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 arXiv:1608.03665v4 [cs.NE] 18 Oct 2016 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 5.1and 3.1speedups of convolutional layer computation ofAlexNetagainst 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 originalResNetwith 32 layers. ForAlexNet, structure regularization by SSL also reduces the error by1%. 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 re- markable 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 impactspracticalacceleration in hardware platforms. Figure 1 depicts practical speedup of each layer of aAlexNet, which is non-structurally sparsified by‘1 -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 Speedup 1.5 1 Quadro K600 1 Sparsity Tesla K40c GTX Titan 0.5 Sparsity 0 0 conv1 conv2 conv3 conv4 conv5 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 todirectlylearn a compressed structure of deep CNNs by group Lasso regularization during the training. SSL is a generic regularization to adaptively adjust mutiple 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.5.1on CPU and 3.1on GPU forAlexNet). 2 Related works Connection pruning and weight sparsifying. Hanet al.[7][8] reduced number of parameters of AlexNetby 9andVGG-16by 13using connection pruning. Since most reduction is achieved on fully-connected layers, the authors obtained 3to 4layer-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 ofResNet-152in [5] are from fully-connected layers, compression and acceleration on convolutional layers become essential. Liuet al.[6] achieved >90% sparsity of convolutional layers inAlexNetwith 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 4.5speedup on CPUs for scene text character recognition and Dentonet al.[10] achieved 2speedups on both CPUs and GPUs for the first two layers. Both of the works usedLow Rank Approximation(LRA) with1% 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 havelower 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 2 W(l) (1) nl ,:,:,: channel-wise W(l) (2) shortcut :,c l ,:,: W(l) (3) :,c l ,m l ,k l W(l) (4) W(l) (1) nl ,:,:,: W(l) (1) nl ,:,:,: … W(l) (2) :,c l ,:,: W(l) (2) shape-wise :,c l ,:,: W(l) (3) W(l) (3) :,c l ,m l ,k l :,c filter-wise W(l) (1) l ,m l ,k l depth-wise W(l) (4) nl ,:,:,: W(l) (4) Figure 2: The proposed structured sparsity learning (SSL) for DNNs. Weights in filters are split W(l) (2) 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 W(l) (3) structured sparsity that were explored in the work. :,c l ,m l ,k l W(l) (4) 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 athttps://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 1 Suppose weights of convolutional layers in a DNN form a sequence of 4-D tensors 1 W(l) 2 RNl Cl Ml Kl , whereNl ,Cl ,Ml andKl are the dimensions of thel-th(1lL) weight tensor along the axes of filter, channel, spatial height and spatial width, respectively.Ldenotes the number of convolutional layers. Then the proposed generic optimization target of a DNN with structured sparsity regularization can be formulated as: 1 XL E(W) =ED (W) +R(W) +g Rg W(l) : (1) l=1 HereWrepresents 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.,‘2 -norm; andRg ()is the structured sparsity regularization on each layer. BecauseGroup Lassocan 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 weightsPwcan be represented asRg (w) = G jjw(g) jj g=1 g , wherew(g) is a group of partial weights inw andGis the total number of groups. Different groups may overlap. Herer jj jj g is the group Lasso, or P 2 jjw(g) jj g)g = jw(g) j w( , wherejw(g) jis the number of weights inw(g) .i=1 i 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, anddepth-wisestructured sparsity in Figure 2. For simplicity, theR()term of Eq. (1) is omitted in the following formulation expressions. Penalizing unimportant filers and channels. SupposeW(l) (l)nl ;:;:;: is thenl -thfilter andW:;c l ;:;: is the cl -thchannel of all filters in thel-thlayer. The optimization target of learning the filter-wise and 3 channel-wise structured sparsity can be defined as 0 1 0 1 XL XNl XL XCl E(W) =ED (W) +n @ jjW(l) jjA (2) nl ;:;:;:g +c @ jjW(l) jjA:;c l ;:;:g : l=1 nl =1 l=1 cl =1 As indicated in Eq. (2), our approach tends to remove less important filters and channels. Note that zeroing out a filter in thel-thlayer results in a dummy zero output feature map, which in turn makes a corresponding channel in the(l+ 1)-thlayer 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,W(l) denotes the vector of :;c l ;m l ;k all corresponding weights located at spatial position of(m l l ;k l )in the 2D filters across thecl -th channel. Thus, we defineW(l) as theshape fiberrelated to learning arbitrary filter shape :;c l ;m l ;k 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: 0 1 XL XCl XMl XKl E(W) =ED (W) +s @ jjW(l) jjA:;c l ;m l ;k g : (3) l l=1 cl =1ml =1kl =1 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 isPE(W) =E )D (W)+d L jjW(l jj l=1 g . 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-wisesparsity, 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 ofGEneral Matrix Multiplication(GEMM) by lowering weight tensors and feature tensors to matrices [18]. For example, in Caffe [19], a 3D filterW(l)nl ;:;:;: is reshaped to a row in the weight matrix where each column is the collection of weightsW(l) :;c related to shape-wise sparsity. Combining filter-wise and shape-wise sparsity can directly reduce the l ;m l ;k l dimension of weight matrix in GEMM by removing zero rows and columns. In this context, we use row-wiseandcolumn-wisesparsity as the interchangeable terminology offilter-wiseandshape-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 . 4 Table 1: Results after penalizing unimportant filters and channels inLeNet LeNet# Error Filter # § Channel # § FLOP § Speedup § 1 (baseline) 0.9% 20—50 1—20 100%—100% 1.00—1.00 2 0.8% 5—19 1—4 25%—7.6% 1.64—5.23 3 1.0% 3—12 1—3 15%—3.6% 1.99—7.44 § In the order ofconv1—conv2 Table 2: Results after learning filter shapes inLeNet LeNet# Error Filter size § Channel # FLOP Speedup 1 (baseline) 0.9% 25—500 1—20 100%—100% 1.00—1.00 4 0.8% 21—41 1—2 8.4%—8.2% 2.33—6.93 5 1.0% 7—14 1—1 1.4%—2.8% 5.19—10.82 § The sizes of filters after removing zero shape fibers, in the order ofconv1—conv2 4.1 LeNetand 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 learnedconv1filters 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 onLeNetis summarized in Table 2. The baselineLeNet 1 hasconv1filters with a regular55square (size = 25) whileLeNet 5reduces the dimension that can be constrained by a24rectangle (size = 7). The 3D shape ofconv2filters in the baseline is also regularized to the 2D shape inLeNet 5within only one channel, indicating that only one filter in conv1is needed. This fact significantly saves FLOP and computation time. Figure 3: Learnedconv1filters inLeNet 1(top),LeNet 2(middle) andLeNet 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: 5 Table 2: Results after learning filter shapes inLeNet LeNet# Error Filter size § Channel # FLOP Speedup 1(baseline) 0.9% 25–500 1–20 100%–100% 1.00⇥–1.00⇥ 4 0.8% 21–41 1–2 8.4%–8.2% 2.33⇥–6.93⇥ 5 1.0% 7–14 1–1 1.4%–2.8% 5.19⇥–10.82⇥ § The sizes of filters after removing zero shape fibers, in the order ofconv1–conv2 50 50 50 % Reconstruction error conv1 conv1 conv1 40 conv2 40 conv2 40 conv2 conv3 conv3 30 30 30 conv4 conv5 20 20 20 10 10 10 0 0 0 0 50 100 0 50 100 0 50 100 % ranks 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.% ranksof eigenvec- tors 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:AlexNet4 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 ofconv1filters, 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 inconv1is 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 MLP# Error Neuron # per layer § FLOP per layer § 1 291 1(baseline) 1.43% 784–500–300–10 100%–100%–100% 2 1.34% 469–294–166–10 35.18%–32.54%–55.33% 3 1.53% 434–174–78–10 19.26%–9.05%–26.00% § In the order ofinput layer–hidden layer 1–hidden layer 2–output layer 281 028 (a) (b) Figure 4: (a) Results of learning the number of neurons inMLP. (b) the connection numbers of input neurons (i.e.pixels) inMLP 2after SSL. 6 Table 3: Learning row-wise and column-wise sparsity ofConvNeton CIFAR-10 ConvNet # Error Row sparsity § Column sparsity § Speedup § 1 (baseline) 17.9% 12.5%–0%–0% 0%–0%–0% 1.00–1.00–1.00 2 17.9% 50.0%–28.1%–1.6% 0%–59.3%–35.1% 1.43–3.05–1.57 3 16.9% 31.3%–0%–1.6% 0%–42.8%–9.8% 1.25–2.01–1.18 § in the order ofconv1–conv2–conv3 handwriting digits are usually written in the center and pixels close to the boundary contain little discriminative classification information. 4.2 ConvNetandResNeton CIFAR-10 We implemented theConvNetof [1] anddeep 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 ofResNetto 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 learnedconv1filters. 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, fourconv1filters of the baseline are actually all-zeros as shown in Figure 5, demonstrating the great potential of filter sparsity. When SSL is applied, half ofconv1filters 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: Learnedconv1filters 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 shortcutsare utilized to connect the feature maps with the same dimension while 11 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 ofResNet-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 6 10 SSL ResNet−20 ResNet−32 % error 9 8 7 12 14 16 18 20SSL−ResNet−# 10 20 SSL ResNet−20 ResNet−32 18 32×32 16×16 8×8 # conv layers16 % error 9 141210 8 8642 7 0 12 14 16 18 20 12 14 16 18 20 SSL−ResNet−# SSL−ResNet−# 20 18 Figure 6: Error vs. layer number after depth regularization by SSL. 32×32 16×16 8×8 # conv layers 16 ResNet-#is the originalResNet in [ 1412 5] with # layers.SSL-ResNet-#is the depth-regularizedResNetby SSL with # layers, including the last fully-connected layer. 32 108 32 indicates the convolutional layers with an output map size of 3264 32, and so forth. 20 with all zero weights are removed and the net is finally fine-tuned with a base learning rate of 0.01, 12 14 16 18 20SSL−ResNet−# which is lower than that (i.e., 0.1) in the baseline. Figure 6 plots the trend of the error vs. the number of layers under different strengths of depth regularizations. Compared with originalResNetin [5], SSL learns aResNetwith 14 layers (SSL- ResNet-14) that reaching a lower error than the one of the baseline with 20 layers (ResNet-20); SSL-ResNet-18andResNet-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 ofSSL-ResNet-20from SSL-ResNet-18shows the suboptimal selection of the depth in the group of “3232”. 4.3 AlexNeton ImageNet To show the generalization of our method to large scale DNNs, we evaluate SSL usingAlexNetwith ILSVRC 2012.CaffeNet[19] – the replication ofAlexNet[1] with mirror changes, is used in our experiment. All training images are rescaled to the size of 256256. A 227227 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,AlexNetis 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 com- putation 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 % Sparsity % FLOP reduction 100 100 50 6 speedup % Reconstruction error conv1 conv1 l1 SSL 80 conv2 80 40 conv2 5 conv3 conv3 4 60 conv4 60 30 conv4 conv5 40 FLOP 40 20 conv5 3 2 20 20 10 1 0 0 0 41.5 42 42.5 43 43.5 44 0 50 100 0 % top-1 error % dimensionality Quadro Tesla Titan Xeon Xeon Xeon Xeon Black T8 T4 T2 T1 (a) (b) (c) Figure 7: (a) 2D-filter-wise sparsity and FLOP reduction vs. top-1 error. Vertical dash line shows the error of originalAlexNet; (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 theAlexNet 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 1andAlexNet 2in Table 4 are used as testbenches. 7 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 ofAlexNetby1% 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 1.4layer-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-regularizedAlexNet. The results show that the smoothness restric- tion enforces parameter searching in lower-dimensional space and enableslowerrank 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 aAlexNetwith 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 ofAlexNet 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 5.1and 3.1layer-wise acceleration on CPU and GPU, respectively. Instead,‘1 -norm achieves on average only 3.0and 0.9layer-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 ofAlexNeton ILSVRC 2012 # Method Top1 err. Statistics conv1 conv2 conv3 conv4 conv5 sparsity 67.6% 92.4% 97.2% 96.6% 94.3% 1 ‘1 44.67% CPU 0.80 2.91 4.84 3.83 2.76 GPU 0.25 0.52 1.38 1.04 1.36 column sparsity 0.0% 63.2% 76.9% 84.7% 80.7% row sparsity 9.4% 12.9% 40.6% 46.9% 0.0%2 SSL 44.66% CPU 1.05 3.37 6.27 9.73 4.93 GPU 1.00 2.37 4.94 4.03 3.05 3 pruning[7] 42.80% sparsity 16.0% 62.0% 65.0% 63.0% 63.0% sparsity 14.7% 76.2% 85.3% 81.5% 76.3% 4 ‘1 42.51% CPU 0.34 0.99 1.30 1.10 0.93 GPU 0.08 0.17 0.42 0.30 0.32 column sparsity 0.00% 20.9% 39.7% 39.7% 24.6% 5 SSL 42.53% CPU 1.00 1.27 1.64 1.68 1.32 GPU 1.00 1.25 1.63 1.72 1.36 8 and Titan) and CPU (Intel Xeon E5-2630). SSL can achieve on average3speedup 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 averagely2speedup 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. 9 [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. 10