Martin Popel, Martin Pot.thast, Milan Straka, Filip Ginter, Joakim Nivre, and Slav Petrov. 2018a. CoNLL 2018 shared task: Mul. tilingual parsing from raw text to universal depen. dencies. In Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Univer.sal Dependencies, pages 1�21, Brussels, Belgium. Association for Computational Linguistics. Daniel Zeman, Jan Hajic,� Martin Popel, Martin Pot.thast, Milan Straka, Filip Ginter, Joakim Nivre, and Slav Petrov. 2018b. Conll 2018 shared task: mul.tilingual parsing from raw text to universal depen.dencies. In Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Univer.sal Dependencies, pages 1�21. Meng Zhang, Yang Liu, Huanbo Luan, and Maosong Sun. 2017. Earth mover�s distance minimization for unsupervised bilingual lexicon induction. In Pro.ceedings of the 2017 Conference on Empirical methods in Natural Language Processing, pages 1934� 1945. Appendix A Questionnaire An example of the first questionnaire from our user case study is shown below. The second sheet also included the results in 44 more language pairs. We provide an answer key after the second sheet. Please provide your prediction of the BLEU score based on the language pair and dataset features (the domain of the training and test sets is TED talks). After you �nish, please go to sheet v2. <

> Please provide your prediction of the BLEU score in the yellow area given all the information in this sheet. Note that all experiments are trained with the same model. <> B Representative datasets In this section, we show the searching results of most/least representative subsets for the rest of the <

> Figure 4: Beam search results (beam size=100) for up to the 5 most (and least) representative datasets for the remaining NLP tasks. We also show random search results of corresponding sizes. C New Model In this section, we show the extrapolation performance for new models on BLI, MA and the remaining systems of UD. <

> Figure 5: RMSE scores of BLI task from dataset-wise mean value predictor (the dashed black line in each graph) and predictors trained with experimental records of other models and 0�5 records from a new model (as indicated by the title of each graph). <

> Figure 6: RMSE scores of MA task from dataset-wise mean value predictor (the dashed black line in each graph) and predictors trained with experimental records of other models and 0�5 records from a new model (as indicated by the title of each graph) <

> Figure 7: RMSE scores of UD task from dataset-wise mean value predictor (the dashed black line in each graph) and predictors trained with experimental records of other models and 0�5 records from a new model (as indicated by the title of each graph). D Feature importance In this section, we show the plots of feature importance for all the tasks. <> <> <> <> <> <> Predicting trends in the quality of state-of-the-art neural networks without access to training or testing data Charles H. Martin Tongsu (Serena) Peng y Michael W. Mahoney z Abstract In many practical applications, one works with deep neural network (DNN) models trained by someone else. For such pretrained models, one typically does not have access to training data or test data. Moreover, one does not know many details about the model, such as the specifics of the training data, the loss function, the hyperparameter values, etc. Given one or many pretrained models, can one say anything about the expected performance or quality of the models? Here, we present and evaluate empirical quality metrics for pretrained DNN models at scale. Using the open-source Weight Watcher tool, we analyze hundreds of publicly-available pretrained models, including older and current state-of-the-art models in computer vision (CV) and natural language processing (NLP). We examine both familiar norm-based capacity control metrics (Frobenius and Spectral norms) as well as newer Power Law (PL) based metrics (including fitted PL exponents, <>, and the Weighted Alpha metric, <>, from the recently-developed Theory of Heavy-Tailed Self Regularization (HT-SR). We also introduce the -Shatten Norm metric. We find that norm-based metrics correlate well with reported test accuracies for well-trained models across nearly all CV architecture series. On the other hand, we find that norm-based metrics can not distinguish \good-versus-bad" models|which, arguably is the point of needing quality metrics. Indeed, they may give spurious results. We also find that PL-based metrics do much better|quantitatively better at discriminating among a series of \good-better-best" models, and qualitatively better at discriminating \good-versus-bad" models. PL-based metrics can also be used to characterize fine-scale properties of these models, and we introduce the layer-wiseCorrelation Flow as new quality assessment. We show how poorly-trained (and/or poorly fine-tuned) models may exhibit bothScale Collapse and unusually large PL exponents,6, in particular for recent NLP models. Our techniques, as implemented in the Weight Watcher tool, can be used to identify when a pretrained DNN has problems that can not be detected simply by examining training/test accuracies. 1 Introduction A common problem in machine learning (ML) is to evaluate the quality of a given model. A popular way to accomplish this is to train a model and then evaluate its training/testing error. There are many problems with this approach. The training/testing curves give very limited insight into the overall properties of the model; they do not take into account the (often large human and CPU/GPU) time for hyperparameter fiddling; they typically do not correlate with other properties of interest such as robustness or fairness or interpretability; and so on. A less well- known problem, but one that is increasingly important, in particular in industrial-scale artificial intelligence (AI), arises when the model user is not the model developer. Here, one may not have access to either the training data or the testing data. Instead, one may simply be given a model that has already been trained a pretrained model|and need to use it as-is, or to fine-tune and/or compress it and then use it. Natively|but in our experience commonly, among ML practitioners and ML theorists|if one does not have access to training or testing data, then one can say absolutely nothing about the quality of a ML model. This may be true in worst-case theory, but models are used in practice, and there is a need for a practical theory to guide that practice. Moreover, if ML is to become an industrial process, then that process will become siloed: some groups will gather data, other groups will develop models, and other groups will use those models. Users of models can not be expected to know the precise details of how models were built, the specifics of data that were used to train the model, what was the loss function or hyperparameter values, how precisely the model was regularized, etc. Moreover, for many large scale, practical applications, there is no obvious way to define an ideal test metric. For example, models that generate fake text or conversational chatbots may use a proxy, like perplexity, as a test metric. In the end, however, they really require human evaluation. Alternatively, models that cluster user profiles, which are widely used in areas such as marketing and advertising, are unsupervised and have no obvious labels for comparison and/or evaluation. In these and other areas, ML objectives can be poor proxies for downstream goals. Most importantly, in industry, one faces unique practical problems such as: do we have enough data for this model? Indeed, high quality, labeled data can be very expensive to acquire, and this cost can make or break a project. Methods that are developed and evaluated on any well-defined publicly-available corpus of data, no matter how large or diverse or interesting, are clearly not going to be well-suited to address problems such as this. It is of great practical interest to have metrics to evaluate the quality of a trained model|in the absence of training/testing data and without any detailed knowledge of the training/testing process. We seek a practical theory for pretrained models which can predict how, when, and why such models can be expected to perform well or poorly. In this paper, we present and evaluate quality metrics for pretrained deep neural network (DNN) models, and we do so at scale. We consider a large suite of hundreds of publicly-available models, mostly from computer vision (CV) and natural language processing (NLP). By now, there are many such state-of-the-art models that are publicly-available, e.g., there are now hundreds of pretrained models in CV (500) and NLP (100). 1 These provide a large corpus of models that by some community standard are state-of-the-art. 2 Importantly, all of these models have been trained by someone else and have been viewed to be of sufficient interest/quality to be made publicly-available; and, for all of these models, we have no access to training data or testing data, and we have no knowledge of the training/testing protocols. The quality metrics we consider are based on the spectral properties of the layer weight matrices. They are based on norms of weight matrices (such norms have been used in traditional statistical learning theory to bound capacity and construct regularizers) and/or parameters of power law (PL) fits of the eigenvalues of weight matrices (such PL fits are based on statistical mechanics approaches to DNNs). Note that, while we use traditional norm-based and PL-based metrics, our goals are not the traditional goals. Unlike more common ML approaches,we do not seek a bound on the generalization(e.g., by evaluating training/test error during training), 1 When we began this work in 2018, there were fewer than tens of such models; now in 2020, there are hundreds of such models; and we expect that in a year or two there will be an order of magnitude or more of such models. 2 Clearly, there is a selection bias or survivorship bias here|people tend not to make publicly-available their poorly-performing models|but these models are things in the world that (like social networks or the internet) can be analyzed for their properties. we do not seek a new regularizer, and we do not aim to evaluate a single model(e.g., as with hyperparameter optimization). 3 Instead, we want to examine different models across common architecture series, and we want to compare models between different architectures themselves, and in both cases, we ask: Can we predict trends in the quality of pretrained DNN models without access to training or testing data? To answer this question, we analyze hundreds of publicly-available pretrained state-of-the-art CV and NLP models. Here is a summary of our main results. Norm-based metrics and well-trained models.Norm-based metrics do a reasonably good job at predicting quality trends in well-trained CV/NLP models. Norm-based metrics and poorly-trained models.Norm-based metrics may give spurious results when applied to poorly-trained models (e.g., models trained without enough data, etc.), exhibitingScale Collapse for these models. PL-based metrics and model quality.PL-based metrics do much better at predicting quality trends in pretrained CV/NLP models. They are quantitatively better at discriminating good-better-best trends, and qualitatively better at distinguishing \good-versus- bad" models. PL-based metrics and model diagnostics. PL-based metrics can also be used to characterize fine-scale model properties (including layer-wiseCorrelation Flow) in well- trained and poorly-trained models, and they can be used to evaluate model enhancements (e.g., distillation, fine-tuning, etc.). We emphasize that our goal is a practical theory to predict trends in the quality of state-of-the- art DNN models, i.e., not to make a statement about every publicly-available model. We have examined hundreds of models, and we identify general trends, but we also highlight interesting exceptions. The WeightWatcher Tool. All of our computations were performed with the publicly-available Weight Watcher tool (version 0.2.7) [1]. To be fully reproducible, we only examine publicly- available, pretrained models, and we also provide all Jupyter and Google Colab notebooks used in an accompanying github repository [2]. See Appendix A for details on how to reproduce all results. Organization of this paper. We start in Section2and Section3with background and an overview of our general approach. In Section4, we study three well-known widely-available DNN CV architectures (the VGG, ResNet, and DenseNet series of models); and we provide an illustration of our basic methodology, both to evaluate the different metrics against reported test accuracies and to use quality metrics to understand model properties. Then, in Section5, we look at several variations of a popular NLP DNN architecture (the OpenAI GPT and GPT2 models); and we show how model quality and properties vary between several variants of GPT and GPT2, including how metrics behave similarly and differently. Then, in Section6, we present results based on an analysis of hundreds of pretrained DNN models, showing how well each metric predicts the reported test accuracies, and how the PL-based metrics perform remarkably well. Finally, in Section7, we provide a brief discussion and conclusion. 3 One could of course use these techniques to improve training, and we have been asked about that, but we are not interested in that here. Our main goal here is to use these techniques to evaluate properties of state-of-the-art pretrained DNN models. 2 Background and Related Work Most theory for DNNs is applied to small toy models and assumes access to data. There is very little work asking how to predict, in a theoretically-principled manner, the quality of large-scale state-of-the-art DNNs, and how to do so without access to training data or testing data or details of the training protocol, etc. Our approach is, however, related to two other lines of work. Statistical mechanics theory for DNNs. Statistical mechanics ideas have long had influence on DNN theory and practice [3,4,5]; and our best-performing metrics (those using fitted PL exponents) are based on statistical mechanics [4,6,7,8,9], in particular the recently-developed Theory of Heavy Tailed Self Regularization (HT-SR) [6,7,9]. We emphasize that the way in which we (and HT-SR Theory) use statistical mechanics theory is quite different than the way it is more commonly formulated. Several very good overviews of the more common approach are available [3,5]. We use statistical mechanics in a broader sense, drawing upon techniques from quantitative nance and random matrix theory. Thus, much more relevant for our methodological approach is older work of Bouchaud, Potters, Sornette, and coworkers [10,11,12,13] on the statistical mechanics of heavy tailed and strongly correlated systems. Norm-based capacity control theory. There is also a large body of work on using norm- based metrics to bound generalization error [14,15,16]. In this area, theoretical work aims to prove generalization bounds, and applied work uses these norms to construct regularizers to improve training. While we do find that norms provide relatively good quality metrics, at least for distinguishing good-better-best among well-trained models, we are not interested in proving generalization bounds or developing new regularizers. 3 Methods Let us write the Energy Landscape (or optimization function, parameterized by <> and <>) for a DNN wit <> layers, activation functions <>, and <> weight matrices <> and biases <>, as: <> (1) Each DNN layer contains one or more layer 2D <> weight matrices, <>, or pre-activation maps, <>, extracted from 2D Convolutional layers, and whereN > M.4 (We may drop the i and/or <> subscripts below.) See Appendix A for how we define the Conv2D layer matrixes and for our choices of normalization. Assume we are given several pretrained DNNs, e.g., as part of an architecture series. The models have been trained and evaluated on labeled data fdi <>, using standard techniques. The pretrained pytorch model files are publicly-available, and the test accuracies have been reported online. In this study, we do not have access to this data, and we have not trained any of the models ourselves, nor have we re-evaluated the test accuracies. We expect that most well-trained, production-quality models will employ one or more forms of regularization, such as Batch Normalization (BN), Dropout, etc., and many will also contain additional structure such as Skip Connections, etc. Here, we will ignore these details, and will focus only on the pretrained layer weight matrices Wl . 4 We do not use intra-layer information from the models in our quality metrics, but (as we will describe) our metrics can be used to learn about intra-layer model properties. DNN Empirical Quality Metrics. The best performing empirical quality metrics depend on the norms and/or spectral properties of each weight matrix,W, and/or, equivalently, it’s Empirical Correlation Matrix:X=WT W. Here, we consider the following metrics. <> Here, <> is the i th eigenvalue of the X, and <> is the maximum eigenvalue. Recall that the eigenvalues are squares of the singular values <> of <>. Also, note that we do not i normalize X by <>; see Appendix A for a discussion of this issue. The first two norms are well-known in ML; the last two deserve special mention. The empirical parameter is the Power Law (PL) exponent that arises in the recently-developed HT-SR Theory [6,7,9]. Operationally, is determined by using the publicly-available Weight Watcher tool [1] to fit the Empirical Spectral Density (ESD) of X, i.e., a histogram of the eigenvalues, call it <>, to a truncated PL, <> (2) Each of these quantities is defined for a given layer W matrix. For norm-based metrics, we use the average of the log norm, and to the appropriate power. Informally, this amounts to assuming that the layer weight matrices are statistically independent, in which case we can estimate the model complexityC, or test accuracy, with a standard Product Norm (which resembles a data dependent VC complexity), <>; (3) where <> is a matrix norm. The log complexity, <>; (4) takes the form of an average Log Norm. For the Frobenius Norm metric and Spectral Norm metric, we can use Eqn. (4) directly. 6 The Weighted Alpha metric is an average of <> over all layers <>, weighted by the size, or scale, or each matrix, <>; (5) where L is the total number of layer weight matrices. The Weighted Alpha metric was introduced previously [9], where it was shown to correlate well with trends in reported test accuracies of pretrained DNNs, albeit on a limited set of models. Based on this, in this paper, we introduce and evaluate the -Shatten Norm metric. Notice for the -Shatten Norm metric, however,l varies from layer to layer, and so in Eqn. (6) it can not be taken out of the sum: 5 Notice <>. 6 When taking <>, the 2 comes down and out of the sum, and thus ignoring it only changes the metric F by a constant factor. We use X to emphasize that <> depends on the ESD of X.2 <> (6) For small <>, the Weighted Alpha metric approximates the Log -Shatten norm, as can be shown with a statistical mechanics and random matrix theory derivation [17]; and the Weighted Alpha and -Shatten norm metrics often behave like an improved, weighted average Log Spectral Norm, and may track this metric in some cases. To avoid confusion, let us clarify the relationship between <> and <>. We fit the ESD of the correlation matrix X to a truncated PL, parameterized by 2 values: the PL exponent <>, and the maximum eigenvalue <>. (Technically, we also need the minimum eigenvalue <>, but this detail does not affect our analysis.) The PL exponent <> measures of the amount of correlation in a DNN layer weight matrixW. It is valid for <>, and it is scale-invariant, i.e., it does not depend on the normalization ofWorX. The <> is a measure of the size, or scale, of W. Multiplying each <> by the corresponding log <> weighs \bigger" layers more, and averaging this product leads to a balanced, Weighted Alpha metric for the entire DNN. Convolutional Layers and Normalization issues. There are several technical issues (regarding spectral analysis of convolutional layers and normalization of empirical matrices) that are important for reproducibility of our results. See Appendix A for a discussion. 4 Comparison of CV models In this section, we examine empirical quality metrics described in Section3for several CV model architecture series. This includes the VGG, ResNet, and DenseNet series of models, each of which consists of several pretrained DNN models, trained on the full ImageNet [18] dataset, and each of which is distributed with the current open source pyTorch framework (version 1.4) [19]. This also includes a larger set of ResNet models, trained on the ImageNet-1K dataset [18], provided on the OSMR \Sandbox for training convolutional networks for computer vision" [20], which we call the ResNet-1K series. We perform coarse model analysis, comparing and contrasting the four model series, and predicting trends in model quality. We also perform fine layer analysis, as a function of depth for these models, illustrating that PL-based metrics can provide novel insights among the VGG, ResNet/ResNet-1K, and DenseNet architectures. Average Quality Metrics versus Reported Test Accuracies. We have examined the performance of the four quality metrics (Log Frobenius norm, Log Spectral norm, Weighted Alpha, and Log -Norm) applied to each of the VGG, ResNet, ResNet-1K, and DenseNet series. To start, Figure1considers the VGG series (in particular, the pretrained models VGG11, VGG13, VGG16, and VGG19, with and without BN), and it plots the four quality metrics versus the reported test accuracies [19], 7 as well as a basic linear regression line. All four metrics correlate quite well with the reported Top1 accuracies, with smaller norms and smaller values of <> implying better generalization (i.e., greater accuracy, lower error). While all four metrics perform well, notice that the Log -Norm metric (<>) performs best (with an RMSE of 0:42, see Table 1); <> and the Weighted Alpha metric (<>), which is an approximation to the Log -Norm metric [17], performs second best (with an RMSE of 0:48, see Table1). 7 That is, these test accuracies have been previously reported and made publicly-available by others. We take them as given, and we do not attempt to reproduce/verify them, since we do not permit ourselves any access to training/test data. <

> Figure 1: Comparison of Average Log Norm and Weighted Alpha quality metrics versus reported test accuracy for pretrained VGG models (with and without BN), trained on ImageNet, available in pyTorch (v1.4). Metrics fit by linear regression, RMSE reported. See Table1for a summary of results for Top1 accuracies for all four metrics for the VGG, ResNet, and DenseNet series. Similar results (not shown) are obtained for the Top5 accuracies. Overall, for the the ResNet, ResNet-1K, and DenseNet series, all metrics perform relatively well, the Log-Norm metric performs second best, and the Weighted Alpha metric performs best. These model series are all well-trodden, and our results indicate that norm-based metrics and PL-based metrics can both distinguish among a series of \good-better-best" models, with PL- based metrics performing somewhat (i.e., quantitatively) better. The DenseNet series has similar behavior to what we see in Figures1and2for the other models. However, as noted in Table1, it has only 4 data points. In our larger analysis, in Section6, we will only include series with 5 or more models. (Note that these and many other such plots can be seen on our publicly-available repo.) Variation in Data Set Size. We are interested in how our four quality metrics depend on data set size. To examine this, we look at results on ResNet versus ResNet-1K. See Figure2, which plots and compares the Log-Norm metric for the full ResNet model, trained on the full ImageNet dataset, against the ResNet-1K model, which has been trained on a much smaller ImageNet-1K data set. The Log-Norm is much better than the Log Frobenius/Spectral norm metrics (although, as Table1shows, it is actually slightly worse than the Weighted Alpha metric). The ResNet series has strong correlation, with an RMSE of 0:66, whereas the ResNet-1K series <

> Table 1: RMSE (smaller is better) for linear fits of quality metrics to reported Top1 test error for pretrained models in each architecture series. Column # refers to number of models. VGG, ResNet, and DenseNet were pretrained on ImageNet, and ResNet-1K was pretrained on ImageNet- 1K. also shows good correlation, but has a much larger RMSE of 1:9. (Other metrics exhibit similar behavior.) As expected, the higher quality data set shows a better fit, even with fewer data points. Layer Analysis: Metrics as a Function of Depth. We can learn much more about a pretrained model by going beyond average values of quality metrics to examining quality metrics for each layer weight matrix,W, as a function of depth (or layer id). For example, we can plot (just) the PL exponent, , for each layer, as a function of depth. See Figure3, which plots for each layer (the first layer corresponds to data, the last layer to labels) for the least accurate (shallowest) and most accurate (deepest) model in each of the VGG (no BN), ResNet, and DenseNet series. (Again, a much more detailed set of plots is available at our repo; but note that the corresponding layer-wise plots for Frobenius and Spectral norms are much less interesting than the results we present here.) In the VGG models, Figure3(a)shows that the PL exponent systematically increases as we move down the network, from data to labels, in the Conv2D layers, starting with <> and reaching all the way to <> and then, in the last three, large, fully-connected (FC) layers, stabilizes back down to <>. This is seen for all the VGG models (again, only the shallowest and deepest are shown in this figure), indicating that the main effect of increasing depth is to increase the range over which increases, thus leading to larger values in later Conv2D layers of the VGG models. This is quite different than the behavior of either the ResNet-1K models or the DenseNet models. For the ResNet-1K models, Figure 3 (b) shows that also increases in the last few layers (more dramatically, in fact, than for VGG, observe the differing scales on the Y axes). However, <

> Figure 3: PL exponent () versus layer id, for the least and the most accurate models in VGG (a), ResNet (b), and DenseNet (c) series. (VGG is without BN; and note that the Y axes on each plot are different.) Subfigure (d) displays the ResNet models (b), zoomed in to 2 [1;5], and with the layer ids overlaid on the X-axis, from smallest to largest, to allow a more detailed analysis of the most strongly correlated layers. Notice that ResNet152 exhibits different and much more stable behavior of across layers. This contrasts with how both VGG models gradually worsen in deeper layers and how the DenseNet models are much more erratic. In the text, this is interpreted in terms ofCorrelation Flow. as the ResNet-1K models get deeper, there is a wide range over which values tend to remain quite small. This is seen for other models in the ResNet-1K series, but it is most pronounced for the larger ResNet-1K (152) model, whereremains relatively stable at <>, from the earliest layers all the way until we reach close to the final layers. For the DenseNet models, Figure 3 (c) shows that fi tends to increase as the layer id increases, in particular for layers toward the end. While this is similar to what is seen in the VGG models, with the DenseNet models, values increase almost immediately after the first few layers, and the variance is much larger (in particular for the earlier and middle layers, where it can range all the way to <>) and much less systematic throughout the network. Comparison of VGG, ResNet, and DenseNet Architectures. We can interpret these observations by recalling the architectural differences between the VGG, ResNet, and DenseNet architectures, and, in particular, the number of of residual connections. VGG resembles the traditional convolutional architectures, such as LeNet5, and consists of several [Conv2D-Maxpool- <

> Figure 4: ResNet20, distilled with Group Regularization, as implemented in the distiller (4D regularized 5L removed) pretrained models. Log Spectral Norm (<>) and PL exponent (<>) for individual layers, versus layer id, for both baseline (before distillation, green) and fine- tuned (after distillation, red) pretrained models. ReLu] blocks, followed by 3 large Fully Connected (FC) layers. ResNet greatly improved on VGG by replacing the large FC layers, shrinking the Conv2D blocks, and introducing residual connections. This optimized approach allows for greater accuracy with far fewer parameters (and GPU memory requirements), and ResNet models of up to 1000 layers have been trained [21]. We conjecture that the efficiency and effectiveness of ResNet is reflected in the smaller and more stable <>, across nearly all layers, indicating that the inner layers are very well correlated and strongly optimized. Contrast this with the DenseNet models, which contains many connections between every layer. Our results (large , meaning they even a PL model is probably a poor fit) suggest that DenseNet has too many connections, diluting high quality interactions across layers, and leaving many layers very poorly optimized. Correlation Flow. More generally, we can understand the results presented in Figure3in terms of what we will call theCorrelation Flow of the model. Recall that the average Log - Norm metric and the Weighted Alpha metric are based on HT-SR Theory [6,7,9], which is in turn based on ideas from the statistical mechanics of heavy tailed and strongly correlated systems [10,11,12,13]. There, one expects the weight matrices of well-trained DNNs will exhibit correlations over many size scales. Their ESDs can be well-fit by a (truncated) PL, with exponents <>. Much larger values (<>) may reflect poorer PL fits, whereas smaller values (<>), are associated with models that generalize better. Informally, one would expect a DNN model to perform well when it facilitates the propagation of information/features across layers. Previous work argues this by computing the gradients over the input data. In the absence of training/test data, one might hope that this leaves empirical signatures on weight matrices, and thus we can to try to quantify this by measuring the PL properties of weight matrices. In this case, smaller values correspond to layers in which correlations across multiple scales are better captured [6,11], and we expect that small values that are stable across multiple layers enable better correlation flow through the network. We have seen this in many models, including those shown in Figure3. Scale Collapse; or How Distillation May Break Models. The similarity between norm- based metrics and PL-based metrics suggests a question: is the Weighted Alpha metric just a variation of the more familiar norm-based metrics? More generally, do fitted values contain information not captured by norms? In examining hundreds of pretrained models, we have found several anomalies that demonstrate the power of our approach. In particular, to show that does capture something different, consider the following example, which looks at a compressed/distilled DNN model [22]. In this example, we show that some distillation methods may actually break models unexpectedly by introducing what we callScale Collapse, where several distilled layers have unexpectedly small Spectral Norms. We consider ResNet20, trained on CIFAR10, before and after applying the Group Regularization distillation technique, as implemented in the distiller package [23]. We analyze the pretrained 4D regularized 5L removed baseline and fine-tuned models. The reported baseline test accuracies (Top1= 91:45 and Top5= 99:75) are better than the reported fine-tuned test accuracies (Top1= 91:02 and Top5= 99:67). Because the baseline accuracy is greater, the previous results on ResNet (Table1and Figure2) suggest that the baseline Spectral Norms should be smaller on average than the fine-tuned ones.The opposite is observed.Figure4presents the Spectral Norm (here denoted <> ) and PL exponent () for each individual layer weight matrixW.8 On the other hand, the values (in Figure 4 (b)) do not differ systematically between the baseline and fine-tuned models. Also (not shown), the average (unweighted) baseline is smaller than the fine-tuned average (as predicted by HT-SR Theory, the basis of <>). That being said, Figure4(b)also depicts two very large 6 values for the baseline, but not for the fine-tuned, model. This suggests the baseline model has at least two over- parameterized/under-trained layers, and that the distillation method does, in fact, improve the fine-tuned model by compressing these layers. The pretrained models in the distiller package have passed some quality metric, but they are much less well trodden than any of the VGG, ResNet, or DenseNet series. While norms make good regularizers for a single model, there is no reason a priori to expect them correlate so well with test accuracies across different models. We do expect, however, the PL fit o do so because it effectively measures the amount of correlation in the model [6,7,9]. The reason for the anomalous behavior shown in Figure4is that the distiller Group Regularization technique causes the norms of the W pre-activation maps for two Conv2D layers to increase spuriously. This is difficult to diagnose by analyzing training/test curves, but it is easy to diagnose with our approach. 5 Comparison of NLP Models In this section, we examine empirical quality metrics described in Section3for several NLP model architectures. Within the past two years, nearly 100 open source, pretrained NLP DNNs based on the revolutionary Transformer architecture have emerged. These include variants of BERT, Transformer-XML, GPT, etc. The Transformer architectures consist of blocks of so-called Attention layers, containing two large, Feed Forward (Linear) weight matrices [24]. In contrast to smaller pre-Activation maps arising in Cond2D layers, Attention matrices are significantly larger. In general, we have found that they have larger PL exponents . Based on HT-SR Theory (in particular, the interpretation of values of 2 as modeling systems with good correlations over many size scales [10,11]), this suggests that these models fail to capture successfully many of the correlations in the data (relative to their size) and thus are substantially under-trained. More generally, compared to the CV models of Section4, modern NLP models have larger weight matrices and display different spectral properties. Thus, they provide a very different test for our empirical quality metrics. While norm-based metrics perform reasonably well on well-trained NLP models, they often behave anomalously on poorly-trained models. Indeed, for such \bad" models, weight matrices may display rank collapse, decreased Frobenius mass, or unusually small Spectral norms. (This may be misinterpreted as \smaller is better.") In contrast, PL-based metrics, including the Log -Norm metric (<>) and the Weighted Alpha metric (<>) display consistent behavior, even on poorly trained models. Indeed, we can use these metrics to help identify when architectures need repair and when more and/or better data are needed. What do large values of mean? Many NLP models, such as GPT and BERT, have some weight matrices with unusually large PL exponents (e.g.,6). This indicates these matrices may be under-correlated (i.e., over-parameterized, relative to the amount of data). In this regime, the truncated PL fit itself may not be very reliable because the MLE estimator it uses is unreliable in this range (i.e., the specific values returned by the truncated PL fits are less reliable, but having large versus small values of is reliable). Phenomenologically, if we examine the ESD visually, we can usually describe theseWas in the Bulk-Decayor Bulk-plus-Spikes phase [6,7]. Previous work [6,7] has conjectured that very well-trained DNNs would not have many outlier 6; and improved versions of GPT (shown below) and BERT (not shown) confirm this. OpenAI GPT Models. The OpenAI GPT and GPT2 models provide us with the opportunity to analyze two effects: training the same model with different data set sizes; and increasing the sizes of both the data set and the architectures simultaneously. These models have the remarkable ability to generate fake text that appears to the human to be real, and they have generated significant media attention because of the potential for their misuse. For this reason, the original GPT model released by OpenAI was trained on a deficient data set, rendering the model interesting but not fully functional. Later, OpenAI released a much improved model, GPT2-small, which has the same architecture and number of layers as GPT, but which has been trained on a larger and better data set (and with other changes), making it remarkably good at generating (near) human-quality fake text. By comparing the poorly-trained (i.e., \bad") GPT to the well-trained (i.e., \good") GPT2-small, we can identify empirical indicators for when a model has in fact been poorly-trained and thus may perform poorly when deployed. By comparing GPT2-medium to GPT2-large to GPT2-xl, we can examine the effect of increasing data set and model size simultaneously, an example of what we call a series of \good-better-best" models. The GPT models we analyze are deployed with the popular HuggingFace PyTorch library [25]. GPT has 12 layers, with 4 Multi-head Attention Blocks, giving 48 layer Weight Matrices,W. Each Block has 2 components, the Self Attention (attn) and the Projection (proj) matrices. The self-attention matrices are larger, of dimension (2304x768) or (3072x768). The projection layer concatenates the self-attention results into a vector (of dimension 768). This gives 50 large matrices. Because GPT and GPT2 are trained on different data sets, the initial Embedding matrices differ in shape. GPT has an initial Token and Positional Embedding layers, of dimension (40478x768) and (512x768), respectively, whereas GPT2 has input Embeddings of shape (50257x768) and (1024x768), respectively. The OpenAI GPT2 (English) models are: GPT2- small, GPT2-medium, GPT2-large, and GPT2-xl, having 12, 24, 36, and 48 layers, respectively, with increasingly larger weight matrices. Average Quality Metrics for GPT and GPT2. We have analyzed the four quality metrics described in Section3for the OpenAI GPT and GPT2 pretrained models. See Table2for a summary of results. We start by examining trends between GPT and GPT2-small. Observe that all four metrics increase when going from GPT to GPT2-small, i.e., they are smaller for the higher-quality model (higher quality since GPT was trained to better data), when the number of layers is held xed. Notice that in the GPT model, being poorly trained, the norm metrics all exhibitScale Collapse, compared to GPT2-small. <

> Table 2: Average value for the average Log Norm and Weighted Alpha metrics for pretrained OpenAI GPT and GPT2 models. Column # refers to number of layers treated. Note that the averages do not include the first embedding layer(s) because they are not (implicitly) normalized. We next examine trends between GPT2-medium to GPT2-large to GPT2-xl. Observe that (with one minor exception involving the log Frobenius norm metric) all four metrics decrease as one goes from medium to large to xl, indicating that the larger models indeed look better than the smaller models. Notice that, for these well-trained models, the norm metrics now behave as expected, decreasing with increasing accuracy. Going beyond average values, Figure5(a)shows the histogram (empirical density), for all layers, of for GPT and GPT2-small. These two histograms are very different. The older deficient GPT has numerous unusually large exponents meaning they are not really well- described by a PL fit. Indeed, we expect that a poorly-trained model will lack good (i.e., small) PL behavior in many/most layers. On the other hand, as expected, the newer improved GPT2- small model has, on average, smaller values than the older GPT, with all 6 and with smaller mean/median. It also has far fewer unusually-large outlying values than GPT. From this (and other results not shown), we see that provides a good quality metric for comparing these two models, the \bad" GPT versus the \good" GPT2-small. This should be contrasted with the behavior displayed by the Frobenius norm (not shown) and the Spectral norm. Scale Collapse in Poorly Trained Models. We next describe the behavior of the Spectral norm in GPT versus GPT2-small. In Figure5(b), the \bad" GPT model has a smaller mean/median Spectral norm as well as, spuriously, many much smaller Spectral norms, com- pared to the \good" GPT2-small, violating the conventional wisdom that smaller Spectral norms are better. Indeed, because there are so many anonymously small Spectral norms, it appears that the GPT model may be exhibiting a kind ofScale Collapse, like that observed in the distilled CV models (in Figure4). This is important because it demonstrates that, while the Spectral (or Frobenius) norm may correlate well with predicted test error, it is not a good indicator of the overall model quality. It can mispredict good-versus-bad questions in ways not seen with PL-based metrics. Using it as an empirical quality metric may give spurious results when applied to poorly-trained or otherwise deficient models. (Note that Figure5(b)also shows some unusually large Spectral Norms. Upon examination, e.g., from Figure6(b)(below), we see that these correspond to the first embedding layer(s). These layers have a different effective normalization, and therefore a different scale. We discuss this further in AppendixA. Here, we do not include them in our computed average metrics in Table2, and we do not include them in the histogram plot in Figure5(b).) Layer Analysis: Correlation Flow and Scale Collapse in GPT and GPT2. We also examine in Figure 6 the PL exponent and Log Spectral Norm versus layer id, for GPT and GPT2-small. Let’s start with Figure6(a), which plots versus the depth (i.e., layer id) for each model. The deficient GPT model displays two trends in , one stable with 4, and one <

> Figure 5: Histogram of PL exponents (<>) and Log Spectral Norms (<>) for weight matrices from the OpenAI GPT and GPT2-small pretrained models. increasing with layer id, with reaching as high as 12. In contrast, the well-trained GPT2-small model shows consistent and stable patterns, again with one stable <> (and below the GPT trend), and the other only slightly trending up, with 6. The scale-invariant metric lets us identify potentially poorly-trained models. These results show that the Correlation Flow differs significantly between GPT and GPT2-small (with the better GPT2-small looking more like the better ResNet-1K from Figure3(b)). These results should be contrasted with the corresponding results for Spectral Norms, shown in Figure6(b). Attention models have two types of layers, one small and large; and the Spectral Norm, in particular, displays unusually small values for some of these layers for GPT. This Scale Collapse for the poorly-trained GPT is similar to what we observed for the distilled ResNet20 model in Figure4(b). Because of the anomalous scale collapse that is frequently observed in poorly-trained models, these results suggest that scale-dependent norm metrics should not be directly applied to distinguish good-versus-bad models. <

> Figure 6: PL exponents (<>) (in (a)) and Log Spectral Norms (<>) (in (b)) for weight matrices from the OpenAI GPT and GPT2-small pretrained models. (Note that the quantities being shown on each Y axis are different.) In the text, this is interpreted in terms ofCorrelation Flow and Scale Collapse. GPT2: medium, large, xl. We now look across series of increasingly improving GPT2 models (i.e., we consider good-better-best questions), by examining both the PL exponent as well as the Log Norm metrics. In general, as we move from GPT2-medium to GPT2-xl, histograms for both exponents and the Log Norm metrics downshift from larger to smaller values. For example, see Figure7, which shows the histograms over the layer weight matrices for fitted PL exponent (<>) and the Log Alpha Norm (<>) metric. We see that the average decreases with increasing model size, although the differences are less noticeable between the differing good-better-best GTP2 models than between the good- versus-bad GPT and GPT2-small models. Unlike GPT, however, the layer Log Alpha Norms behave more as expected for GPT2 layers, with the larger models consistently having smaller norms. Similarly, the Log Spectral Norm also decreases on average with the larger models (not shown). As expected, the norm metrics can indeed distinguish among good-better-best models among a series well-trained models. We do notice, however, that while the peaks of the are getting smaller, towards 2:0, the tails of the distribution shifts right, with larger GPT2 models having more usually large (also not shown). We suspect this indicates that these larger GPT2 models are still under-optimized/over- parameterized (relative to the data on which they were trained) and that they have capacity to support datasets even larger than the recent XL 1.5B release [26]. <

> Figure 7: Histogram of PL exponents (<>) and Log Alpha Norm (<>) for weight matrices from models of different sizes in the GPT2 architecture series. (Plots omit the first 2 (embedding) layers, because they are normalized differently giving anomalously large values.) 6 Comparing Hundreds of CV Models In this section, we summarize results from a large-scale analysis of hundreds of CV models, including models developed for image classification, segmentation, and a range of related tasks. Our aim is to complement the detailed results from Sections4and5by providing broader conclusions. The models we consider have been pretrained on nine datasets. We provide full details about how to reproduce these results in AppendixA. We choose ordinary least squares (OLS) regression to quantify the relationship between quality metrics (computed with the Weight Watcher tool ) and the reported test error and/or accuracy metrics. We regress the metrics on the Top1 (and Top5) reported errors (as dependent variables). These include Top5 errors for the ImageNet-1K model, percent error for the CIFAR-10/100, SVHN, CUB-200-2011 models, and Pixel accuracy (Pix.Acc.) and Intersection-Over-Union (IOU) for other models. We regress them individually on each of the norm-based and PL-based metrics, as described in Section4. Our results are summarized in Table3. For the mean, largerR2 and smaller MSE are desirable; and for the standard deviation, smaller values are desirable. Taken as a whole, over the entire corpus of data, PL-based metrics are somewhat better for both theR2 mean and standard deviation; and PL-based metrics are much better for MSE mean and standard deviation. These <

> Table 3: Comparison of linear regression fits for different average Log Norm and Weighted Alpha metrics across 5 CV datasets, 17 architectures, covering 108 (out of over 400) different pretrained DNNs. We include regressions only for architectures with five or more data points, and which are positively correlated with test error. These results can be readily reproduced using the Google Colab notebooks (see AppendixA). (and other) results suggest our conclusions from Sections4and5hold much more generally, and they suggest obvious questions for future work. 7 Conclusion We have developed (based on strong theory) and evaluated (on a large corpus of publicly-available pretrained models from CV and NLP) methods to predict trends in the quality of state-of-the-art neural networks|without access to training or testing data. Prior to our work, it was not obvious that norm-based metrics would perform well to predict trends in quality across models (as they are usually used within a given model or parameterized model class, e.g., to bound generalization error or to construct regularizers). Our results are the first to demonstrate that they can be used for this important practical problem. That PL-based metrics perform better (than norm-based metrics) should not be surprising|at least to those familiar with the statistical mechanics of heavy tailed and strongly correlated systems [10,11,12,13] (since our use of PL exponents is designed to capture the idea that well-trained models capture correlations over many size scales in the data). Again, though, our results are the first to demonstrate this. It is also gratifying that our approach can be used to provide fine-scale insight (such as rationalizing the flow of correlations or the collapse of size scale) throughout a network. We conclude with a few comments on what a practical theory of DNNs should look like. To do so, we distinguish between two types of theories:non-empirical or analogical theories, in which one creates, often from general principles, a very simple toy model that can be analyzed rigorously, and one then argues that the model is relevant to the system of interest; and semi-empirical theories, in which there exists a rigorous asymptotic theory, which comes with parameters, for the system of interest, and one then adjusts or fits those parameters to the finite non-asymptotic data. A drawback of the former approach is that it typically makes very strong assumptions on the data, and the strength of those assumptions can limit the practical applicability of the theory. Nearly all of the work on the theory of DNNs focuses on the former type of theory. Our approach focuses on the latter type of theory. Our results, which are based on using sophisticated statistical mechanics theory and solving important practical DNN problems, suggests that the latter approach should be of interest more generally for those interested in developing a practical DNN theory. Acknowledgements. MWM would like to acknowledge ARO, DARPA, NSF, and ONR as well as the UC Berkeley BDD project and a gift from Intel for providing partial support of this work. We would also like to thank Amir Khosrowshahi and colleagues at Intel for helpful discussion regarding the Group Regularization distillation technique. References [1]WeightWatcher, 2018.https://pypi.org/project/WeightWatcher/. [2]https://github.com/CalculatedContent/ww-trends-2020. [3]A. Engel and C. P. L. Van den Broeck.Statistical mechanics of learning. Cambridge University Press, New York, NY, USA, 2001. [4]C. H. Martin and M. W. Mahoney. Rethinking generalization requires revisiting old ideas: statistical mechanics approaches and complex learning behavior. Technical Report Preprint:arXiv:1710.09553, 2017. [5]Y. Bahri, J. Kadmon, J. Pennington, S. Schoenholz, J. Sohl-Dickstein, and S. Ganguli. Statistical mechanics of deep learning.Annual Review of Condensed Matter Physics, pages 000{000, 2020. [6]C. H. Martin and M. W. Mahoney. Implicit self-regularization in deep neural networks: Evidence from random matrix theory and implications for learning. Technical Report Preprint:arXiv:1810.01075, 2018. [7]C. H. Martin and M. W. Mahoney. Traditional and heavy-tailed self regularization in neural network models. InProceedings of the 36th International Conference on Machine Learning, pages 4284{4293, 2019. [8]C. H. Martin and M. W. Mahoney. Statistical mechanics methods for discovering knowledge from modern production quality neural networks. InProceedings of the 25th Annual ACM SIGKDD Con- ference, pages 3239{3240, 2019. [9]C. H. Martin and M. W. Mahoney. Heavy-tailed Universality predicts trends in test accuracies for very large pre-trained deep neural networks. InProceedings of the 20th SIAM International Conference on Data Mining, 2020. [10]J. P. Bouchaud and M. Potters.Theory of Financial Risk and Derivative Pricing: From Statistical Physics to Risk Management. Cambridge University Press, 2003. [11]D. Sornette.Critical phenomena in natural sciences: chaos, fractals, selforganization and disorder: concepts and tools. Springer-Verlag, Berlin, 2006. [12]J. P. Bouchaud and M. Potters. Financial applications of random matrix theory: a short review. In G. Akemann, J. Baik, and P. Di Francesco, editors,The Oxford Handbook of Random Matrix Theory. Oxford University Press, 2011. [13]J. Bun, J.-P. Bouchaud, and M. Potters. Cleaning large correlation matrices: tools from random matrix theory.Physics Reports, 666:1{109, 2017. [14]B. Neyshabur, R. Tomioka, and N. Srebro. Norm-based capacity control in neural networks. In Proceedings of the 28th Annual Conference on Learning Theory, pages 1376{1401, 2015. [15]P. Bartlett, D. J. Foster, and M. Telgarsky. Spectrally-normalized margin bounds for neural networks. Technical Report Preprint:arXiv:1706.08498, 2017. [16]Q. Liao, B. Miranda, A. Banburski, J. Hidary, and T. Poggio. A surprising linear relationship predicts test performance in deep networks. Technical Report Preprint:arXiv:1807.09659, 2018. [17]C. H. Martin and M. W. Mahoney. Unpublished results, 2020. [18]O. Russakovsky et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211{252, 2015. [19]A. Paszke et al. Pytorch: An imperative style, high-performance deep learning library. InAnnual Advances in Neural Information Processing Systems 32: Proceedings of the 2019 Conference, pages 8024{8035, 2019. [20]Sandbox for training convolutional networks for computer vision. https://github.com/osmr/ imgclsmob. [21]K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual networks. Technical Report Preprint:arXiv:1603.05027, 2016. [22]Y. Cheng, D. Wang, P. Zhou, and T. Zhang. A survey of model compression and acceleration for deep neural networks. Technical Report Preprint:arXiv:1710.09282, 2017. [23]Intel Distiller package.https://nervanasystems.github.io/distiller. [24]A. Vaswani et al. Attention is all you need. Technical Report Preprint:arXiv:1706.03762, 2017. [25]T. Wolf et al. Huggingface’s transformers: State-of-the-art natural language processing. Technical Report Preprint:arXiv:1910.03771, 2019. [26]OpenAI GPT-2: 1.5B Release.https://openai.com/blog/gpt-2-1-5b-release/. [27]H. Sedghi, V. Gupta, and P. M. Long. The singular values of convolutional layers. Technical Report Preprint:arXiv:1805.10408, 2018. [28]X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. InProceedings of the 13th International Workshop on artificial Intelligence and Statistics, pages 249{256, 2010. A Appendix In this appendix, we provide more details on several issues that are important for the reproducibility of our results. All of our computations were performed with the Weight Watcher tool (version 0.2.7) [1]. More details and more results are available in an accompanying github repository [2]. A.1 Reproducibility Considerations SVD of Convolutional 2D Layers. There is some ambiguity in performing spectral analysis on Conv2D layers. Each layer is a 4-index tensor of dimension (<>), with an (wxh) filter (or kernel) and (in;out) channels. When w=h=k, it gives (kxk) tensor slices, or pre-Activation Maps Wi;L of dimension (in/out) each. We identify 3 different approaches for running SVD on a Conv2D layer: 1.run SVD on each pre-Activation MapWi;L , yielding (kxk) sets of M singular values; 2.stack the maps into a single matrix of, say, dimension (<>), and run SVD to get in singular values; 3.compute the 2D Fourier Transform (FFT) for each of the (in;out) pairs, and run SVD on the Fourier coefficients [27], leading to <> non-zero singular values. Each method has tradeoffs. Method (3) is mathematically sound, but computationally expensive. Method (2) is ambiguous. For our analysis, because we need thousands of runs, we select method (1), which is the fastest (and is easiest to reproduce). Normalization of Empirical Matrices. Normalization is an important, if underappreciated, practical issue. Importantly, the normalization of weight matrices does not affect the PL fits because is scale-invariant. Norm-based metrics, however, do depend strongly on the scale of the weight matrix|that is the point. To apply RMT, we usually define X with a <> normalization, assuming variance of <>. Pretrained DNNs are typically initialized with random weight matrices <>, with <>, or some variant, e.g., the Glorot/Xavier normalization [28], or <> normalization for Convolutional 2D Layers. With this implicit scale, we do not renormalize the empirical weight matrices, i.e., we use them as-is. The only exception is that <> Table 4: Jupyter notebooks used to reproduce all results in Sections 4 and 5. pwe do rescale the Conv2D pre-activation mapsWi;L byk= 2 so that they are on the same scale as the Linear / Fully Connected (FC) layers. Special consideration for NLP models. NLP models, and other models with large initial p embeddings, require special care because the embedding layers frequently lack the implicit <> normalization present in other layers. For example, in GPT, for most layers, the maximum eigenvalue <>, but in the first embedding layer, the maximum eigenvalue is of orderN(the number of words in the embedding), or <>). For GPT and GPT2, we treat all layers as-is (although one may want to normalize the first 2 layers X by <>, or to treat them as outliers). A.2 Reproducing Sections 4 and 5 We provide a github repository for this paper that includes Jupyter notebooks that fully reproduce all results (as well as many other results) [2]. All results have been produced using the Weight- Watcher tool (v0.2.7) [1]. The ImageNet and OpenAI GPT pretrained models are provided in the current pyTorch [19] and Huggingface [25] distributions, as specified in the requirements.txt file. A.3 Reproducing Figure 4, for the Distiller Model In the distiller folder of our github repo, we provide the original Jupyter Notebooks, which use the Intel distiller framework [23]. Figure4is from the‘‘...-Distiller-ResNet20.ipynb’’ notebook (see Table4). For completeness, we provide both the results described here, as well as additional results on other pretrained and distilled models using the Weight Watcher tool . A.4 Reproducing Table 3 in Section 6 In the ww-colab folder of our github repo, we provide several Google Colab notebooks which can be used to reproduce the results of Section6. The ImageNet-1K and other pretrained models are taken from the pytorch models in theomsr/imgclsmob\Sandbox for training convolutional net- works for computer vision" github repository [20]. The data for each regression can be generated in parallel by running each Google Colab notebook (i.e.,wwcolab0100.ipynb) simultaneously on the same account. The data generated are analyzed withww colabresults.ipynb, which runs all regressions and which tabulates the results presented in Table3. We attempt to run linear regressions for all pyTorch models for each architecture series for all datasets provided. There are over 450 models in all, and we note that theosmr/imgclsmob repository is constantly being updated with new models. We omit the results for CUB-200-2011, <> Table 5: Datasets used <> Table 6: Architectures used Pascal-VOC2012, ADE20K, and COCO datasets, as there are fewer than 15 models for those datasets. Also, we filter out regressions with fewer than 5 datapoints. We remove the following outliers, as identified by visual inspection:efficientb0,b2. We also remove the entirecifar100 ResNeXTseries, which is the only example to show no trends with the norm metrics. The final datasets used are shown in Table 5. The final architecture series used are shown in Table6, with the number of models in each. To explain further how to reproduce our analysis, we run three batches of linear regressions. First, at the global level, we divide models by datasets and run regressions separately on all models of a certain dataset, regardless of the architecture. At this level, the plots are quite noisy and clustered, as each architecture has its own accuracy trend; but one can still see that most plots show positive relationship with positive coefficients. Example regressions are shown in Figure8, as available in the results notebook. To generate the results in Table3, we run linear regressions for each architecture series in Table6, regressing each empirical Log Norm metric against the reported Top1 (and Top5) errors (as listed on theosmr/imgclsmobgithub repository README file [20], with the relevant data extracted and provided in our github repo aspytorchcv.html). We record theR2 andMSE for each metric, averaged over all regressions for all architectures and datasets. See Table7and Table8. In the repo, plots are provided for every regression, and more fine grained results may be computed by the reader by analyzing the data in thedf all.xlsxfile. The final analysis includes 108 regressions in all, those with 4 or more models, with a positive R2. <> Table 7: MSEResults for all CV model regressions. <> Table 8: R2 Results for all CV model regressions. <

> Figure 8: PL exponentfiversus reported Top1 Test Accuracies for pretrained DNNs available for five different data sets. <> <> <> <> <> <> Pruning neural networks without any data by iteratively conserving synaptic flow Hidenori Tanaka Daniel Kunin Physics & Informatics Laboratories Institute for Computational and NTT Reserach, Inc. Mathematical Engineering Department of Applied Physics Stanford University Stanford University Daniel L. K. Yamins Surya Ganguli Department of Psychology Department of Applied Physics Department of Computer Science Stanford University Stanford University Abstract Pruning the parameters of deep neural networks has generated intense interest due to potential savings in time, memory and energy both during training and at test time. Recent works have identified, through an expensive sequence of training and pruning cycles, the existence of winning lottery tickets or sparse trainable subnetworks at initialization. This raises a foundational question: can we identify highly sparse trainable subnetworks at initialization, without ever training, or indeed without ever looking at the data? We provide an affirmative answer to this question through theory driven algorithm design. We first mathematically formulate and experimentally verify a conservation law that explains why existing gradient- based pruning algorithms at initialization suffer from layer-collapse, the premature pruning of an entire layer rendering a network untrainable. This theory also elucidates how layer-collapse can be entirely avoided, motivating a novel pruning algorithmIterative Synaptic Flow Pruning (SynFlow). This algorithm can be interpreted as preserving the total flow of synaptic strengths through the network at initialization subject to a sparsity constraint. Notably, this algorithm makes no reference to the training data and consistently outperforms existing state-of-the-art pruning algorithms at initialization over a range of models (VGG and ResNet), datasets (CIFAR-10/100 and Tiny ImageNet), and sparsity constraints (up to99:9percent). Thus our data-agnostic pruning algorithm challenges the existing paradigm that data must be used to quantify which synapses are important. 1 Introduction Network pruning, or the compression of neural networks by removing parameters, has been an important subject both for reasons of practical deployment [1,2,3,4,5,6,7] and for theoretical understanding of artificial [8] and biological [9] neural networks. Conventionally, pruning algorithms have focused on compressing pre-trained models [1,2,3,5,6]. However, recent works [10,11] have identified through iterative training and pruning cycles (iterative magnitude pruning) that there exist sparse subnetworks (winning tickets) in randomly-initialized neural networks that, when trained in isolation, can match the test accuracy of the original network. Moreover, its been shown that some of these winning ticket subnetworks can generalize across datasets and optimizers [12]. While these results suggest training can be made more efficient by identifying winning ticket subnetworks at initialization, they do not provide efficient algorithms to find them. Typically, it requires significantly more computational costs to identify winning tickets through iterative training and pruning cycles than simply training the original network from scratch [10,11]. Thus, the fundamental unanswered question is: can we identify highly sparse trainable subnetworks at initialization, without ever training, or indeed without ever looking at the data? Towards this goal, we start by investigating the limitations of existing pruning algorithms at initialization [13,14], determine simple strategies for avoiding these limitations, and provide a novel data-agnostic algorithm that improves upon state-of-the-art results. Our main contributions are: 1.We study layer-collapse, the premature pruning of an entire layer making a network untrainable, and formulate the axiomMaximal Critical Compression that posits a pruning algorithm should avoid layer-collapse whenever possible (Sec. 3). 2.We demonstrate theoretically and empirically that synaptic saliency, a general class of gradient-based scores for pruning, is conserved at every hidden unit and layer of a neural network (Sec. 4). 3.We show that the se conservation law simply parameters in large layers receive lower scores than parameters in small layers, which elucidates why single-shot pruning disproportionately prunes the largest layer leading to layer-collapse (Sec. 4). 4.We hypothesize that iterative magnitude pruning[10] avoids layer-collapse because gradient descent effectively encourages the magnitude scores to observe a conservation law, which combined with iteration results in the relative scores for the largest layers increasing during pruning (Sec. 5). 5.We prove that a pruning algorithm avoids layer-collapse entirely and satisfies Maximal Critical Compression if it uses iterative, positive synaptic saliency scores (Sec. 6). 6.We introduce a new data-agnostic algorithmIterative Synaptic Flow Pruning (SynFlow)that satisfies Maximal Critical Compression (Sec. 6) and demonstrate empirically 2 that this algorithm achieves state-of-the-art pruning performance on 12 distinct combinations of models and datasets (Sec. 7). 2 Related work While there are a variety of approaches to compressing neural networks, such as novel design of micro- architectures [15,16,17], dimensionality reduction of network parameters [18,19], and training of dynamic sparse networks [20, 21], in this work we will focus on neural network pruning. Pruning after training.Conventional pruning algorithms assign scores to parameters in neural networks after training and remove the parameters with the lowest scores [5,22,23]. Popular scoring metrics include weight magnitudes [4,6], its generalization to multi-layers [24], first- [1,25,26,27] and second-order [2,3,27] Taylor coefficients of the training loss with respect to the parameters, and more sophisticated variants [28,29,30]. While these pruning algorithms can indeed compress neural networks at test time, there is no reduction in the cost of training. Pruning before Training.Recent works demonstrated that randomly initialized neural networks can be pruned before training with little or no loss in the final test accuracy [10,13,31]. In particular, the Iterative Magnitude Pruning (IMP) algorithm [10,11] repeats multiple cycles of training, pruning, and weight rewinding to identify extremely sparse neural networks at initialization that can be trained to match the test accuracy of the original network. While IMP is powerful, it requires multiple cycles of expensive training and pruning with very specific sets of hyperparameters. Avoiding these difficulties, a different approach uses the gradients of the training loss at initialization to prune the network in a single-shot [13,14]. While these single-shot pruning algorithms at initialization are much more efficient, and work as well as IMP at moderate levels of sparsity, they suffer from layer-collapse, or the premature pruning of an entire layer rendering a network untrainable [32,33]. Understanding and circumventing this layer-collapse issue is the fundamental motivation for our study. 3 Layer-collapse: the key obstacle to pruning at initialization Broadly speaking, a pruning algorithm at initialization is defined by two steps. The first step scores the parameters of a network according to some metric and the second step masks the parameters (removes or keeps the parameter) according to their scores. The pruning algorithms we consider will always mask the parameters by simply removing the parameters with the smallest scores. This ranking process can be applied globally across the network, or layer-wise. Empirically, its been shown that global-masking performs far better than layer-masking, in part because it introduces fewer hyperparameters and allows for flexible pruning rates across the network [23]. However, recent works [32,14,33] have identified a key failure mode,layer-collapse, for existing pruning algorithms using global-masking. Layer-collapse occurs when an algorithm prunes all parameters in a single weight layer even when prunable parameters remain elsewhere in the network. This renders the network untrainable, evident by sudden drops in the achievable accuracy for the network as shown in Fig. 1. To gain insight into the phenomenon of layer-collapse we will define some useful terms inspired by a recent paper studying the failure mode [33]. Given a network,compression ratio (FORMULA) is the number of parameters in the original network divided by the number of parameters remaining after pruning. For example, when the compression ratio <>, then only one out of a thousand of the parameters remain after pruning.Max compression (<>) is the maximal <

> possible compression ratio for a network that doesn’t lead to layer-collapse. For example, for a network with L layers and N parameters, <>, which is the compression ratio associated with pruning all but one parameter per layer. Critical compression (<>) is the maximal compression ratio a given algorithm can achieve without inducing layer-collapse. In particular, the critical compression of an algorithm is always upper bounded by the max compression of the network:<>. Figure 1:Layer-collapse leads to a suddenThis inequality motivates the following axiom we postulate any drop in accuracy.Top-1 test accuracy as a successful pruning algorithm should satisfy. function of the compression ratio for a VGG- Axiom.Maximal Critical Compression.The critical compression 1 6 model pruned at initialization and trained of a pruning algorithm applied to a network should always on CIFAR-100. Colored arrows represent equal the max compression of that network. the critical compression of the corresponding pruning algorithm. Only our algorithm, In other words, this axiom implies a pruning algorithm should SynFlow, reaches the theoretical limit of max never prune a set of parameters that results in layer-collapse if compression (black dashed line) without col- there exists another set of the same cardinality that will keep lapsing the network. See Sec. 7 for more the network trainable. To the best of our knowledge, no exist- details on the experiments. ing pruning algorithm with global-masking satisfies this simple axiom. Of course any pruning algorithm could be modified to satisfy the axiom by introducing specialized layer-wise pruning rates. However, to retain the benefits of global-masking [23], we will formulate an algorithm, Iterative Synaptic Flow Pruning (SynFlow), which satisfies this property by construction. SynFlow is a natural extension of magnitude pruning, that preserves the total flow of synaptic strengths from input to output rather than the individual synaptic strengths themselves. We will demonstrate that not only does the SynFlow algorithm achieve Maximal Critical Compression, but it consistently outperforms existing state-of-the-art pruning algorithms (as shown in Fig. 1 and in Sec. 7), all while not using the data. Throughout this work, we benchmark our algorithm, SynFlow, against two simple baselines, random scoring and scoring based on weight magnitudes, as well as two state-of-the-art single-shot pruning algorithms, Single- shot Network Pruning based on Connection Sensitivity (SNIP) [13] and Gradient Signal Preservation (GraSP) [14]. SNIP [13] is a pioneering algorithm to prune neural networks at initialization by scoring weights based on the gradients of the training loss. GraSP [14] is a more recent algorithm that aims to preserve gradient flow at initialization by scoring weights based on the Hessian-gradient product. Both SNIP and GraSP have been thoroughly benchmarked by [14] against other state-of-the-art pruning algorithms that involve training [2, 34, 10, 11, 35, 21, 20], demonstrating competitive performance. 4 Conservation laws of synaptic saliency In this section, we will further verify that layer-collapse is a key obstacle to effective pruning at initialization and explore what is causing this failure mode. As shown in Fig. 2, with increasing compression ratios, existing random, magnitude, and gradient-based pruning algorithms will prematurely prune an entire layer making the network untrainable. Understanding why certain score metrics lead to layer-collapse is essential to improve the design of pruning algorithms. Random pruning prunes every layer in a network by the same amount, evident by the horizontal lines in Fig. 2. With random pruning the smallest layer, the layer with the least parameters, is the first to be fully pruned. Conversely, magnitude pruning prunes layers at different rates, evident by the staircase pattern in Fig. 2. Magnitude pruning effectively prunes parameters based on the variance of their initialization, which for common network initializations, such as Xavier [36] or Kaiming [37], are inversely proportional to the width of a layer [33]. With magnitude pruning the widest layers, the layers with largest input or output dimensions, are the <

> Figure 2:Where does layer-collapse occur Fraction of parameters remaining at each layer of a VGG-19 model pruned at initialization with ImageNet over a range of compression ratios (<>). A higher transparency represents a higher compression ratio. A dashed line indicates that there is at least one layer with no parameters, implying layer-collapse has occurred. first to be fully pruned. Gradient-based pruning algorithms SNIP [13] and GraSP [14] also prune layers at different rates, but it is less clear what the root cause for this preference is. In particular, both SNIP and GraSP aggressively prune the largest layer, the layer with the most trainable parameters, evident by the sharp peaks in Fig. 2. Based on this observation, we hypothesize that gradient-based scores averaged within a layer are inversely proportional to the layer size. We examine this hypothesis by constructing a theoretical framework grounded in flow networks. We first define a general class of gradient-based scores, prove a conservation law for these scores, and then use this law to prove that our hypothesis of inverse proportionality between layer size and average layer score holds exactly. A general class of gradient-based scores.Synaptic saliency is any score metric that can be expressed as the Hadamard product <> (1) whereRis a scalar loss function of the output of a feed-forward network parameterized by . When R is the training loss L, the resulting synaptic saliency metric is equivalent (modulo sign) to <>, the score metric used in Skeletonization [1], one of the first network pruning algorithms. The resulting metric is also closely related to <> the score used in SNIP [13], <> the score used in GraSP, and <> the score used in the pruning after training algorithm Taylor-FO [27]. This general class of score metrics, while not encompassing, exposes key properties of gradient-based scores used for pruning. The conservation of synaptic saliency.All synaptic saliency metrics respect two surprising conservation laws that hold at any initialization and step in training. Theorem 1.Neuron-wise Conservation of Synaptic Saliency.For a feedforward neural network with homogenous activation functions, <>, (e.g. ReLU, Leaky ReLU, linear), the sum of the synaptic saliency for the incoming parameters to a hidden neuron (Sin ) is equal to the sum of the synaptic saliency for the outgoing parameters from the hidden neuron (S_out). Proof.Consider the jth hidden neuron of a network with outgoing parameters out and incoming parameters P <>, such that <> and <>. The sum of the synaptic saliency for the outgoing parameters is <> (2) The sum of the synaptic saliency for the incoming parameters is <> (3) When is homogeneous, then <> <

> Figure 3: Total score in Neuron-wise conservation of score.Each dot represents a hidden unit from the feature-extractor of a VGG-19 model pruned at initialization with ImageNet. The location of each dot corresponds to the total score for the unit’s incoming and outgoing parameters, <>. The black dotted line represents exact neuron-wise conservation of score. <> Figure 4:Inverse relationship between layer size and average layer score.Each dot represents a layer from a VGG-19 model pruned at initialization with ImageNet. The location of each dot corresponds to the layer’s average score 4 and inverse number of elements. The black dotted line represents a perfect linear relationship. The neuron-wise conservation of synaptic saliency implies network conservation as well. Theorem 2.Network-wise Conservation of Synaptic Saliency.The sum of the synaptic saliency across any set of parameters that exactly 3 separates the input neurons x from the output neurons y of a feedforward neural network with homogenous activation functions equals <> We prove this theorem in Appendix 10 by applying the neuron-wise conservation law recursively. Similar conservation properties have been noted in the neural network interpretability literature and have motivated the construction of interpretability methods such as Conductance [38] and Layer-wise Relevance Propagation [39], which have recently been modified for network pruning [9,40]. While the interpretability literature has focused on attribution to the input pixels and hidden neuron activations, we have formulated conservation laws that are more general and applicable to any parameter and neuron in a network. Remarkably, these conservation laws of synaptic saliency apply to modern neural network architectures and a wide variety of neural network layers (e.g. dense, convolutional, batchnorm, pooling, residual) as visually demonstrated in Fig. 3. Conservation and single-shot pruning leads to layer-collapse.The conservation laws of synaptic saliency provide us with the theoretical tools to validate our earlier hypothesis of inverse proportionality between layer size and average layer score as a root cause for layer-collapse of gradient-based pruning methods. Consider the set of parameters in a layer of a simple, fully connected neural network. This set would exactly separate the input neurons from the output neurons. Thus, by the network-wise conservation of synaptic saliency (theorem 2), the total score for this set is constant for all layers, implying the average is inversely proportional to the layer size. We can empirically evaluate this relationship at scale for existing pruning methods by computing the total score for each layer of a model, as shown in Fig. 4. While this inverse relationship is exact for synaptic saliency, other closely related gradient-based scores, such as the scores used in SNIP and GraSP, also respect this relationship. This validates the empirical observation that for a given compression ratio, gradient-based pruning methods will disproportionately prune the largest layers. Thus, if the compression ratio is large enough and the pruning score is only evaluated once, then a gradient-based pruning method will completely prune the largest layer leading to layer-collapse. 3 Every element of the set is needed to separate the input neurons from the output neurons. 4 For GraSP we negated the average layer score so that we could plot on a log-log plot. 5 Magnitude pruning avoids layer-collapse with conservation and iteration Having demonstrated and investigated the cause of layer-collapse in single-shot pruning methods at initialization, we now explore an iterative pruning method that appears to avoid the issue entirely. Iterative Magnitude Pruning (IMP) is a recently proposed pruning algorithm that has proven to be successful in finding extremely sparse trainable neural networks at initialization (winning lottery tickets) [10,11,12,41,42,43,44]. The algorithm follows three simple steps. First train a network, second prune parameters with the smallest magnitude, third reset the unpruned parameters to their initialization and repeat until the desired compression ratio. While simple and powerful, IMP is impractical as it involves training the network several times, essentially defeating <

> the purpose of constructing a sparse initialization. That being said it does not suffer from the same catastrophic layer-collapse that other pruning at initialization methods are susceptible to. Thus, understanding better how IMP avoids layer-collapse might shed light on how to improve pruning at initialization. As has been noted previously [10,11], iteration is essential for stabilizing IMP. In fact, without sufficient pruning iterations, IMP will suffer from layer-collapse, evident in the sudden accuracy drops for the darker curves in Fig. 5a. However, the number of layer-collapse. Notice that if IMP didn’t train the network during each prune cycle, then, no matter the number of pruning iterations, it would be equivalent to single-shot magnitude pruning. Thus, something very critical must happen to the magnitude of the parameters during training, that when coupled with sufficient pruning iterations allows IMP to avoid layer-collapse. We hypothesize that gradient descent training effectively encourages the scores to observe an approximate layer-wise conservation law, which when coupled with sufficient pruning iterations allows IMP to avoid layer-collapse. Gradient descent encourages conservation. To better understand the dynamics of the IMP algorithm during training, we will consider a differentiable score <> algorithmically equivalent to the magnitude score. Consider these scores throughout training with gradient descent on a loss function L using an infinitesimal step size (i.e. gradient flow). In this setting, the temporal derivative of the parameters is equivalent to <>, and thus the temporal derivative of the score is <> (4) Surprisingly, this is a form of synaptic saliency and thus the neuron-wise and layer-wise conservation laws from Sec. 4 apply. In particular, this implies that for any two layers l and k of a simple, fully connected network, then <>. This invariance has been noticed before by [45] as a form of implicit regularization and used to explain the empirical phenomenon that trained multi-layer models can have similar layer-wise magnitudes. In the context of pruning, this phenomenon implies that gradient descent training, with a small enough learning rate, encourages the squared magnitude scores to converge to an approximate layer-wise conservation, as shown in Fig. 5b. Conservation and iterative pruning avoids layer-collapse.As explained in section 4, conservation alone leads to layer-collapse by assigning parameters in the largest layers with lower scores relative to parameters in smaller layers. However, if conservation is coupled with iterative pruning, then when the largest layer is pruned, becoming smaller, then in subsequent iterations the remaining parameters of this layer will be assigned higher relative scores. With sufficient iterations, conservation coupled with iteration leads to a self-balancing pruning strategy allowing IMP to avoid layer-collapse. This insight on the importance of conservation and iteration applies more broadly to other algorithms with exact or approximate conservation properties (e.g. Skeletonization, SNIP, and GraSP as demonstrated in Sec. 3). Indeed, very recent work empirically confirms that iteration improves the performance of SNIP [46]. 6 A data-agnostic algorithm satisfying Maximal Critical Compression In the previous section we identified two key ingredients of IMP’s ability to avoid layer-collapse: (i) approximate layer-wise conservation of the pruning scores, and (ii) the iterative re-evaluation of these scores. While these properties allow the IMP algorithm to identify high performing and highly sparse, trainable neural networks, it requires an impractical amount of computation to obtain them. Thus, we aim to construct a more efficient pruning algorithm while still inheriting the key aspects of IMP’s success. So what are the essential ingredients for a pruning algorithm to avoid layer-collapse and provably attain Maximal Critical Compression? We prove the following theorem in Appendix 10. Theorem 3.Iterative, positive, conservative scoring achieves Maximal Critical Compression.If a pruning algorithm, with global-masking, assigns positive scores that respect layer-wise conservation and if the algorithm re-evaluates the scores every time a parameter is pruned, then the algorithm satisfies the Maximal Critical Compression axiom. The Iterative Synaptic Flow Pruning (SynFlow) algorithm Theorem 3 directly motivates the design of our novel pruning algorithm, SynFlow, that provably reaches Maximal Critical Compression. First, the necessity for iterative score evaluation discourages algorithms that involve backpropagation on batches of data, and instead motivates the development of an efficient data-independent scoring procedure. Second, positivity and conservation motives the construction of a loss function that yields positive synaptic saliency scores. We combine these insights to introduce a new loss function (where 1 is the all ones vector and <> the element-wise absolute value of parameters in the lth layer), <> (5) that yields the positive, synaptic saliency scores ( @RSF ) we term Synaptic Flow. For a simple, fully connected network (i.e. <>), we can factor the Synaptic Flow score for a parameter <> as <> <> (6) This perspective demonstrates that Synaptic Flow score is a generalization of magnitude score (jw[l] j), where ij the scores consider the product of synaptic strengths flowing through each parameter, taking the inter-layer interactions of parameters into account. We use the Synaptic Flow score in the Iterative Synaptic Flow Pruning (SynFlow) algorithm summarized in the pseudocode below. Algorithm 1:Iterative Synaptic Flow Pruning (SynFlow). <> Given a network <> and specified compression ratio , the SynFlow algorithm requires only one additional hyperparameter, the number of pruning iterations n. We demonstrate in Appendix 11, that an exponential pruning schedule <> with n=100 pruning iterations essentially prevents layer-collapse whenever avoidable (Fig. 1), while remaining computationally feasible, even for large networks. 7 Experiments We empirically benchmark the performance of our algorithm, SynFlow (red), against the baselines random pruning and magnitude pruning, as well as the state-of-the-art algorithms SNIP [13] and GraSP [14]. In Fig. 6, we test the five algorithms on 12 distinct combinations of modern architectures (VGG-11, VGG-16, ResNet- 18, WideResNet-18) and datasets (CIFAR-10, CIFAR-100, Tiny ImageNet) over an exponential sweep of compression ratios (<>). See Appendix 12 for more details and hyperparameters of the experiments. Consistently, SynFlow outperforms the other algorithms in the high compression regime (10 1:5 < ) and demonstrates significantly more stability, as indicated by its tight intervals. Furthermore, SynFlow is the only algorithm that reliably shows better performance to the random pruning baseline: SNIP and GraSP perform significantly worse than random pruning with ResNet-18 and WideResNet-18 trained on Tiny ImageNet. SynFlow is also quite competitive in the low compression regime (<>). Although magnitude pruning can partially outperform SynFlow in this regime with models trained on Tiny ImageNet, it suffers from catastrophic layer-collapse as indicated by the sharp drops in accuracy. <> Figure 6:SynFlow consistently outperforms other pruning methods.Top-1 test accuracy as a function of different compression ratios over 12 distinct combinations of models and datasets. We performed three runs with the same hyperparameter conditions and different random seeds. The solid line represents the mean, the shaded region represents area between minimum and maximum performance of the three runs. 8 Conclusion In this paper, we developed a unifying theoretical framework that explains why existing single-shot pruning algorithms at initialization suffer from layer-collapse. We applied our framework to elucidate how iterative magnitude pruning [10] overcomes layer-collapse to identify winning lottery tickets at initialization. Building on the theory, we designed a new data-agnostic pruning algorithm, SynFlow, that provably avoids layer-collapse and reaches Maximal Critical Compression. Finally, we empirically confirmed that our SynFlow algorithm consistently performs better than existing algorithms across 12 distinct combinations of models and datasets, despite the fact that our algorithm is data-agnostic and requires no pre-training. Promising future directions for this work are to (i) explore a larger space of potential pruning algorithms that satisfy Maximal Critical Compression, (ii) harness SynFlow as an efficient way to compute appropriate per-layer compression ratios to combine with existing scoring metrics, and (iii) incorporate pruning as a part of neural network initialization schemes. Overall, our data-agnostic pruning algorithm challenges the existing paradigm that data must be used to quantify which synapses of a neural network are important. 9 Acknowledgements We thank Jonathan M. Bloom, Weihua Hu, Javier Sagastuy-Brena, Chengxu Zhuang, and members of the Stanford Neuroscience and Artificial Intelligence Laboratory for helpful discussions. We thank the Stanford Data Science Scholars program (DK), the Burroughs Welcome, Simons and James S. McDonnell foundations, and an NSF career award (SG) for support. References [1]Michael C Mozer and Paul Smolensky. Skeletonization: A technique for trimming the fat from a network via relevance assessment. InAdvances in neural information processing systems, pages 107–115, 1989. [2]Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. InAdvances in neural information processing systems, pages 598–605, 1990. [3]Babak Hassibi and David G Stork. Second order derivatives for network pruning: Optimal brain surgeon. InAdvances in neural information processing systems, pages 164–171, 1993. [4]Steven A Janowsky. Pruning versus clipping in neural networks.Physical Review A, 39(12):6600, 1989. [5]Russell Reed. Pruning algorithms-a survey.IEEE transactions on Neural Networks, 4(5):740–747, 1993. [6]Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. InAdvances in neural information processing systems, pages 1135–1143, 2015. [7]Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. Efficient processing of deep neural networks: A tutorial and survey.Proceedings of the IEEE, 105(12):2295–2329, 2017. [8]Sanjeev Arora, Rong Ge, Behnam Neyshabur, and Yi Zhang. Stronger generalization bounds for deep nets via a compression approach. In Jennifer Dy and Andreas Krause, editors,Proceedings of the 35th International Conference on Machine Learning, volume 80 ofProceedings of Machine Learning Research, pages 254–263. PMLR, 2018. [9]Hidenori Tanaka, Aran Nayebi, Niru Maheswaranathan, Lane McIntosh, Stephen Baccus, and Surya Ganguli. From deep learning to mechanistic understanding in neuroscience: the structure of retinal prediction. InAdvances in Neural Information Processing Systems, pages 8535–8545, 2019. [10]Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. InInternational Conference on Learning Representations, 2019. [11]Jonathan Frankle, G Karolina Dziugaite, DM Roy, and M Carbin. Stabilizing the lottery ticket hypothesis. arXiv, page. [12] Ari Morcos, Haonan Yu, Michela Paganini, and Yuandong Tian. One ticket to win them all: generalizing lottery ticket initializations across datasets and optimizers. InAdvances in Neural Information Processing Systems, pages 4933–4943, 2019. [13]Namhoon Lee, Thalaiyasingam Ajanthan, and Philip Torr. SNIP: SINGLE-SHOT NETWORK PRUNING BASED ON CONNECTION SENSITIVITY. InInternational Conference on Learning Representations, 2019. [14]Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. InInternational Conference on Learning Representations, 2020. [15]Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and< 0.5 mb model size.arXiv preprint arXiv:1602.07360, 2016. [16]Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications.arXiv preprint arXiv:1704.04861, 2017. [17]Ameya Prabhu, Girish Varma, and Anoop Namboodiri. Deep expander networks: Efficient deep networks from graph theory. InProceedings of the European Conference on Computer Vision (ECCV), pages 20–35, 2018. [18]Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. InProceedings of the British Machine Vision Conference. BMVA Press, 2014. doi: http://dx.doi.org/10.5244/C.28.88. [19]Alexander Novikov, Dmitrii Podoprikhin, Anton Osokin, and Dmitry P Vetrov. Tensorizing neural networks. InAdvances in neural information processing systems, pages 442–450, 2015. [20]Guillaume Bellec, David Kappel, Wolfgang Maass, and Robert Legenstein. Deep rewiring: Training very sparse deep networks. InInternational Conference on Learning Representations, 2018. [21]Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H Nguyen, Madeleine Gibescu, and Antonio Liotta. Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science.Nature communications, 9(1):1–12, 2018. [22]Trevor Gale, Erich Elsen, and Sara Hooker. The state of sparsity in deep neural networks.arXiv preprint arXiv:1902.09574, 2019. [23]Davis Blalock, Jose Javier Gonzalez Ortiz, Jonathan Frankle, and John Guttag. What is the state of neural network pruning?arXiv preprint arXiv:2003.03033, 2020. [24]Sejun Park*, Jaeho Lee*, Sangwoo Mo, and Jinwoo Shin. Lookahead: A far-sighted alternative of magnitude-based pruning. InInternational Conference on Learning Representations, 2020. [25]Ehud D Karnin. A simple procedure for pruning back-propagation trained neural networks. IEEE transactions on neural networks, 1(2):239–242, 1990. [26]Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference.arXiv preprint arXiv:1611.06440, 2016. [27]Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11264–11272, 2019. [28]Yiwen Guo, Anbang Yao, and Yurong Chen. Dynamic network surgery for efficient dnns. InAdvances in neural information processing systems, pages 1379–1387, 2016. [29]Xin Dong, Shangyu Chen, and Sinno Pan. Learning to prune deep neural networks via layer-wise optimal brain surgeon. InAdvances in Neural Information Processing Systems, pages 4857–4867, 2017. [30]Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching- Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9194–9203, 2018. [31]Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. InInternational Conference on Learning Representations, 2019. [32]Namhoon Lee, Thalaiyasingam Ajanthan, Stephen Gould, and Philip H. S. Torr. A signal propaga- tion perspective for pruning neural networks at initialization. InInternational Conference on Learning Representations, 2020. [33]Haoran You, Chaojian Li, Pengfei Xu, Yonggan Fu, Yue Wang, Xiaohan Chen, Richard G. Baraniuk, Zhangyang Wang, and Yingyan Lin. Drawing early-bird tickets: Toward more efficient training of deep networks. InInternational Conference on Learning Representations, 2020. [34]Wenyuan Zeng and Raquel Urtasun. Mlprune: Multi-layer pruning for automated neural network compres- sion. 2018. [35]Hesham Mostafa and Xin Wang. Parameter efficient training of deep convolutional neural networks by dynamic sparse reparameterization. InProceedings of the 36th International Conference on Machine Learning, volume 97 ofProceedings of Machine Learning Research, pages 4646–4655. PMLR, 2019. [36]Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. InProceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256, 2010. [37]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. [38]Kedar Dhamdhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron. InInternational Conference on Learning Representations, 2019. [39]Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.PloS one, 10(7), 2015. [40]Seul-Ki Yeom, Philipp Seegerer, Sebastian Lapuschkin, Simon Wiedemann, Klaus-Robert Müller, and Wojciech Samek. Pruning by explaining: A novel criterion for deep neural network pruning.arXiv preprint arXiv:1912.08881, 2019. [41]Hattie Zhou, Janice Lan, Rosanne Liu, and Jason Yosinski. Deconstructing lottery tickets: Zeros, signs, and the supermask. InAdvances in Neural Information Processing Systems, pages 3592–3602, 2019. [42]Haoran You, Chaojian Li, Pengfei Xu, Yonggan Fu, Yue Wang, Xiaohan Chen, Richard G. Baraniuk, Zhangyang Wang, and Yingyan Lin. Drawing early-bird tickets: Toward more efficient training of deep networks. InInternational Conference on Learning Representations, 2020. [43]Jonathan Frankle, Gintare Karolina Dziugaite, Daniel M Roy, and Michael Carbin. Linear mode connectiv- ity and the lottery ticket hypothesis.arXiv preprint arXiv:1912.05671, 2019. [44]Haonan Yu, Sergey Edunov, Yuandong Tian, and Ari S. Morcos. Playing the lottery with rewards and multiple languages: lottery tickets in rl and nlp. InInternational Conference on Learning Representations, 2020. [45]Simon S Du, Wei Hu, and Jason D Lee. Algorithmic regularization in learning deep homogeneous models: Layers are automatically balanced. InAdvances in Neural Information Processing Systems, pages 384–395, 2018. [46]Stijn Verdenius, Maarten Stol, and Patrick Forré. Pruning via iterative ranking of sensitivity statistics. arXiv preprint arXiv:2006.00896, 2020. Appendix 10 Proofs We provide a proof for Theorem 2 which we rewrite below. Theorem 2.Network-wise Conservation of Synaptic Saliency.The sum of the synaptic saliency across any set of parameters that exactly separates the input neuronsxfrom the output neuronsyof a feedforward neural network with homogenous activation functions equals <> Proof.We begin by defining the set of neurons (V) and the set of prunable parameters (E) for a neural network. Consider a subset of the neurons <>, such that all output neuronsyc 2Sand all input neuronsxi 2VnS. Consider the set of parameters cut by this partition <> (7) By theorem 1, we know that that sum of the synaptic saliency over C(S) is equal to the sum of the synaptic saliency over the set of parameters adjacent toC(S)and between neurons in <>. Continuing this argument, then eventually we get that this sum must be equal to the sum of the synaptic saliency over the set of parameters incident to the output neuronsy, which is <> (8) We can repeat this argument iterating through the setVnStill we reach the input neuronsxto show that this sum is also equal to <> We provide a proof for Theorem 3 which we rewrite below. Theorem 3. Iterative, positive, conservative scoring achieves Maximal Critical Compression.If a pruning algorithm, with global-masking, assigns positive scores that respect layer-wise conservation and if the algorithm re-evaluates the scores every time a parameter is pruned, then the algorithm satisfies the Maximal Critical Compression axiom. Proof.We prove this theorem by contradiction. Assume that a pruning algorithm with global-masking and iterative, positive, conservative scoring does not satisfy the Maximal Critical Compression axiom. This implies that at some iteration, the algorithm will prune the last parameter in a layer (layer l), despite there existing more than one parameters <> in another layer (layer k). Because the algorithm uses global-masking, then the score for the last parameter in layer l,S[l] , is less than or equal to the scores for each parameter, S[k], in layer k <> (9) P Because the scores respect a layer-wise conservation, then S[l] = N[k] S[k]. This implies, by the positivity of i=1 the scores and because N[k]>1, that for all i, <> (10) This is a contradiction to the previous inequality. 11 Hyperparameters choices for the SynFlow algorithm Theorem 3 required that an algorithm re-evaluates the scores every time a parameter is pruned. However, theorem 2 provides a theoretical insight to drastically reduce the number of iterations needed to practically attain Maximal Critical Compression. We now introduce a modification to theorem 3 that motivates practical hyperparameter choices used in the SynFlow algorithm. Theorem 4.Achieving Maximal Critical Compression practically. If a pruning algorithm, with global- masking, assigns positive scores that respect layer-wise conservation and if the prune size, the total score for the parameters pruned at any iteration, is strictly less than the cut size, the total score for an entire layer, whenever possible, then the algorithm satisfies the Maximal Critical Compression axiom. Proof.We prove this theorem by contradiction. Assume there is an iterative pruning algorithm that uses positive, layer-wise conserved scores and maintains that the prune size at any iteration is less than the cut size whenever possible, but doesn’t satisfy the Maximal Critical Compression axiom. At some iteration the algorithm will prune a set of parameters containing a subset separating the input neurons from the output neurons, despite there existing a set of the same cardinality that does not lead to layer-collapse. By theorem 2, the total score for the separating subset is <>, which implies by the positivity of the scores, that the total prune size is at @y least <>. This contradicts the assumption that the algorithm maintains that the prune size at any iteration is @y always strictly less than the cut size whenever possible. Motivated by Theorem 4, we can now choose a practical, yet effective, number of pruning iteration (n) and schedule for the compression ratios <> applied at each iteration (k) for the SynFlow algorithm. Two natural candidates for a compression schedule would be either linear <> or exponential <>. Empirically we find that the SynFlow algorithm with 100 pruning iterations and an exponential compression schedule satisfies the conditions of theorem 4 over a reasonable range of compression ratios <>, as shown in Fig. 7b. This is not true if we use a linear schedule for the compression ratios, as shown in Fig. 7a. Interestingly, Iterative Magnitude Pruning also uses an exponential compression schedule, but does not provide a thorough explanation for this hyperparameter choice [10]. <

> Figure 7:Choosing the number of pruning iterations and compression schedule for SynFlow.Maximum ratio of prune size with cut size for increasing number of pruning iterations for SynFlow with a linear (left) or exponential (right) compression schedule. Higher transparency represents higher compression ratios. The black dotted line represents the maximal prune size ratio that can be obtained while still satisfying the conditions of theorem 4. All data is from a VGG-19 model at initialization using ImageNet. Potential numerical instability. The SynFlow algorithm involves computing the SynFlow objective, <>, whose singular values may vanish or explode exponentially with depthL. This may lead to l=1 potential numerical instability for very deep networks, although we did not observe this for the models presented in this paper. One way to address this potential challenge would be to appropriately scale network parameters at each layer to maintain stability. Because the SynFlow algorithm is scale invariant at each layer <>, this modification will not effect the performance of the algorithm. 12 Experimental details An open source version of our code and the data used to generate all the figures in this paper are available at github.com/ganguli-lab/Synaptic-Flow. 12.1 Pruning algorithms All pruning algorithms we considered in our experiments use the following two steps: (i) scoring parameters, and (ii) masking parameters globally across the network with the lowest scores. Here we describe details of how we computed scores used in each of the pruning algorithms. Random:We sampled independently from a standard Gaussian. Magnitude:We computed the absolute value of the parameters. SNIP:We computed the score <> using a random subset of the training dataset with a size ten times the number of classes, namely 100 for CIFAR-10, 1000 for CIFAR-100,2000 for Tiny ImageNet, and 10000 for ImageNet. The score was computed on a batch of size 256 for CIFAR-10/100, 64 for Tiny ImageNet, and 16 for ImageNet, then summed across batches to obtain the score used for pruning. GraSP: We computed the score <> using a random subset of the training dataset with a size ten times the number of classes, namely 100 for CIFAR-10,1000 for CIFAR-100,2000 for Tiny ImageNet, and 10000for ImageNet. The score was computed on a batch of size 256 for CIFAR-10/100, 64 for Tiny ImageNet, and 16 for ImageNet, then summed across batches to obtain the score used for pruning. SynFlow:We applied the pseudocode 1 with 100 pruning iterations motivated by the theoretical and empirical results discussed in Sec 11. 12.2 Model architectures We adapted standard implementations of VGG-11 and VGG-16 from OpenLTH, and ResNet-18 and WideResNet- 18 from PyTorch models. We considered all weights from convolutional and linear layers of these models as prunable parameters, but did not prune biases nor the parameters involved in batchnorm layers. For convolutional and linear layers, the weights were initialized with a Kaiming normal strategy and biases to be zero. 12.3 Training hyperparameters Here we provide hyperparameters that we used to train the models presented in Fig. 1 and Fig. 6. These hyperparameters were chosen for the performance of the original model and were not optimized for the performance of the pruned networks. <

> <> <> <>