Identity Mappings in Deep Residual Networks arXiv:1603.05027v3 [cs.CV] 25 Jul 2016 Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun Microsoft Research AbstractDeep 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: yl =h(xl ) +F(xl ;Wl ); xl+1 =f(yl ); wherexl andxl+1 are input and output of thel-th unit, andFis a residual function. In [1],h(xl ) =xl is an identity mapping andfis 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 toh(xl ), with a key choice of using an identity mappingh(xl ) =xl . 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 thatif bothh(xl )and f(yl )are identity mappings, the signal could bedirectlypropagated 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 ofh(xl ). We nd that the identity mappingh(xl ) =xl chosen in [1] 2 2 20 x x ResNet−1001, original (error: 7.61%)l l ResNet−1001, proposed (error: 4.92%) weight BN 15 BN ReLU 0.2 ReLU weight Test Error (% s Training Los 10 weight BN BN ReLU ) 0.02 additton weight 5 ReLU additton xl+1 xl+1 0.002 0 (a) original (b) proposed 0 1 2 3 4 5 6 Iterations x 10 4 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 11 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 mappingf(yl ) =yl , 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 overt. 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] aremodularizedarchitectures 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: yl =h(xl ) +F(xl ;Wl ); (1) xl+1 =f(yl ): (2) Herexl is the input feature to thel-th Residual Unit.Wl =fWl;k j1kK gis a set of weights (and biases) associated with thel-th Residual Unit, andKis the number of layers in a Residual Unit (Kis 2 or 3 in [1]).Fdenotes the residual function,e.g., a stack of two 33 convolutional layers in [1]. The functionfis the operation after element-wise addition, and in [1]fis ReLU. The functionh is set as an identity mapping:h(xl ) =xl .1 Iffis also an identity mapping:xl+1 yl , we can put Eqn.(2) into Eqn.(1) and obtain: xl+1 =xl +F(xl ;Wl ): (3) Recursively ( xl+2 =xl+1 +F(xl+1 ;Wl+1 ) =xl +F(xl ;Wl ) +F(xl+1 ;Wl+1 ) , etc.) we will have: L X1 xL =xl + F(xi ;Wi ); (4) i=l forany deeper unitLandany shallower unitl. Eqn.(4) exhibits some nice properties.(i)The featurexL of any deeper unitLcan be represented as thePfeaturexl of any shallower unitlplus a residual function in a form of L1 F,i=l indicating that the model is in aresidualfashion between any unitsLandl.(ii)PThe featurexL =x0 + L1 F(x Wi=0 i ; i ), of any deep unitL, is thesummation of the outputs of all preceding residual functions (plusx0 ). This is in contrast to Qa \plain network" where a featurexL is a series of matrix-vectorproducts, say, L1 Wi=0 i x0 (ignoring BN and ReLU). Eqn.(4) also leads to nice backward propagation properties. Denoting the loss function asE, from the chain rule of backpropagation [9] we have: ! @E @E @x @E @ L X1 = L = 1 + F(x@xl @xL @x i ;Wi ) : (5) l @xL @xli=l Eqn.(5) indicates that the gradient @E can be decomposed into two additive @x terms: a term of l @E that propagates information directly without concern- @xL ing any weight layers, and another term of P@E @ L1 F that propagates @xL @xl i=l through the weight layers. The additive term of @E ensures that information is @x directly propagated back toany shallower unitl. Eqn.( L 5) also suggests that it 1 It is noteworthy that there are Residual Units for increasing dimensions and reducing feature map sizes [1] in whichhis 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. 4 is unlikely for the gradient @E to be canceled out for a mini-batch, because in @x general the term P l @ L1 F cannot be always -1 for all samples in a mini-batch. @xl i=l This implies that the gradient of a layer does not vanish even when the weights are arbitrarily small. 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 connectionh(xl ) =xl , and (ii) the condition thatfis an identity mapping. These directly propagated information ows 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 modication,h(xl ) =l xl , to break the identity shortcut: xl+1 =l xl +F(xl ;Wl ); (6) wherel is a modulating scalar (for simplicity we still assumefis identity). Recursively applying this formulation we obtain an equation similar to Eqn. (4):Q P Qx 1L = ( L1 ( L1 i=l i )xl + L i=l j=i+1 j )F(xi ;Wi ), or simply: L Y1 L X1 xL = ( i )xl + F^(xi ;Wi ); (7) i=l i=l where the notationF^absorbs the scalars into the residual functions. Similar to Eqn.(5), we have backpropagation of the following form: ! @E @E L Y1 @ L X1 = ( @x i ) + F^(xi ;Wi ) : (8) l @xL @xli=l i=l Unlike Eqn.(5), in Eqn.(8) the rst additive term is modulated by a factorQL1 1 for alli, this i=l i . For an extremely deep network (Lis large), ifi > factor can be exponentially large; ifi <1 for alli, this factor can be expo- nentially small and vanish, which blocks the backpropagated signal from the shortcut and forces it to ow through the weight layers. This results in opti- mization diculties as we show by experiments. In the above analysis, the original identity skip connection in Eqn.(3) is re- placed with a simple scalingh(xl ) =l xl . If the skip connectionh(xl ) represents more complicated transforms (such as gating and 11 convolutions), in Eqn.(8)Qthe rst term becomes L1 h0 whereh0 is the derivative ofh. This product i=l i may also impede information propagation and hamper the training procedure as witnessed in the following experiments. 5 3x3 conv 3x3 conv ReLU ReLU 3x3 conv 3x3 conv 0.5 0.5 addition addition ReLU (a) original (b) constant scaling ReLU 3x3 conv 3x3 conv ReLU ReLU 1x1 conv 3x3 conv 1x1 conv 3x3 conv sigmoid sigmoid 1- 1- addition addition ReLU (c) exclusive gating ReLU (d) shortcut-only gating 3x3 conv 3x3 conv ReLU ReLU 1x1 conv 3x3 conv dropout 3x3 conv addition addition ReLU (e) conv shortcut ReLU (f) dropout shortcut 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 dierent 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 33 convolutional layers) and is challenging for optimization. Our implementa- tion details (see appendix) are the same as [1]. Throughout this paper we report the median accuracy of5 runsfor each architecture on CIFAR, reducing the impacts of random variations. Though our above analysis is driven by identityf, the experiments in this section are all based onf= ReLU as in [1]; we address identityfin the next sec- tion. Our baseline ResNet-110 has 6.61% error on the test set. The comparisons of other variants (Fig.2and Table1) are summarized as follows: Constant scaling. We set= 0:5 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 1= 0:5, 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 diculties when the shortcut signal is scaled down. 6 Table 1.Classication error on the CIFAR-10 test set using ResNet-110 [1], with dierent types of shortcut connections applied to all Residual Units. We report \fail" when the test error is higher than 20%. case Fig. on shortcut onF error (%) remark original [1] Fig.2(a) 1 1 6.61 0 1 fail This is a plain net constant Fig.2(b)scaling 0.5 1 fail 0.5 0.5 12.35 frozen gating 1g(x) g(x) fail initb exclusive g =0 to5 Fig.2(c)gating 1g(x) g(x) 8.70 initbg =-6 1g(x) g(x) 9.81 initbg =-7 shortcut-only 1g(x) 1 12.86Fig.2(d) initbg =0 gating 1g(x) 1 6.91 initbg =-6 11 conv shortcut Fig.2(e) 11 conv 1 12.22 dropout shortcut Fig.2(f) dropout 0.5 1 fail Exclusive gating. Following the Highway Networks [6,7] that adopt a gating mechanism [5], we consider a gating functiong(x) =(W g x+bg ) where a transform is represented by weights W g and biasesbg followed by the sigmoid function(x) = 1 . In a convolutional networkg(x) is realized by a 111+ex convolutional layer. The gating function modulates the signal by element-wise multiplication. We investigate the \exclusive" gates as used in [6,7] | theFpath is scaled byg(x) and the shortcut path is scaled by 1g(x). See Fig2(c). We nd that the initialization of the biasesbg is critical for training gated models, and following the guidelines 2 in [6,7], we conduct hyper-parameter search on the initial value of bg 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. Fig3(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 whenbg is not appropriately initialized. The impact of the exclusive gating mechanism is two-fold. When 1g(x) approaches 1, the gated shortcut connections are closer to identity which helps information propagation; but in this caseg(x) approaches 0 and suppresses the functionF. To isolate the eects 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 1g(x). See Fig2(d). The initialized value ofbg is still essential in this case. When the initializedbg is 0 (so initially the expectation of 1g(x) is 0.5), the network converges to a poor result of 12.86% (Table1). This is also caused by higher training error (Fig3(c)). 2 See also:people.idsia.ch/~rupesh/very_deep_learning/by [6,7]. 7 2 20 2 20 15 15 0.2 0.2 Test Error (% Test Error (% s s Training Los Training Los 10 10 ) ) 0.02 0.02 5 5 110, original 110, original 110, const scaling (0.5, 0.5) 110, exclusive gating (init b=−6) 0.002 0 0.002 00 1 2 3 4 5 6 0 1 2 3 4 5 6 Iterations Iterations (a) x 10 4 (b) x 10 4 2 20 2 20 15 15 0.2 0.2 Test Error (% Test Error (% s s Training Los Training Los 10 10 ) ) 0.02 0.02 5 5 110, original 110, original 110, shortcut−only gating (init b=0) 110, 1x1 conv shortcut 0.002 0 0.002 00 1 2 3 4 5 6 0 1 2 3 4 5 6 Iterations (c) x 10 4 Iterations (d) x 10 4 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 initializedbg is very negatively biased (e.g.,6), the value of 1g(x) 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. 11 convolutional shortcut. Next we experiment with 11 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 11 shortcut connections could be useful. But we nd 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 11 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 11 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. 8 Table 2.Classication error (%) on the CIFAR-10 test set using dierent activation functions. case Fig. ResNet-110 ResNet-164 original Residual Unit [1] Fig.4(a) 6.61 5.93 BN after addition Fig.4(b) 8.17 6.50 ReLU before addition Fig.4(c) 7.84 6.14 ReLU-only pre-activation Fig.4(d) 6.71 5.91 full pre-activation Fig.4(e) 6.37 5.46 xl xl xl xl xl weight weight weight ReLU BN BN BN BN weight ReLU ReLU ReLU ReLU BN weight weight weight weight ReLU BN BN addition BN weight ReLU addition BN ReLU BN weight ReLU ReLU addition addition addition xl+1 xl+1 xl+1 xl+1 xl+1 (b) BN after (c) ReLU before (d) ReLU-only(a) original (e) full pre-activationaddition addition pre-activation Figure 4.Various usages of activation in Table2. All these units consist of the same components | only the orders are dierent. 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.Multiplicativemanipulations (scaling, gating, 11 convolutions, and dropout) on the shortcuts can hamper information propagation and lead to optimization problems. It is noteworthy that the gating and 11 convolutional shortcuts introduce more parameters, and should have strongerrepresentationalabilities than iden- tity shortcuts. In fact, the shortcut-only gating and 11 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 activationf 9 is the identity mapping. But in the above experimentsfis 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 makefan 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 11 layer for reducing dimension, a 33 layer, and a 11 layer for restoring dimension. As designed in [1], its computational complexity is similar to the two-33 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 turningfinto an identity mapping, we go the opposite way by adopting BN after addition (Fig.4(b)). In this casefinvolves 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 reected by the diculties on reducing training loss at the beginning of training (Fib.6left). ReLU before addition. A nave choice of makingfinto an identity map- ping is to move the ReLU before addition (Fig.4(c)). However, this leads to a non-negativeoutput from the transformF, while intuitively a \residual" func- tion should take values in (1;+1). As a result, the forward propagated sig- nal 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 satised by other Residual Units including the following ones. Post-activation or pre-activation?In the original design (Eqn.(1) and Eqn.(2)), the activationxl+1 =f(yl ) aectsboth pathsin thenextResidual Unit:yl+1 =f(yl ) +F(f(yl );Wl+1 ). Next we develop anasymmetricform where an activationf^only aects theFpath:yl+1 =yl +F(f^(yl );Wl+1 ), for anyl(Fig.5(a) to (b)). By renaming the notations, we have the following form: xl+1 =xl +F(f^ (xl );Wl );: (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 activationf^ is asymmetrically adopted, it is equivalent to recastingf^as thepre-activationof the next Residual Unit. This is illustrated in Fig.5. 10 ... ... ... act. original act. act. Residual weight asymmetric weight weight Unit output act. activation act. act. weight weight weight addition addition addition act. pre-activation act. Residual Unit act. weight weight weight act. act. act. weight weight weight addition addition addition act. ... adopt output activation ... ... equivalent to only to weight path (a) (b) (c) Figure 5.Using asymmetric after-addition activation is equivalent to constructing a pre-activationResidual Unit. Table 3.Classication error (%) on the CIFAR-10/100 test set using the original Residual Units and our pre-activation Residual Units. dataset network baseline unit pre-activation unit ResNet-110 (1layer skip) 9.90 8.91 ResNet-110 6.61 6.37CIFAR-10 ResNet-164 5.93 5.46 ResNet-1001 7.61 4.92 ResNet-164 25.16 24.33CIFAR-100 ResNet-1001 27.82 22.71 The distinction between post-activation/pre-activation is caused by the pres- ence of the element-wiseaddition. For a plain network that hasNlayers, there areN1 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. Table2shows 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 benets 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 2 20 2 20 164, original 164, proposed (pre−activation) 15 15 0.2 0.2 Test Error (% Test Error (% s s Training Los Training Los 10 10 ) ) 0.02 0.02 5 5 110, original 110, BN after add 0.002 0 0.002 0 0 1 2 3 4 5 6 0 1 2 3 4 5 6 Iterations Iterations x 10 4 x 10 4 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 nd the impact of pre-activation is twofold. First, the optimization is further eased (comparing with the baseline ResNet) becausefis an identity mapping. Second, using BN as pre-activation improves regularization of the models. Ease of optimization. This eect 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 eect becomes prominent and Eqn.(3) (so Eqn.(5)) is not a good approximation. On the other hand, whenfis 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 nd 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 suer 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 thatyl in Eqn.(1) is more frequently above zero andfdoes not truncate it (xl is always non-negative due to the pre- vious ReLU, soyl is below zero only when the magnitude ofFis very negative). The truncation, however, is more frequent when there are 1000 layers. 12 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). CIFAR-10 error (%) CIFAR-100 error (%) NIN [15] 8.81 NIN [15] 35.68 DSN [16] 8.22 DSN [16] 34.57 FitNet [17] 8.39 FitNet [17] 35.04 Highway [7] 7.72 Highway [7] 32.39 All-CNN [14] 7.25 All-CNN [14] 33.71 ELU [12] 6.55 ELU [12] 24.28 FitResNet, LSUV [18] 5.84 FitNet, LSUV [18] 27.66 ResNet-110 [1] (1.7M) 6.61 ResNet-164 [1] (1.7M) 25.16 ResNet-1202 [1] (19.4M) 7.93 ResNet-1001 [1] (10.2M) 27.82 ResNet-164 [ours] (1.7M) 5.46 ResNet-164 [ours] (1.7M) 24.33 ResNet-1001 [ours] (10.2M) 4.92 (4.890.14) ResNet-1001 [ours] (10.2M) 22.71 (22.680.22) ResNet-1001 [ours] (10.2M) y 4.62 (4.690.20) Reducing overtting. 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 reg- ularization eect [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 lter sizes, nor use regularization techniques (such as dropout) which are very eective for these small datasets. We obtain these results via a simple but essential concept | going deeper. These results demonstrate the potential ofpushing 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 diculties. The training error of these non-identity shortcut networks is obviously higher than the original ResNet at the rst 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]. method augmentation train crop test crop top-1 top-5 ResNet-152, original Residual Unit [1] scale 224224224224 23.0 6.7 ResNet-152, original Residual Unit [1] scale 224224320320 21.3 5.5 ResNet-152,pre-actResidual Unit scale 224224320320 21.1 5.5 ResNet-200, original Residual Unit [1] scale 224224320320 21.8 6.0 ResNet-200,pre-actResidual Unit scale 224224320320 20.7 5.3 ResNet-200,pre-actResidual Unit scale+asp ratio 22422432032020.1 y 4.8 y Inception v3 [19] scale+asp ratio 299299299299 21.2 5.6 (similar to Fig.3), and we decided to halt training due to limited resources. But we did nish 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 (224224) 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 sides2[256;480], and so the test of a 224224 crop ons= 256 (as did in [1]) is negatively biased. Instead, we test a single 320320 crop froms= 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 299299 used by Inception v3 [19], allowing a fairer comparison. The original ResNet-152 [1] has top-1 error of 21.3% on a 320320 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 diculties. However, the original ResNet-200 has an error rate of 21.8%, higher than the baseline ResNet-152. But we nd that the original ResNet-200 haslowertraining error than ResNet-152, suggesting that it suers from overtting. 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 2828. 14 depth (so a 1001-layer net is10complex 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 ipping 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 rst and the last Residual Units of the entire net- work. For the rst Residual Unit (that follows a stand-alone convolutional layer, conv 1 ), we adopt the rst activation right after conv 1 and before splitting into two paths; for the last Residual Unit (followed by average pooling and a fully- connected classier), 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 modication procedure as shown in Fig.5. The bottleneck Residual Units (for ResNet-164/1001 on CIFAR) are con- structed following [1]. For example, a " # 33, 16 unit in ResNet-110 is replaced 33, 162 3 with a 11, 166 7 unit in ResNet-164, both of which have roughly the same num- 433, 165 11, 64 ber 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 classication. In: ICCV. (2015)