Datasets:

jamescalam

/

llama-2-arxiv-papers-chunked

like 17

1605.07681

13

the matrix multiplication efficient and exact, but it is also easy to integrate into the traditional FCN framework for both training and prediction schemes. Additionally, due to the use of standard convolution and matrix multiplication operations, our RWN can be trivially trained via standard back-propagation in an end-to-end fashion. 4. Convolutional Random Walk Networks In this work, our goal is to integrate a random walk process into the FCN architecture to encourage coherent semantic segmentation among pixels that are similar to each other. Such a process introduces an explicit grouping mechanism, which should be beneficial in addressing the issues of (1) poor localization around the boundaries, and (2) spatially fragmented segmentations. A schematic illustration of our proposed RWN architecture is presented in Fig. 2. Our RWN is a network composed of two branches: (1) one branch that predicts semantic segmentation potentials, and (2) another branch devoted to predicting pixel-level affinities. These two branches are merged via a novel random walk layer that encourages spatially coherent semantic segmentation. The entire RWN can be jointly optimized end-to-end. We now describe each of the components of the RWN architecture in more detail. 4.1. Semantic Segmentation Branch For the semantic segmentation branch, we present results

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

14

the components of the RWN architecture in more detail. 4.1. Semantic Segmentation Branch For the semantic segmentation branch, we present results for several variants of the DeepLab segmentation systems, including DeepLab-LargeFOV [6], DeepLab-attention [7], and DeepLab-v2, which is one of the top performing segmentation systems. DeepLab-largeFOV is a fully convolutional adaptation of the VGG [26] architecture, which contains 16convolutional layers. DeepLab-attention [7], is a multi-scale VGG based network, for which each multi-scale branch focuses on a specific part of the image. Finally, DeepLab-v2 is a multi-scale network based on the residual network [11] implementation. We note that even though we use a DeepLab architecture in our experiments, other architectures such as [2] and many others could be integrated into our framework. 4.2. Pixel-Level Affinity Branch To learn the pairwise pixel-level affinities, we employ a separate affinity learning branch with its own learning objective (See Fig. 2). The affinity branch is connected with the inputnn3RGB image, and low-level conv1_1and

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

15

the inputnn3RGB image, and low-level conv1_1and conv1_2layers. The feature maps corresponding to these layers arennin width and height but they have a different number of channels ( 3;64;and64respectively). Let k be the total number of affinity learning parameters (in our casek= 3 + 64 + 64 = 131 ). Then, let Fbe a sparse n2n2kmatrix that stores L1distances between each pixel and all of its neighbors within a radius R, according to each channel. Note that the distances are notsummed up across the kchannels, but instead they are computed and stored separately for each channel. The resulting matrix F is then used as an input to the affinity branch, as shown in Figure 2. The affinity branch consists of a 11kconvolutional layer and an exponential layer. The output of the exponential layer is then attached to the Euclidean loss layer and is optimized to predict the ground truth pixel affinities, which are obtained from the original semantic segmentation anInput DeepLab_v2 RWN_v2 Figure 3: A figure illustrating the segmentation results of

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

16

are obtained from the original semantic segmentation anInput DeepLab_v2 RWN_v2 Figure 3: A figure illustrating the segmentation results of our RWN and the DeepLab-v2 network. Note that RWN produced segmentations are spatially smoother and produce less false positive predictions than the DeepLab-v2 system. notations. Specifically, we set the ground truth affinity between two pixels to 1if the pixels share the same semantic label and have distance less than Rfrom each other. Note thatF, which is used as an input to the affinity branch, is a sparse matrix, as only a small fraction of all the entries in F are populated with non-zero values. The rest of the entries are ignored during the computation. Also note that we only use features from RGB, conv1_1 andconv1_2layers, because they are not affected by pooling, and thus, preserve the original spatial resolution. We also experimented with using features from deeper FCN layers such as fc6, and fc7. However, we observed that features from deeper layers are highly correlated to the predicted semantic segmentation unary potentials, which causes redundancy and little improvement in the segmentation performance. We also experimented with using more

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

17

features from deeper layers are highly correlated to the predicted semantic segmentation unary potentials, which causes redundancy and little improvement in the segmentation performance. We also experimented with using more than one convolutional layer in the affinity learning branch, but observed that additional layers provide negligible improvements in accuracy. 4.3. Random Walk Layer To integrate the semantic segmentation potentials and our learned pixel-level affinities, we introduce a novel random walk layer, which propagates the semantic segmentation information based on the learned affinities. The random walk layer is connected to the two bottom layers: (1) the fc8 layer containing the semantic segmentation potentials, and (2) the affinity layer that outputs a sparse n2n2random walk transition matrix A. Then, let fdenote the activation values from the fc8layer, reshaped to the dimensions ofn2m, wheren2refers to the number of pixels, and mis the number of object classes in the dataset. A single random walk layer implements one step of the random walk Method aero bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv mean overall

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

18

random walk layer implements one step of the random walk Method aero bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv mean overall DeepLab-largeFOV 79.8 71.5 78.9 70.9 72.1 87.9 81.2 85.7 46.9 80.9 56.5 82.6 77.9 79.3 80.1 64.4 77.6 52.7 80.3 70.0 73.8 76.0 RWN-largeFOV 81.6 72.1 82.3 72.0 75.4 89.1 82.5 87.4 49.1 83.6 57.9 84.8 80.7 80.2 81.2 65.7 79.7 55.5 81.5 74.0 75.8 77.9 DeepLab-attention 83.4 76.0 83.0 74.2 77.6 91.6 85.2 89.1 54.4 86.1 62.9 86.7 83.8 84.2 82.4 70.2 84.7 61.0 84.8 77.9 79.0 80.5

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

19

RWN-attention 84.7 76.6 85.5 74.0 79.0 92.4 85.6 90.0 55.6 87.4 63.5 88.2 85.0 84.8 83.4 70.1 85.9 62.6 85.1 79.3 79.9 81.5 DeepLab-v2 85.5 50.6 86.9 74.4 82.7 93.1 88.4 91.9 62.1 89.7 71.5 90.3 86.2 86.3 84.6 75.1 87.6 72.2 87.8 81.3 81.4 83.4 RWN-v2 86.0 50.0 88.4 73.5 83.9 93.4 88.6 92.5 63.9 90.9 72.6 90.9 87.3 86.9 85.7 75.0 89.0 74.0 88.1 82.3 82.1 84.3 Table 2: Semantic segmentation results on the SBD dataset according to the per-pixel Intersection over Union evaluation metric. From the results, we observe that our proposed RWN consistently outperforms DeepLab-LargeFOV , DeepLabattention, and DeepLab-v2 baselines.

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

20

metric. From the results, we observe that our proposed RWN consistently outperforms DeepLab-LargeFOV , DeepLabattention, and DeepLab-v2 baselines. process, which can be performed as ^y=Af, where ^yindicates the diffused segmentation predictions, and Adenotes the random walk transition matrix. The random walk layer is then attached to the softmax loss layer, and is optimized to predict ground truth semantic segmentations. One of the advantages of our proposed random walk layer is that it is implemented as a matrix multiplication, which makes it possible to back-propagate the gradients to both (1) the affinity branch and (2) the segmentation branch. Let the softmax-loss gradient be an n2m matrix@L @^y, wheren2is the number of pixels in the fc8layer, andmis the number of predicted object classes. Then the gradients, which are back-propagated to the semantic segmentation branch are computed as@L @f=AT@L @^y, whereAT is the transposed random walk transition matrix. Also, the gradients, that are back-propagated to the affinity branch are computed as@L @A=@L

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

21

gradients, that are back-propagated to the affinity branch are computed as@L @A=@L @^yfT, wherefTis amn2matrix that contains transposed activation values from the fc8layer of the segmentation branch. We note that@L @Ais asparse n2n2matrix, which means that the above matrix multiplication only considers the pixel pairs that correspond to the non-zero entries in the random walk transition matrix A. 4.4. Random Walk Prediction at Testing In the previous subsection, we mentioned that the prediction in the random walk layer can be done via a simple matrix multiplication operation ^y=Af, whereAdenotes the random walk transition matrix, and fdepicts the activation values from the fc8layer. Typically, we want to apply multiple random walk steps until convergence is reached. However, we also do not want to deviate too much from our initial segmentation predictions, in case the random walk transition matrix is not perfectly accurate, which is a reasonable expectation. Thus, our prediction scheme needs to balance two effects: (1) propagating the segmentation information across the nodes using a random walk transition matrix, and (2) not deviating too much from the initial segmentation.

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

22

balance two effects: (1) propagating the segmentation information across the nodes using a random walk transition matrix, and (2) not deviating too much from the initial segmentation. This tradeoff between the two quantities is very similar to the idea behind MRF and CRF models, which try to minimize an energy formed by unary and pairwise terms. However, as discussed earlier, MRF and CRF methods tend to use 1) grid structure graphs and 2) various approximateinference techniques to propagate segmentation information globally. In comparison, our random walk approach is advantageous because it can use 1) any arbitrary graph structure and 2) an exact matrix multiplication operation to achieve the same goal. Let us first denote our segmentation prediction after t+1 random walk steps as ^yt+1. Then our general prediction scheme can be written as: ^yt+1=A^yt+ (1)f (1) wheredenotes a parameter [0;1]that controls the tradeoff between (1) diffusing segmentation information along the connections of a random walk transition matrix and (2) not deviating too much from initial segmentation values (i.e. the outputs of the last layer of the FCN). Let us now initialize ^y0to contain the output values from the fc8

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

23

values (i.e. the outputs of the last layer of the FCN). Let us now initialize ^y0to contain the output values from the fc8 layer, which we denoted with f. Then we can write our prediction equation by substituting the recurrent expressions: ^yt+1= (A)t+1f+ (1)tX i=0(A)if (2) Now, because we want to apply our random walk procedure until convergence we set t=1. Then, because our random walk transition matrix is stochastic we know thatlimt!1(A)t+1= 0 . Furthermore, we can write St=Pt i=0Ai=I+A+A2+:::+At, whereIis an identity matrix, and where Stdenotes a partial sum of random walk transitions until iteration t. We can then write StASt=IAt+1, which implies: (IA)St=IAt+1(3) From our previous derivation, we already know that limt!1(A)t+1= 0, which implies that S1= (IA)1(4) Thus, our final prediction equation, which corresponds to applying repeated random walk steps until convergence,

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

24

S1= (IA)1(4) Thus, our final prediction equation, which corresponds to applying repeated random walk steps until convergence, can be written as ^y1= (IA)1f (5) Input DeepLab_v2-CRF RWN_v2 Figure 4: Comparison of segmentation results produced by our RWN versus the DeepLab-v2-CRF system. It can be noticed that, despite not using any post-processing steps, our RWN predicts fine object details (e.g., bike wheels or plane wings) more accurately than DeepLab-v2-CRF, which fails to capture some these object parts. In practice, the random walk transition matrix Ais pretty large, and inverting it is impractical. To deal with this problem, we shrink the matrix (IA)using a simple and efficient technique presented in [1], and then invert it to compute the final segmentation. In the experimental section, we show that such a prediction scheme produces solid results and is still pretty efficient ( 1 second per image). We also note that we use this prediction scheme only during testing. During training we employ a scheme that uses a single

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

25

note that we use this prediction scheme only during testing. During training we employ a scheme that uses a single random walk step (but with a much larger radius), which is faster. We explain this procedure in the next subsection. 4.5. Implementation Details We jointly train our RWN in an end-to-end fashion for 2000 iterations, with a learning rate of 105,0:9momentum, the weight decay of 5
105, and 15samples per batch. For the RWN model, we set the tradeoff parameter to 0:01. During testing we set the random walk connectivity radiusR= 5 and apply the random walk procedure until convergence. However, during training we set R= 40 , and apply a single random walk step. This training strategy works well because increasing the radius size eliminates the need for multiple random walk steps, which speeds up the training. However, using R= 5 and applying an infinite number of random walk steps until convergence still yields slightly better results (see study in 5.4), so we use it during testing. For all of our experiments, we use a Caffe library [13]. During training, we also employ data augmentation techniques such as cropping, and mirroring. 5. Experimental Results

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

26

5. Experimental Results In this section, we present our results for semantic segmentation on the SBD [10] dataset, which contains objects and their per-pixel labels for 20Pascal VOC classes (excluding the background class). We also include scene label-Method mean IOU overall IOU DeepLab-largeFOV-CRF 75.7 77.7 RWN-largeFOV 75.8 77.9 DeepLab-attention-CRF 79.9 81.6 RWN-attention 79.9 81.5 DeepLab-v2-CRF 81.9 84.2 RWN-v2 82.1 84.3 DeepLab-DT 76.6 78.7 RWN 76.7 78.8 Table 3: Quantitative comparison between our RWN model and several variants of the DeepLab system that use a dense CRF or a domain-transfer (DT) filter for post-processing. These results suggest that our RWN acts as an effective globalization scheme, since it produces results that are similar or even better than the results achieved by postprocessing the DeepLab outputs with a CRF or DT. ing results on the commonly used Stanford background [9] and Sift Flow [18] datasets. We evaluate our segmentation results on these tasks using the standard metric of the

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

27

ing results on the commonly used Stanford background [9] and Sift Flow [18] datasets. We evaluate our segmentation results on these tasks using the standard metric of the intersection-over-union (IOU) averaged per pixels across all the classes from each dataset. We also include the classagnostic overall pixel intersection-over-union score, which measures the per-pixel IOU across all classes. We experiment with several variants of the DeepLab system [6, 7] as our main baselines throughout our experiments: DeepLab-LargeFOV [6], DeepLab-attention [7], and DeepLab-v2. Our evaluations provide evidence for four conclusions: In subsections 5.1, 5.2, we demonstrate that our proposed RWN outperforms DeepLab baselines for both semantic segmentation, and scene labeling tasks. In subsection 5.1, we demonstrate that, compared to the dense CRF approaches, RWN predicts segmentations that are spatially smoother. In Subsection 5.3, we show that our approach is more efficient than the denseCRF inference. Finally, in Subsection, 5.4, we demonstrate that our random walk layer is beneficial and that our model is flexible to use different graph structures.

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

28

random walk layer is beneficial and that our model is flexible to use different graph structures. 5.1. Semantic Segmentation Task Standard Evaluation. In Table 2, we present semantic segmentation results on the Pascal SBD dataset [10], which contains 8055 training and 2857 testing images. These results indicate that RWN consistently outperforms all three of the DeepLab baselines. In Figure 3, we also compare qualitative segmentation results of a DeepLab-v2 network and our RWN model. We note that the RWN segmentations contain fewer false positive predictions and are also spatially smoother across the object regions. 0 2 4 6 8 10 12 14 16 18 200.20.250.30.350.40.450.50.55 Trimap Width (in Pixels)Pixel Classification Error (%)Localization Around the Boundaries Error DeepLab RWNFigure 5: Localization error around the object boundaries within a trimap. Compared to the DeepLab system (blue), our RWN (red) achieves lower segmentation error around object boundaries for all trimap widths. Furthermore, in Table 3, we present experiments where we compare RWN with models using dense CRFs [15] to post-process the predictions of DeepLab systems. We also

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

29

Furthermore, in Table 3, we present experiments where we compare RWN with models using dense CRFs [15] to post-process the predictions of DeepLab systems. We also include DeepLab-DT [5], which uses domain-transfer filtering to refine the segmentations inside an FCN. Based on these results, we observe that, despite not using any postprocessing, our RWN produces results similar to or even better than the DeepLab models employing post-processing. These results indicate that RWN can be used as a globalization mechanism to ensure spatial coherence in semantic segmentation predictions. In Figure 4 we present qualitative results where we compare the final segmentation predictions of RWN and the DeepLab-v2-CRF system. Based on these qualitative results, we observe that RWN captures more accurately the fine details of the objects, such as the bike wheels, or plane wings. The DeepLab-v2-CRF system misses some of these object parts. Localization Around the Boundaries. Earlier we claimed that due to the use of large receptive fields and many pooling layers, FCNs tend to produce blobby segmentations that lack fine object boundary details. We want

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

30

many pooling layers, FCNs tend to produce blobby segmentations that lack fine object boundary details. We want to show that our RWN produces more accurate segmentations around object boundaries the traditional FCNs. Thus, adopting the practice from [15], we evaluate segmentation accuracy around object boundaries. We do so by counting the relative number of misclassified pixels within a narrow band (“trimap”) surrounding the ground truth object boundaries. We present these results in Figure 5. The results show that RWN achieves higher segmentation accuracy than the DeepLab (DL) system for all trimap widths considered in this test. Spatial Smoothness. We also argued that applying the dense CRF [15] as a post-processing technique often leads to spatially fragmented segmentations (see the right side of Fig. 1). How can we evaluate whether a given method produces spatially smooth or spatially fragmented segmenta-Method MF AP DeepLab-largeFOV-CRF 0.676 0.457 RWN-largeFOV 0.703 0.494 DeepLab-attention-CRF 0.722 0.521 RWN-attention 0.747 0.556 DeepLab-v2-CRF 0.763 0.584

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

31

RWN-attention 0.747 0.556 DeepLab-v2-CRF 0.763 0.584 RWN-v2 0.773 0.595 Table 4: Quantitative comparison of spatial segmentation smoothness. We extract the boundaries from the predicted segmentations and evaluate them against ground truth object boundaries using max F-score (MF) and average precision (AP) metrics. These results suggest that RWN segmentations are spatially smoother than the DeepLab-CRF segmentations across all baselines. tions? Intuitively, spatially fragmented segmentations will produce many false boundaries that do not correspond to actual object boundaries. Thus, to test the spatial smoothness of a given segmentation, we extract the boundaries from the segmentation and then compare these boundaries against ground truth object boundaries using the standard maximum F-score (MF) and average precision (AP) metrics, as done in the popular BSDS benchmark [22]. We perform this experiment on the Pascal SBD dataset and present these results in Table 4. We can see that the boundaries extracted from the RWN segmentations yield better MF and AP results compared to the boundaries extracted from the different variants of the DeepLab-CRF system. Thus, these results suggest that RWN produces spatially smoother segmentations than DeepLab-CRF. 5.2. Scene Labeling

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

32

results suggest that RWN produces spatially smoother segmentations than DeepLab-CRF. 5.2. Scene Labeling We also tested our RWN on the task of scene labeling using two popular datasets: Stanford Background [9] and Sift Flow [18]. Stanford Background is a relatively small dataset for scene labeling. It contains 715images, which we randomly split into 600training images and 115testing images. In contrast, the Sift Flow dataset contains 2489 training examples and 201testing images. For all of our experiments, we use the DeepLab-largeFOV [6] architecture since it is smaller and more efficient to train and test. To evaluate scene labeling results, we use the overall IOU evaluation metric which is a commonly used metric for this task. In Table 5, we present our scene labeling results on both of these datasets. Our results indicate that our RWN method outperforms the DeepLab baseline by 2:57%, and 2:54% on these two datasets, respectively. 5.3. Runtime Comparisons We also include the runtime comparison of our RWN approach versus the denseCRF inference. We note that using a single core of a 2:7GHz Intel Core i7 processor, the denseCRF inference requires 3.301 seconds per image on Input Image Iteration 0Iteration 50

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

33

denseCRF inference requires 3.301 seconds per image on Input Image Iteration 0Iteration 50 Figure 6: A figure illustrating how the probability predictions change as we apply more random walk steps. Note that the RWN predictions become more refined and better localized around the object boundaries as more random walk steps are applied. average on a Pascal SBD dataset. In comparison, a single iteration of a random walk, which is simply a sparse matrix multiplication, takes 0.032 seconds on average on the same Pascal SBD dataset. A DeepLab_v2 post-processed with denseCRF achieves 81:9%IOU score on this same Pascal SBD dataset. In comparison, RWN_v2 with a single random walk iteration and with R=40 (radius) achieves 82:2% IOU, which is both more accurate and more than 100times more efficient than the denseCRF inference. 5.4. Ablation Experiments Optimal Number of Random Walk Steps. In Figure 7, we illustrate how the IOU accuracy changes when we use a different number of random walk steps. We observe that the segmentation accuracy keeps increasing as we apply more random walk steps, and that it reaches its peak performance when the random walk process converges, which indicates

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

34

segmentation accuracy keeps increasing as we apply more random walk steps, and that it reaches its peak performance when the random walk process converges, which indicates the effectiveness of our random walk step procedure. In Figure 6, we also illustrate how the predicted object segmentation probabilities change as we apply more random walk steps. We observe that the object boundaries become much better localized as more iterations of random walk are applied. Radius Size. To analyze the effect of a radius size in the RWN architecture, we test alternative versions of our model with different radii sizes. Our results indicate, that the RWN model produces similar results with different radii in the interval of R > 3andR < 20if the random walk step process is applied until convergence. We also note that, if we select R= 40 , and apply a random walk step only once , we can achieve the segmentation accuracy of 75:5%and77:6%according to the two evaluation metrics, 0 5 10 15 20 25 30 40 50 60 inf0.7350.740.7450.750.7550.760.765 Number of Random Walk StepsIOU AccuracyAccuracy versus the Number of Random Walk Steps

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

35

Number of Random Walk StepsIOU AccuracyAccuracy versus the Number of Random Walk Steps Mean−class IOUFigure 7: IOU accuracy as a function of the number of random walk steps. From this plot we observe that the segmentation accuracy keeps improving as we apply more random walk steps and that it reaches its peak when the random walk process converges. DeepLab-largeFOV RWN-largeFOV Stanford-BG 65.74 68.31 Sift-Flow 67.31 69.85 Table 5: Scene labeling results on the Stanford Background and Sift-Flow datasets measured according to the overall IOU evaluation metric. We use a DeepLab-largeFOV network as base model, and show that our RWN yields better results on both of these scene labeling datasets. respectively. In comparison, choosing R= 5 and applying random walk until convergence yields the accuracies of 75:8%and77:9%, which is slightly better. However, note that selecting R= 40 , and applying multiple random walk steps does not yield any improvement in segmentation accuracy. These experiments show the flexibility of our model compared to the MRF or CRF models, which typically use graphs with a fixed grid structure. Our model has the ability to use different graph structures depending on the problem. 6. Conclusion

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

36

graphs with a fixed grid structure. Our model has the ability to use different graph structures depending on the problem. 6. Conclusion In this work, we introduced Random Walk Networks (RWNs), and showed that, compared to traditional fully convolutional networks (FCNs), they produce improved accuracy for the same model complexity. Our RWN addresses the issues of 1) poor localization around the segmentation boundaries and 2) spatially disjoint segmentations. Additionally, our implementation of RWN uses only 131additional learnable parameters ( 0:0008% of the original number of the parameters in the network) and it can be easily integrated into the standard FCN learning framework for a joint end-to-end training. Finally, RWN provides a more efficient alternative to the denseCRF approaches. Our future work includes experimenting with alternative RWN architectures and applying RWN to new domains such as language processing or speech recognition. References [1] Pablo Arbelaez, J. Pont-Tuset, Jon Barron, F. Marqués, and Jitendra Malik. Multiscale combinatorial grouping. In Computer Vision and Pattern Recognition (CVPR) , 2014. 6 [2] Vijay Badrinarayanan, Ankur Handa, and Roberto Cipolla.

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

37

[2] Vijay Badrinarayanan, Ankur Handa, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for robust semantic pixel-wise labelling. arXiv preprint arXiv:1505.07293 , 2015. 4 [3] Bahman Bahmani, Abdur Chowdhury, and Ashish Goel. Fast incremental and personalized pagerank. Proc. VLDB Endow., 4(3):173–184, December 2010. 2, 3 [4] Gedas Bertasius, Jianbo Shi, and Lorenzo Torresani. Semantic segmentation with boundary neural fields. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2016. 2 [5] Liang-Chieh Chen, Jonathan T. Barron, George Papandreou, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with task-specific edge detection using cnns and a discriminatively trained domain transform. CVPR , 2016. 1, 2, 7 [6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos,

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

38

2, 7 [6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Semantic image segmentation with deep convolutional nets and fully. In ICLR , 2015. 1, 2, 4, 6, 7 [7] Liang-Chieh Chen, Yi Yang, Jiang Wang, Wei Xu, and Alan L. Yuille. Attention to scale: Scale-aware semantic image segmentation. CVPR , 2016. 2, 4, 6 [8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In The IEEE International Conference on Computer Vision (ICCV) , December 2015. 1 [9] S. Gould, R. Fulton, and D. Koller. Decomposing a scene into geometric and semantically consistent regions. In Proceedings of the International Conference on Computer Vision (ICCV) , 2009. 6, 7 [10] Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision (ICCV) , 2011. 6

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

39

Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision (ICCV) , 2011. 6 [11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2016. 4 [12] Seunghoon Hong, Hyeonwoo Noh, and Bohyung Han. Decoupled deep neural network for semi-supervised semantic segmentation. In NIPS) , December 2015. 1, 2 [13] 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. 2, 6 [14] Iasonas Kokkinos. Surpassing humans in boundary detection using deep learning. CoRR , abs/1511.07386, 2015. 2[15] Philipp Krähenbühl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

40

in fully connected crfs with gaussian edge potentials. In J. Shawe-Taylor, R.S. Zemel, P.L. Bartlett, F. Pereira, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems 24 , pages 109–117. Curran Associates, Inc., 2011. 1, 7 [16] John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML ’01, pages 282–289, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. 3 [17] Guosheng Lin, Chunhua Shen, Ian D. Reid, and Anton van den Hengel. Efficient piecewise training of deep structured models for semantic segmentation. CoRR , abs/1504.01013, 2015. 1, 2, 3 [18] Ce Liu, Jenny Yuen, and Antonio Torralba. Nonparametric Scene Parsing via Label Transfer , pages 207–236. Springer International Publishing, Cham, 2016. 6, 7

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

41

Scene Parsing via Label Transfer , pages 207–236. Springer International Publishing, Cham, 2016. 6, 7 [19] Ziwei Liu, Xiaoxiao Li, Ping Luo, Chen Change Loy, and Xiaoou Tang. Semantic image segmentation via deep parsing network. In ICCV , 2015. 1, 2, 3 [20] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2015. 1, 2 [21] Laszlo Lovasz. Random walks on graphs: A survey, 1993. 2, 3 [22] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proc. 8th Int’l Conf. Computer Vision , volume 2, pages 416–423, July 2001. 7 [23] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. InComputer Vision (ICCV), 2015 IEEE International Conference on , 2015. 1, 2 [24] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

42

InComputer Vision (ICCV), 2015 IEEE International Conference on , 2015. 1, 2 [24] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical Report 1999-66, Stanford InfoLab, November 1999. Previous number = SIDL-WP-1999-0120. 2, 3 [25] Xiaojuan Qi, Jianping Shi, Shu Liu, Renjie Liao, and Jiaya Jia. Semantic segmentation with object clique potential. InThe IEEE International Conference on Computer Vision (ICCV) , December 2015. 2 [26] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR , abs/1409.1556, 2014. 4 [27] Shuai Zheng, Sadeep Jayasumana, Bernardino RomeraParedes, Vibhav Vineet, Zhizhong Su, Dalong Du, Chang Huang, and Philip Torr. Conditional random fields as recurrent neural networks. In International Conference on Computer Vision (ICCV) , 2015. 1, 2, 3

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1606.04474

0

Learning to learn by gradient descent by gradient descent Marcin Andrychowicz1, Misha Denil1, Sergio Gómez Colmenarejo1, Matthew W. Hoffman1, David Pfau1, Tom Schaul1, Brendan Shillingford1;2, Nando de Freitas1;2;3 1Google DeepMind2University of Oxford3Canadian Institute for Advanced Research marcin.andrychowicz@gmail.com {mdenil,sergomez,mwhoffman,pfau,schaul}@google.com brendan.shillingford@cs.ox.ac.uk ,nandodefreitas@google.com Abstract The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

1

demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art. 1 Introduction Frequently, tasks in machine learning can be expressed as the problem of optimizing an objective functionf()defined over some domain 2. The goal in this case is to find the minimizer = arg min2f(). While any method capable of minimizing this objective function can be applied, the standard approach for differentiable functions is some form of gradient descent, resulting in a sequence of updates t+1=ttrf(t): The performance of vanilla gradient descent, however, is hampered by the fact that it only makes use of gradients and ignores second-order information. Classical optimization techniques correct this behavior by rescaling the gradient step using curvature information, typically via the Hessian matrix of second-order partial derivatives—although other choices such as the generalized Gauss-Newton matrix or Fisher information matrix are possible. Much of the modern work in optimization is based around designing update rules tailored to specific classes of problems, with the types of problems of interest differing between different research

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

2

Much of the modern work in optimization is based around designing update rules tailored to specific classes of problems, with the types of problems of interest differing between different research communities. For example, in the deep learning community we have seen a proliferation of optimization methods specialized for high-dimensional, non-convex optimization problems. These include momentum [Nesterov, 1983, Tseng, 1998], Rprop [Riedmiller and Braun, 1993], Adagrad [Duchi et al., 2011], RMSprop [Tieleman and Hinton, 2012], and ADAM [Kingma and Ba, 2015]. More focused methods can also be applied when more structure of the optimization problem is known [Martens and Grosse, 2015]. In contrast, communities who focus on sparsity tend to favor very different approaches [Donoho, 2006, Bach et al., 2012]. This is even more the case for combinatorial optimization for which relaxations are often the norm [Nemhauser and Wolsey, 1988]. 30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.arXiv:1606.04474v2 [cs.NE] 30 Nov 2016 optimizeroptimizeeparameter updates error signalFigure 1: The optimizer (left) is provided with performance of the optimizee (right) and proposes

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

3

optimizeroptimizeeparameter updates error signalFigure 1: The optimizer (left) is provided with performance of the optimizee (right) and proposes updates to increase the optimizee’s performance. [photos: Bobolas, 2009, Maley, 2011]This industry of optimizer design allows different communities to create optimization methods which exploit structure in their problems of interest at the expense of potentially poor performance on problems outside of that scope. Moreover the No Free Lunch Theorems for Optimization [Wolpert and Macready, 1997] show that in the setting of combinatorial optimization, no algorithm is able to do better than a random strategy in expectation. This suggests that specialization to a subclass of problems is in fact theonly way that improved performance can be achieved in general. In this work we take a different tack and instead propose to replace hand-designed update rules with a learned update rule, which we call the optimizerg, specified by its own set of parameters . This results in updates to the optimizee fof the form t+1=t+gt(rf(t);): (1) A high level view of this process is shown in Figure 1. In what follows we will explicitly model

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

4

A high level view of this process is shown in Figure 1. In what follows we will explicitly model the update rule gusing a recurrent neural network (RNN) which maintains its own state and hence dynamically updates as a function of its iterates. 1.1 Transfer learning and generalization The goal of this work is to develop a procedure for constructing a learning algorithm which performs well on a particular class of optimization problems. Casting algorithm design as a learning problem allows us to specify the class of problems we are interested in through example problem instances. This is in contrast to the ordinary approach of characterizing properties of interesting problems analytically and using these analytical insights to design learning algorithms by hand. It is informative to consider the meaning of generalization in this framework. In ordinary statistical learning we have a particular function of interest, whose behavior is constrained through a data set of example function evaluations. In choosing a model we specify a set of inductive biases about how we think the function of interest should behave at points we have not observed, and generalization corresponds to the capacity to make predictions about the behavior of the target function at novel points. In our setting the examples are themselves problem instances , which means generalization corresponds to the ability to transfer knowledge between different problems. This reuse of problem structure is commonly known as transfer learning , and is often treated as a subject in its own right.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

5

corresponds to the ability to transfer knowledge between different problems. This reuse of problem structure is commonly known as transfer learning , and is often treated as a subject in its own right. However, by taking a meta-learning perspective, we can cast the problem of transfer learning as one of generalization, which is much better studied in the machine learning community. One of the great success stories of deep-learning is that we can rely on the ability of deep networks to generalize to new examples by learning interesting sub-structures. In this work we aim to leverage this generalization power, but also to lift it from simple supervised learning to the more general setting of optimization. 1.2 A brief history and related work The idea of using learning to learn ormeta-learning to acquire knowledge or inductive biases has a long history [Thrun and Pratt, 1998]. More recently, Lake et al. [2016] have argued forcefully for its importance as a building block in artificial intelligence. Similarly, Santoro et al. [2016] frame multi-task learning as generalization, however unlike our approach they directly train a base learner rather than a training algorithm. In general these ideas involve learning which occurs at two different time scales: rapid learning within tasks and more gradual, meta learning across many different tasks.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

6

rather than a training algorithm. In general these ideas involve learning which occurs at two different time scales: rapid learning within tasks and more gradual, meta learning across many different tasks. Perhaps the most general approach to meta-learning is that of Schmidhuber [1992, 1993]—building on work from [Schmidhuber, 1987]—which considers networks that are able to modify their own weights. Such a system is differentiable end-to-end, allowing both the network and the learning 2 algorithm to be trained jointly by gradient descent with few restrictions. However this generality comes at the expense of making the learning rules very difficult to train. Alternatively, the work of Schmidhuber et al. [1997] uses the Success Story Algorithm to modify its search strategy rather than gradient descent; a similar approach has been recently taken in Daniel et al. [2016] which uses reinforcement learning to train a controller for selecting step-sizes. Bengio et al. [1990, 1995] propose to learn updates which avoid back-propagation by using simple parametric rules. In relation to the focus of this paper the work of Bengio et al. could be characterized aslearning to learn without gradient descent by gradient descent . The work of Runarsson and

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

7

aslearning to learn without gradient descent by gradient descent . The work of Runarsson and Jonsson [2000] builds upon this work by replacing the simple rule with a neural network. Cotter and Conwell [1990], and later Younger et al. [1999], also show fixed-weight recurrent neural networks can exhibit dynamic behavior without need to modify their network weights. Similarly this has been shown in a filtering context [e.g. Feldkamp and Puskorius, 1998], which is directly related to simple multi-timescale optimizers [Sutton, 1992, Schraudolph, 1999]. Finally, the work of Younger et al. [2001] and Hochreiter et al. [2001] connects these different threads of research by allowing for the output of backpropagation from one network to feed into an additional learning network, with both networks trained jointly. Our approach to meta-learning builds on this work by modifying the network architecture of the optimizer in order to scale this approach to larger neural-network optimization problems. 2 Learning to learn with recurrent neural networks In this work we consider directly parameterizing the optimizer. As a result, in a slight abuse of notation we will write the final optimizee parameters (f;)as a function of the optimizer parameters and

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

8

we will write the final optimizee parameters (f;)as a function of the optimizer parameters and the function in question. We can then ask the question: What does it mean for an optimizer to be good? Given a distribution of functions fwe will write the expected loss as L() =Efh f (f;)
i : (2) As noted earlier, we will take the update steps gtto be the output of a recurrent neural network m, parameterized by , whose state we will denote explicitly with ht. Next, while the objective function in(2)depends only on the final parameter value, for training the optimizer it will be convenient to have an objective that depends on the entire trajectory of optimization, for some horizon T, L() =Ef"TX t=1wtf(t)# where t+1=t+gt; gt ht+1 =m(rt;ht;):(3) Herewt2R0are arbitrary weights associated with each time-step and we will also use the notation rt=rf(t). This formulation is equivalent to (2)whenwt=1[t=T], but later we will describe

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

9

rt=rf(t). This formulation is equivalent to (2)whenwt=1[t=T], but later we will describe why using different weights can prove useful. We can minimize the value of L()using gradient descent on . The gradient estimate @L()=@ can be computed by sampling a random function fand applying backpropagation to the computational graph in Figure 2. We allow gradients to flow along the solid edges in the graph, but gradients along the dashed edges are dropped. Ignoring gradients along the dashed edges amounts to making the assumption that the gradients of the optimizee do not depend on the optimizer parameters, i.e. @rt @= 0. This assumption allows us to avoid computing second derivatives of f. Examining the objective in (3)we see that the gradient is non-zero only for terms where wt6= 0. If we usewt=1[t=T]to match the original problem, then gradients of trajectory prefixes are zero and only the final optimization step provides information for training the optimizer. This renders Backpropagation Through Time (BPTT) inefficient. We solve this problem by relaxing the objective

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

10

Backpropagation Through Time (BPTT) inefficient. We solve this problem by relaxing the objective such thatwt>0at intermediate points along the trajectory. This changes the objective function, but allows us to train the optimizer on partial trajectories. For simplicity, in all our experiments we use wt= 1for everyt. 3 OptimizeeOptimizert-2t-1t mmm+++ft-1ftft-2∇t-2∇t-1∇tht-2ht-1htht+1gt-1gtθt-2θt-1θtθt+1gt-2Figure 2: Computational graph used for computing the gradient of the optimizer. 2.1 Coordinatewise LSTM optimizer One challenge in applying RNNs in our setting is that we want to be able to optimize at least tens of thousands of parameters. Optimizing at this scale with a fully connected RNN is not feasible as it would require a huge hidden state and an enormous number of parameters. To avoid this difficulty we will use an optimizer mwhich operates coordinatewise on the parameters of the objective function, similar to other common update rules like RMSprop and ADAM. This coordinatewise network

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

11

will use an optimizer mwhich operates coordinatewise on the parameters of the objective function, similar to other common update rules like RMSprop and ADAM. This coordinatewise network architecture allows us to use a very small network that only looks at a single coordinate to define the optimizer and share optimizer parameters across different parameters of the optimizee. Different behavior on each coordinate is achieved by using separate activations for each objective function parameter. In addition to allowing us to use a small network for this optimizer, this setup has the nice effect of making the optimizer invariant to the order of parameters in the network, since the same update rule is used independently on each coordinate. LSTM1fLSTMn∇1θ1θn++ …∇n……… … Figure 3: One step of an LSTM optimizer. All LSTMs have shared parameters, but separate hidden states.We implement the update rule for each coordinate using a two-layer Long Short Term Memory (LSTM) network [Hochreiter and Schmidhuber, 1997], using the now-standard forget gate architecture. The network takes as input the optimizee gradient for a single coordinate as well as the previous hidden state and outputs the update for the corresponding optimizee parameter. We will refer to this architecture, illustrated in

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

12

as the previous hidden state and outputs the update for the corresponding optimizee parameter. We will refer to this architecture, illustrated in Figure 3, as an LSTM optimizer. The use of recurrence allows the LSTM to learn dynamic update rules which integrate information from the history of gradients, similar to momentum. This is known to have many desirable properties in convex optimization [see e.g. Nesterov, 1983] and in fact many recent learning procedures—such as ADAM—use momentum in their updates. Preprocessing and postprocessing Optimizer inputs and outputs can have very different magnitudes depending on the class of function being optimized, but neural networks usually work robustly only for inputs and outputs which are neither very small nor very large. In practice rescaling inputs and outputs of an LSTM optimizer using suitable constants (shared across all timesteps and functions f) is sufficient to avoid this problem. In Appendix A we propose a different method of preprocessing inputs to the optimizer inputs which is more robust and gives slightly better performance. 4 20 40 60 80 10010-1100101LossQuadratics ADAM RMSprop SGD NAG LSTM 20 40 60 80 100 Step100MNIST

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

13

ADAM RMSprop SGD NAG LSTM 20 40 60 80 100 Step100MNIST 120 140 160 180 200MNIST, 200 stepsFigure 4: Comparisons between learned and hand-crafted optimizers performance. Learned optimizers are shown with solid lines and hand-crafted optimizers are shown with dashed lines. Units for the yaxis in the MNIST plots are logits. Left: Performance of different optimizers on randomly sampled 10-dimensional quadratic functions. Center: the LSTM optimizer outperforms standard methods training the base network on MNIST. Right: Learning curves for steps 100-200 by an optimizer trained to optimize for 100 steps (continuation of center plot). 3 Experiments In all experiments the trained optimizers use two-layer LSTMs with 20 hidden units in each layer. Each optimizer is trained by minimizing Equation 3 using truncated BPTT as described in Section 2. The minimization is performed using ADAM with a learning rate chosen by random search. We use early stopping when training the optimizer in order to avoid overfitting the optimizer. After each epoch (some fixed number of learning steps) we freeze the optimizer parameters and evaluate its performance. We pick the best optimizer (according to the final validation loss) and report its average

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

14

performance. We pick the best optimizer (according to the final validation loss) and report its average performance on a number of freshly sampled test problems. We compare our trained optimizers with standard optimizers used in Deep Learning: SGD, RMSprop, ADAM, and Nesterov’s accelerated gradient (NAG). For each of these optimizer and each problem we tuned the learning rate, and report results with the rate that gives the best final error for each problem. When an optimizer has more parameters than just a learning rate (e.g. decay coefficients for ADAM) we use the default values from the optim package in Torch7. Initial values of all optimizee parameters were sampled from an IID Gaussian distribution. 3.1 Quadratic functions In this experiment we consider training an optimizer on a simple class of synthetic 10-dimensional quadratic functions. In particular we consider minimizing functions of the form f() =kWyk2 2 for different 10x10 matrices Wand 10-dimensional vectors ywhose elements are drawn from an IID Gaussian distribution. Optimizers were trained by optimizing random functions from this family and tested on newly sampled functions from the same distribution. Each function was optimized for 100

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

15

Gaussian distribution. Optimizers were trained by optimizing random functions from this family and tested on newly sampled functions from the same distribution. Each function was optimized for 100 steps and the trained optimizers were unrolled for 20 steps. We have not used any preprocessing, nor postprocessing. Learning curves for different optimizers, averaged over many functions, are shown in the left plot of Figure 4. Each curve corresponds to the average performance of one optimization algorithm on many test functions; the solid curve shows the learned optimizer performance and dashed curves show the performance of the standard baseline optimizers. It is clear the learned optimizers substantially outperform the baselines in this setting. 3.2 Training a small neural network on MNIST In this experiment we test whether trainable optimizers can learn to optimize a small neural network on MNIST, and also explore how the trained optimizers generalize to functions beyond those they were trained on. To this end, we train the optimizer to optimize a base network and explore a series of modifications to the network architecture and training procedure at test time. 5 20 40 60 80 100100LossMNIST, 40 units ADAM RMSprop SGD NAG LSTM 20 40 60 80 100 StepsMNIST, 2 layers

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

16

ADAM RMSprop SGD NAG LSTM 20 40 60 80 100 StepsMNIST, 2 layers 20 40 60 80 100MNIST, ReLUFigure 5: Comparisons between learned and hand-crafted optimizers performance. Units for the yaxis are logits. Left: Generalization to the different number of hidden units (40 instead of 20). Center: Generalization to the different number of hidden layers (2 instead of 1). This optimization problem is very hard, because the hidden layers are very narrow. Right: Training curves for an MLP with 20 hidden units using ReLU activations. The LSTM optimizer was trained on an MLP with sigmoid activations. 5 20 35 5010-1100Final Loss LSTM 5 20 35 50 Hidden units ADAM 5 20 35 50 NAG Layers 1 2 Figure 6: Systematic study of final MNIST performance as the optimizee architecture is varied, using sigmoid non-linearities. The vertical dashed line in the left-most plot denotes the architecture at which the LSTM is trained and the horizontal line shows the final performance of the trained optimizer in this setting. In this setting the objective function f()is the cross entropy of a small MLP with parameters .

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

17

optimizer in this setting. In this setting the objective function f()is the cross entropy of a small MLP with parameters . The values of fas well as the gradients @f()=@are estimated using random minibatches of 128 examples. The base network is an MLP with one hidden layer of 20 units using a sigmoid activation function. The only source of variability between different runs is the initial value 0and randomness in minibatch selection. Each optimization was run for 100 steps and the trained optimizers were unrolled for 20 steps. We used input preprocessing described in Appendix A and rescaled the outputs of the LSTM by the factor 0:1. Learning curves for the base network using different optimizers are displayed in the center plot of Figure 4. In this experiment NAG, ADAM, and RMSprop exhibit roughly equivalent performance the LSTM optimizer outperforms them by a significant margin. The right plot in Figure 4 compares the performance of the LSTM optimizer if it is allowed to run for 200 steps, despite having been trained to optimize for 100 steps. In this comparison we re-used the LSTM optimizer from the previous experiment, and here we see that the LSTM optimizer continues to outperform the baseline optimizers on this task.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

18

experiment, and here we see that the LSTM optimizer continues to outperform the baseline optimizers on this task. Generalization to different architectures Figure 5 shows three examples of applying the LSTM optimizer to train networks with different architectures than the base network on which it was trained. The modifications are (from left to right) (1) an MLP with 40 hidden units instead of 20, (2) a network with two hidden layers instead of one, and (3) a network using ReLU activations instead of sigmoid. In the first two cases the LSTM optimizer generalizes well, and continues to outperform the hand-designed baselines despite operating outside of its training regime. However, changing the activation function to ReLU makes the dynamics of the learning procedure sufficiently different that the learned optimizer is no longer able to generalize. Finally, in Figure 6 we show the results of systematically varying the tested architecture; for the LSTM results we again used the optimizer trained using 1 layer of 20 units and sigmoid non-linearities. Note that in this setting where the 6 200 400 600 800 1000100LossCIFAR-10 200 400 600 800 1000 Step100CIFAR-5 200 400 600 800 100010-1CIFAR-2

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

19

200 400 600 800 1000 Step100CIFAR-5 200 400 600 800 100010-1CIFAR-2 ADAM RMSprop SGD NAG LSTM LSTM-subFigure 7: Optimization performance on the CIFAR-10 dataset and subsets. Shown on the left is the LSTM optimizer versus various baselines trained on CIFAR-10 and tested on a held-out test set. The two plots on the right are the performance of these optimizers on subsets of the CIFAR labels. The additional optimizer LSTM-sub has been trained only on the heldout labels and is hence transferring to a completely novel dataset. test-set problems are similar enough to those in the training set we see even better generalization than the baseline optimizers. 3.3 Training a convolutional network on CIFAR-10 Next we test the performance of the trained neural optimizers on optimizing classification performance for the CIFAR-10 dataset [Krizhevsky, 2009]. In these experiments we used a model with both convolutional and feed-forward layers. In particular, the model used for these experiments includes three convolutional layers with max pooling followed by a fully-connected layer with 32 hidden units;

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

20

three convolutional layers with max pooling followed by a fully-connected layer with 32 hidden units; all non-linearities were ReLU activations with batch normalization. The coordinatewise network decomposition introduced in Section 2.1—and used in the previous experiment—utilizes a single LSTM architecture with shared weights, but separate hidden states, for each optimizee parameter. We found that this decomposition was not sufficient for the model architecture introduced in this section due to the differences between the fully connected and convolutional layers. Instead we modify the optimizer by introducing two LSTMs: one proposes parameter updates for the fully connected layers and the other updates the convolutional layer parameters. Like the previous LSTM optimizer we still utilize a coordinatewise decomposition with shared weights and individual hidden states, however LSTM weights are now shared only between parameters of the same type (i.e. fully-connected vs. convolutional). The performance of this trained optimizer compared against the baseline techniques is shown in Figure 7. The left-most plot displays the results of using the optimizer to fit a classifier on a held-out test set. The additional two plots on the right display the performance of the trained optimizer on

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

21

test set. The additional two plots on the right display the performance of the trained optimizer on modified datasets which only contain a subset of the labels, i.e. the CIFAR-2 dataset only contains data corresponding to 2 of the 10 labels. Additionally we include an optimizer LSTM-sub which was only trained on the held-out labels. In all these examples we can see that the LSTM optimizer learns much more quickly than the baseline optimizers, with significant boosts in performance for the CIFAR-5 and especially CIFAR-2 datsets. We also see that the optimizers trained only on a disjoint subset of the data is hardly effected by this difference and transfers well to the additional dataset. 3.4 Neural Art The recent work on artistic style transfer using convolutional networks, or Neural Art [Gatys et al., 2015], gives a natural testbed for our method, since each content and style image pair gives rise to a different optimization problem. Each Neural Art problem starts from a content image ,c, and a style image ,s, and is given by f() =Lcontent (c;) +Lstyle(s;) + Lreg()

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

22

image ,s, and is given by f() =Lcontent (c;) +Lstyle(s;) + Lreg() The minimizer of fis the styled image . The first two terms try to match the content and style of the styled image to that of their first argument, and the third term is a regularizer that encourages smoothness in the styled image. Details can be found in [Gatys et al., 2015]. 7 20 40 60 80 100 120 StepLossNeural art, training resolution 20 40 60 80 100 120 StepDouble resolution ADAM RMSprop SGD NAG LSTMFigure 8: Optimization curves for Neural Art. Content images come from the test set, which was not used during the LSTM optimizer training. Note: the y-axis is in log scale and we zoom in on the interesting portion of this plot. Left: Applying the training style at the training resolution. Right: Applying the test style at double the training resolution. Figure 9: Examples of images styled using the LSTM optimizer. Each triple consists of the content image (left), style (right) and image generated by the LSTM optimizer (center). Left: The result of

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

23

image (left), style (right) and image generated by the LSTM optimizer (center). Left: The result of applying the training style at the training resolution to a test image. Right: The result of applying a new style to a test image at double the resolution on which the optimizer was trained. We train optimizers using only 1 style and 1800 content images taken from ImageNet [Deng et al., 2009]. We randomly select 100 content images for testing and 20 content images for validation of trained optimizers. We train the optimizer on 64x64 content images from ImageNet and one fixed style image. We then test how well it generalizes to a different style image and higher resolution (128x128). Each image was optimized for 128 steps and trained optimizers were unrolled for 32 steps. Figure 9 shows the result of styling two different images using the LSTM optimizer. The LSTM optimizer uses inputs preprocessing described in Appendix A and no postprocessing. See Appendix C for additional images. Figure 8 compares the performance of the LSTM optimizer to standard optimization algorithms. The LSTM optimizer outperforms all standard optimizers if the resolution and style image are the same as the ones on which it was trained. Moreover, it continues to perform very well when both the resolution and style are changed at test time.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

24

as the ones on which it was trained. Moreover, it continues to perform very well when both the resolution and style are changed at test time. Finally, in Appendix B we qualitatively examine the behavior of the step directions generated by the learned optimizer. 4 Conclusion We have shown how to cast the design of optimization algorithms as a learning problem, which enables us to train optimizers that are specialized to particular classes of functions. Our experiments have confirmed that learned neural optimizers compare favorably against state-of-the-art optimization methods used in deep learning. We witnessed a remarkable degree of transfer, with for example the LSTM optimizer trained on 12,288 parameter neural art tasks being able to generalize to tasks with 49,152 parameters, different styles, and different content images all at the same time. We observed similar impressive results when transferring to different architectures in the MNIST task. The results on the CIFAR image labeling task show that the LSTM optimizers outperform handengineered optimizers when transferring to datasets drawn from the same data distribution. References F. Bach, R. Jenatton, J. Mairal, and G. Obozinski. Optimization with sparsity-inducing penalties. Foundations and Trends in Machine Learning , 4(1):1–106, 2012. 8

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

25

and Trends in Machine Learning , 4(1):1–106, 2012. 8 S. Bengio, Y . Bengio, and J. Cloutier. On the search for new learning rules for ANNs. Neural Processing Letters , 2(4):26–30, 1995. Y . Bengio, S. Bengio, and J. Cloutier. Learning a synaptic learning rule . Université de Montréal, Département d’informatique et de recherche opérationnelle, 1990. Y . Bengio, N. Boulanger-Lewandowski, and R. Pascanu. Advances in optimizing recurrent networks. In International Conference on Acoustics, Speech and Signal Processing , pages 8624–8628. IEEE, 2013. F. Bobolas. brain-neurons, 2009. URL https://www.flickr.com/photos/fbobolas/3822222947 . Creative Commons Attribution-ShareAlike 2.0 Generic. N. E. Cotter and P. R. Conwell. Fixed-weight networks can learn. In International Joint Conference on Neural Networks , pages 553–559, 1990. C. Daniel, J. Taylor, and S. Nowozin. Learning step size controllers for robust neural network training. In

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

26

Networks , pages 553–559, 1990. C. Daniel, J. Taylor, and S. Nowozin. Learning step size controllers for robust neural network training. In Association for the Advancement of Artificial Intelligence , 2016. J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. InComputer Vision and Pattern Recognition , pages 248–255. IEEE, 2009. D. L. Donoho. Compressed sensing. Transactions on Information Theory , 52(4):1289–1306, 2006. J. Duchi, E. Hazan, and Y . Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research , 12:2121–2159, 2011. L. A. Feldkamp and G. V . Puskorius. A signal processing framework based on dynamic neural networks with application to problems in adaptation, filtering, and classification. Proceedings of the IEEE , 86(11): 2259–2277, 1998. L. A. Gatys, A. S. Ecker, and M. Bethge. A neural algorithm of artistic style. arXiv Report 1508.06576, 2015.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

27

L. A. Gatys, A. S. Ecker, and M. Bethge. A neural algorithm of artistic style. arXiv Report 1508.06576, 2015. A. Graves, G. Wayne, and I. Danihkela. Neural Turing machines. arXiv Report 1410.5401, 2014. S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation , 9(8):1735–1780, 1997. S. Hochreiter, A. S. Younger, and P. R. Conwell. Learning to learn using gradient descent. In International Conference on Artificial Neural Networks , pages 87–94. Springer, 2001. D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations , 2015. A. Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. B. M. Lake, T. D. Ullman, J. B. Tenenbaum, and S. J. Gershman. Building machines that learn and think like people. arXiv Report 1604.00289, 2016.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

28

people. arXiv Report 1604.00289, 2016. T. Maley. neuron, 2011. URL https://www.flickr.com/photos/taylortotz101/6280077898 . Creative Commons Attribution 2.0 Generic. J. Martens and R. Grosse. Optimizing neural networks with Kronecker-factored approximate curvature. In International Conference on Machine Learning , pages 2408–2417, 2015. G. L. Nemhauser and L. A. Wolsey. Integer and combinatorial optimization . John Wiley & Sons, 1988. Y . Nesterov. A method of solving a convex programming problem with convergence rate o (1/k2). In Soviet Mathematics Doklady , volume 27, pages 372–376, 1983. J. Nocedal and S. Wright. Numerical optimization . Springer Science & Business Media, 2006. M. Riedmiller and H. Braun. A direct adaptive method for faster backpropagation learning: The RPROP algorithm. In International Conference on Neural Networks , pages 586–591, 1993. T. P. Runarsson and M. T. Jonsson. Evolution and design of distributed learning rules. In IEEE Symposium on

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

29

T. P. Runarsson and M. T. Jonsson. Evolution and design of distributed learning rules. In IEEE Symposium on Combinations of Evolutionary Computation and Neural Networks , pages 59–63. IEEE, 2000. A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap. Meta-learning with memory-augmented neural networks. In International Conference on Machine Learning , 2016. J. Schmidhuber. Evolutionary principles in self-referential learning; On learning how to learn: The meta-meta-... hook. PhD thesis, Institut f. Informatik, Tech. Univ. Munich, 1987. J. Schmidhuber. Learning to control fast-weight memories: An alternative to dynamic recurrent networks. Neural Computation , 4(1):131–139, 1992. J. Schmidhuber. A neural network that embeds its own meta-levels. In International Conference on Neural Networks , pages 407–412. IEEE, 1993. J. Schmidhuber, J. Zhao, and M. Wiering. Shifting inductive bias with success-story algorithm, adaptive levin search, and incremental self-improvement. Machine Learning , 28(1):105–130, 1997.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

30

search, and incremental self-improvement. Machine Learning , 28(1):105–130, 1997. N. N. Schraudolph. Local gain adaptation in stochastic gradient descent. In International Conference on Artificial Neural Networks , volume 2, pages 569–574, 1999. R. S. Sutton. Adapting bias by gradient descent: An incremental version of delta-bar-delta. In Association for the Advancement of Artificial Intelligence , pages 171–176, 1992. S. Thrun and L. Pratt. Learning to learn . Springer Science & Business Media, 1998. T. Tieleman and G. Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine Learning , 4:2, 2012. P. Tseng. An incremental gradient (-projection) method with momentum term and adaptive stepsize rule. Journal on Optimization , 8(2):506–531, 1998. D. H. Wolpert and W. G. Macready. No free lunch theorems for optimization. Transactions on Evolutionary Computation , 1(1):67–82, 1997. 9

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

31

Computation , 1(1):67–82, 1997. 9 A. S. Younger, P. R. Conwell, and N. E. Cotter. Fixed-weight on-line learning. Transactions on Neural Networks , 10(2):272–283, 1999. A. S. Younger, S. Hochreiter, and P. R. Conwell. Meta-learning with backpropagation. In International Joint Conference on Neural Networks , 2001. 10 0 64 128−8−24 0 64 128−2616 LSTM ADAM SGDFigure 10: Updates proposed by different optimizers at different optimization steps for two different coordinates. A Gradient preprocessing One potential challenge in training optimizers is that different input coordinates (i.e. the gradients w.r.t. different optimizee parameters) can have very different magnitudes. This is indeed the case e.g. when the optimizee is a neural network and different parameters correspond to weights in different layers. This can make training an optimizer difficult, because neural networks naturally disregard small variations in input signals and concentrate on bigger input values. To this aim we propose to preprocess the optimizer’s inputs. One solution would be to give the

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

32

small variations in input signals and concentrate on bigger input values. To this aim we propose to preprocess the optimizer’s inputs. One solution would be to give the optimizer (log(jrj);sgn(r))as an input, whereris the gradient in the current timestep. This has a problem that log(jrj)diverges forr! 0. Therefore, we use the following preprocessing formula rk!( log(jrj) p;sgn(r) ifjrjep (1;epr) otherwise wherep>0is a parameter controlling how small gradients are disregarded (we use p= 10 in all our experiments). We noticed that just rescaling all inputs by an appropriate constant instead also works fine, but the proposed preprocessing seems to be more robust and gives slightly better results on some problems. B Visualizations Visualizing optimizers is inherently difficult because their proposed updates are functions of the full optimization trajectory. In this section we try to peek into the decisions made by the LSTM optimizer, trained on the neural art task. Histories of updates We select a single optimizee parameter (one color channel of one pixel in the

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

33

trained on the neural art task. Histories of updates We select a single optimizee parameter (one color channel of one pixel in the styled image) and trace the updates proposed to this coordinate by the LSTM optimizer over a single trajectory of optimization. We also record the updates that would have been proposed by both SGD and ADAM if they followed the same trajectory of iterates. Figure 10 shows the trajectory of updates for two different optimizee parameters. From the plots it is clear that the trained optimizer makes bigger updates than SGD and ADAM. It is also visible that it uses some kind of momentum, but its updates are more noisy than those proposed by ADAM which may be interpreted as having a shorter time-scale momentum. Proposed update as a function of current gradient Another way to visualize the optimizer behavior is to look at the proposed update gtfor a single coordinate as a function of the current gradient evaluation rt. We follow the same procedure as in the previous experiment, and visualize the proposed updates for a few selected time steps. These results are shown in Figures 11–13. In these plots, the x-axis is the current value of the gradient for the chosen coordinate, and the y-axis shows the update that each optimizer would propose should the corresponding gradient value be observed. The history of gradient observations is the same for all

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

34

for the chosen coordinate, and the y-axis shows the update that each optimizer would propose should the corresponding gradient value be observed. The history of gradient observations is the same for all methods and follows the trajectory of the LSTM optimizer. 11 The shape of this function for the LSTM optimizer is often step-like, which is also the case for ADAM. Surprisingly the step is sometimes in the opposite direction as for ADAM, i.e. the bigger the gradient, the bigger the update. C Neural Art Shown below are additional examples of images styled using the LSTM optimizer. Each triple consists of the content image (left), style (right) and image generated by the LSTM optimizer (center). D Information sharing between coordinates In previous sections we considered a coordinatewise architecture, which corresponds by analogy to a learned version of RMSprop or ADAM. Although diagonal methods are quite effective in practice, we can also consider learning more sophisticated optimizers that take the correlations between coordinates into effect. To this end, we introduce a mechanism allowing different LSTMs to communicate with each other. Global averaging cells The simplest solution is to designate a subset of the cells in each LSTM layer for communication. These cells operate like normal LSTM cells, but their outgoing activations are averaged at each step

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

35

These cells operate like normal LSTM cells, but their outgoing activations are averaged at each step across all coordinates. These global averaging cells (GACs) are sufficient to allow the networks to implement L2 gradient clipping [Bengio et al., 2013] assuming that each LSTM can compute the square of the gradient. This architecture is denoted as an LSTM+GAC optimizer. NTM-BFGS optimizer We also consider augmenting the LSTM+GAC architecture with an external memory that is shared between coordinates. Such a memory, if appropriately designed could allow the optimizer to learn algorithms similar to (low-memory) approximations to Newton’s method, e.g. (L-)BFGS [see Nocedal and Wright, 2006]. The reason for this interpretation is that such methods can be seen as a set of independent processes working coordinatewise, but communicating through the inverse Hessian approximation stored in the memory. We designed a memory architecture that, in theory, allows the 12 Step 1−10010 Step 2 Step 3 Step 4 Step 5−10010 Step 6 Step 7 Step 8 Step 9−10010 Step 10 Step 11 Step 12 Step 13−10010 Step 14 Step 15 Step 16 Step 17−10010

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

36

Step 6 Step 7 Step 8 Step 9−10010 Step 10 Step 11 Step 12 Step 13−10010 Step 14 Step 15 Step 16 Step 17−10010 Step 18 Step 19 Step 20 Step 21−10010 Step 22 Step 23 Step 24 Step 25−10010 Step 26 Step 27 Step 28 −400 0 400 Step 29−10010 −400 0 400 Step 30−400 0 400 Step 31−400 0 400 Step 32Figure 11: The proposed update direction for a single coordinate over 32 steps. 13 Step 1−10010 Step 2 Step 3 Step 4 Step 5−10010 Step 6 Step 7 Step 8 Step 9−10010 Step 10 Step 11 Step 12 Step 13−10010 Step 14 Step 15 Step 16 Step 17−10010 Step 18 Step 19 Step 20 Step 21−10010 Step 22 Step 23 Step 24 Step 25−10010 Step 26 Step 27 Step 28 −400 0 400 Step 29−10010 −400 0 400 Step 30−400 0 400 Step 31−400 0 400 Step 32Figure 12: The proposed update direction for a single coordinate over 32 steps. 14 Step 1−10010 Step 2 Step 3 Step 4 Step 5−10010 Step 6 Step 7 Step 8 Step 9−10010

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

37

14 Step 1−10010 Step 2 Step 3 Step 4 Step 5−10010 Step 6 Step 7 Step 8 Step 9−10010 Step 10 Step 11 Step 12 Step 13−10010 Step 14 Step 15 Step 16 Step 17−10010 Step 18 Step 19 Step 20 Step 21−10010 Step 22 Step 23 Step 24 Step 25−10010 Step 26 Step 27 Step 28 −400 0 400 Step 29−10010 −400 0 400 Step 30−400 0 400 Step 31−400 0 400 Step 32Figure 13: The proposed update direction for a single coordinate over 32 steps. 15 LSTMLSTMLSTMLSTMLSTMLSTMLSTMr1MLSTMr2r3…………rni1i2i3…………inx= LSTMLSTMLSTMLSTMLSTMLSTMLSTMLSTMa1a2a3…………anb1b2b3…………bnOuterProductMtΔMtMt+1+=Figure 14: Left: NTM-BFGS read operation. Right: NTM-BFGS write operation. network to simulate (L-)BFGS, motivated by the approximate Newton method BFGS, named for Broyden, Fletcher, Goldfarb, and Shanno. We call this architecture an NTM-BFGS optimizer , because

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

38

Broyden, Fletcher, Goldfarb, and Shanno. We call this architecture an NTM-BFGS optimizer , because its use of external memory is similar to the Neural Turing Machine [Graves et al., 2014]. The pivotal differences between our construction and the NTM are (1) our memory allows only low-rank updates; (2) the controller (including read/write heads) operates coordinatewise. In BFGS an explicit estimate of the full (inverse) Hessian is built up from the sequence of observed gradients. We can write a skeletonized version of the BFGS algorithm, using Mtto represent the inverse Hessian approximation at iteration t, as follows gt= read(Mt;t) t+1=t+gt Mt+1= write(Mt;t;gt): Here we have packed up all of the details of the BFGS algorithm into the suggestively named read andwrite operations, which operate on the inverse Hessian approximation Mt. In BFGS these operations have specific forms, for example read(Mt;t) =Mtrf(t)is a specific matrix-vector multiplication and the BFGS write operation corresponds to a particular low-rank update of Mt.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

39

multiplication and the BFGS write operation corresponds to a particular low-rank update of Mt. In this work we preserve the structure of the BFGS updates, but discard their particular form. More specifically the read operation remains a matrix-vector multiplication but the form of the vector used is learned. Similarly, the write operation remains a low-rank update, but the vectors involved are also learned. Conveniently, this structure of interaction with a large dynamically updated state corresponds in a fairly direct way to the architecture of a Neural Turing Machine (NTM), where Mt corresponds to the NTM memory [Graves et al., 2014]. Our NTM-BFGS optimizer uses an LSTM+GAC as a controller; however, instead of producing the update directly we attach one or more read and write heads to the controller. Each read head produces a read vector rtwhich is combined with the memory to produce a read result itwhich is fed back into the controller at the following time step. Each write head produces two outputs, a left write vectoratand a right write vector bt. The two write vectors are used to update the memory state by accumulating their outer product. The read and write operation for a single head is diagrammed in Figure 14 and the way read and write heads are attached to the controller is depicted in Figure 15.

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

40

Figure 14 and the way read and write heads are attached to the controller is depicted in Figure 15. In can be shown that NTM-BFGS with one read head and 3 write heads can simulate inverse Hessian BFGS assuming that the controller can compute arbitrary (coordinatewise) functions and have access to 2 GACs. NTM-L-BFGS optimizer In cases where memory is constrained we can follow the example of L-BFGS and maintain a low rank approximation of the full memory ( vis. inverse Hessian). The simplest way to do this is to store a sliding history of the left and right write vectors, allowing us to form the matrix vector multiplication required by the read operation efficiently. 16 Controller ReadWriteMtrtht+1Mt+1rt+1itatbt∇thtLSTM1LSTM2…gt LSTMk∇kikakbkrkhkhkgkFigure 15: Left: Interaction between the controller and the external memory in NTM-BFGS. The controller is composed of replicated coordinatewise LSTMs (possibly with GACs), but the read and write operations are global across all coordinates. Right: A single LSTM for the kth coordinate in

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1606.04474

41

write operations are global across all coordinates. Right: A single LSTM for the kth coordinate in the NTM-BFGS controller. Note that here we have dropped the time index tto simplify notation. 17

1606.04474

Learning to learn by gradient descent by gradient descent

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

http://arxiv.org/pdf/1606.04474

cs.NE

20160614

20161130

1607.06450

0

Layer Normalization Jimmy Lei Ba University of Toronto jimmy@psi.toronto.eduJamie Ryan Kiros University of Toronto rkiros@cs.toronto.eduGeoffrey E. Hinton University of Toronto and Google Inc. hinton@cs.toronto.edu Abstract Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feedforward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

1

we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques. 1 Introduction Deep neural networks trained with some version of Stochastic Gradient Descent have been shown to substantially outperform previous approaches on various supervised learning tasks in computer vision [Krizhevsky et al., 2012] and speech processing [Hinton et al., 2012]. But state-of-the-art deep neural networks often require many days of training. It is possible to speed-up the learning by computing gradients for different subsets of the training cases on different machines or splitting the neural network itself over many machines [Dean et al., 2012], but this can require a lot of communication and complex software. It also tends to lead to rapidly diminishing returns as the degree of parallelization increases. An orthogonal approach is to modify the computations performed in the forward pass of the neural net to make learning easier. Recently, batch normalization [Ioffe and

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

2

the forward pass of the neural net to make learning easier. Recently, batch normalization [Ioffe and Szegedy, 2015] has been proposed to reduce training time by including additional normalization stages in deep neural networks. The normalization standardizes each summed input using its mean and its standard deviation across the training data. Feedforward neural networks trained using batch normalization converge faster even with simple SGD. In addition to training time improvement, the stochasticity from the batch statistics serves as a regularizer during training. Despite its simplicity, batch normalization requires running averages of the summed input statistics. In feed-forward networks with fixed depth, it is straightforward to store the statistics separately for each hidden layer. However, the summed inputs to the recurrent neurons in a recurrent neural network (RNN) often vary with the length of the sequence so applying batch normalization to RNNs appears to require different statistics for different time-steps. Furthermore, batch normaliza-arXiv:1607.06450v1 [stat.ML] 21 Jul 2016 tion cannot be applied to online learning tasks or to extremely large distributed models where the minibatches have to be small. This paper introduces layer normalization, a simple normalization method to improve the training speed for various neural network models. Unlike batch normalization, the proposed method directly

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

3

This paper introduces layer normalization, a simple normalization method to improve the training speed for various neural network models. Unlike batch normalization, the proposed method directly estimates the normalization statistics from the summed inputs to the neurons within a hidden layer so the normalization does not introduce any new dependencies between training cases. We show that layer normalization works well for RNNs and improves both the training time and the generalization performance of several existing RNN models. 2 Background A feed-forward neural network is a non-linear mapping from a input pattern xto an output vector y. Consider the lthhidden layer in a deep feed-forward, neural network, and let albe the vector representation of the summed inputs to the neurons in that layer. The summed inputs are computed through a linear projection with the weight matrix Wland the bottom-up inputs hlgiven as follows: al i=wl i>hlhl+1 i=f(al i+bl i) (1) wheref(
)is an element-wise non-linear function and wl iis the incoming weights to the ithhidden units andbl iis the scalar bias parameter. The parameters in the neural network are learnt using gradient-based optimization algorithms with the gradients being computed by back-propagation.

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

4

units andbl iis the scalar bias parameter. The parameters in the neural network are learnt using gradient-based optimization algorithms with the gradients being computed by back-propagation. One of the challenges of deep learning is that the gradients with respect to the weights in one layer are highly dependent on the outputs of the neurons in the previous layer especially if these outputs change in a highly correlated way. Batch normalization [Ioffe and Szegedy, 2015] was proposed to reduce such undesirable “covariate shift”. The method normalizes the summed inputs to each hidden unit over the training cases. Specifically, for the ithsummed input in the lthlayer, the batch normalization method rescales the summed inputs according to their variances under the distribution of the data al i=gl i l i al il i
l i=E xP(x) al i l i=r E xP(x)h al il i
2i (2) where al

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

5

l i=r E xP(x)h al il i
2i (2) where al iis normalized summed inputs to the ithhidden unit in the lthlayer andgiis a gain parameter scaling the normalized activation before the non-linear activation function. Note the expectation is under the whole training data distribution. It is typically impractical to compute the expectations in Eq. (2) exactly, since it would require forward passes through the whole training dataset with the current set of weights. Instead, andare estimated using the empirical samples from the current mini-batch. This puts constraints on the size of a mini-batch and it is hard to apply to recurrent neural networks. 3 Layer normalization We now consider the layer normalization method which is designed to overcome the drawbacks of batch normalization. Notice that changes in the output of one layer will tend to cause highly correlated changes in the summed inputs to the next layer, especially with ReLU units whose outputs can change by a lot. This suggests the “covariate shift” problem can be reduced by fixing the mean and the variance of the summed inputs within each layer. We, thus, compute the layer normalization statistics over all the hidden units in the same layer as follows: l=1

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

6

the summed inputs within each layer. We, thus, compute the layer normalization statistics over all the hidden units in the same layer as follows: l=1 HHX i=1al il=vuut1 HHX i=1 al il
2(3) whereHdenotes the number of hidden units in a layer. The difference between Eq. (2) and Eq. (3) is that under layer normalization, all the hidden units in a layer share the same normalization terms and, but different training cases have different normalization terms. Unlike batch normalization, layer normaliztion does not impose any constraint on the size of a mini-batch and it can be used in the pure online regime with batch size 1. 2 3.1 Layer normalized recurrent neural networks The recent sequence to sequence models [Sutskever et al., 2014] utilize compact recurrent neural networks to solve sequential prediction problems in natural language processing. It is common among the NLP tasks to have different sentence lengths for different training cases. This is easy to deal with in an RNN because the same weights are used at every time-step. But when we apply batch normalization to an RNN in the obvious way, we need to to compute and store separate statistics for

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

7

normalization to an RNN in the obvious way, we need to to compute and store separate statistics for each time step in a sequence. This is problematic if a test sequence is longer than any of the training sequences. Layer normalization does not have such problem because its normalization terms depend only on the summed inputs to a layer at the current time-step. It also has only one set of gain and bias parameters shared over all time-steps. In a standard RNN, the summed inputs in the recurrent layer are computed from the current input xtand previous vector of hidden states ht1which are computed as at=Whhht1+Wxhxt. The layer normalized recurrent layer re-centers and re-scales its activations using the extra normalization terms similar to Eq. (3): ht=fhg t att
+bi t=1 HHX i=1at it=vuut1 HHX i=1(at it)2(4) whereWhhis the recurrent hidden to hidden weights and Wxhare the bottom up input to hidden weights.is the element-wise multiplication between two vectors. bandgare defined as the bias and gain parameters of the same dimension as ht.

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

8

weights.is the element-wise multiplication between two vectors. bandgare defined as the bias and gain parameters of the same dimension as ht. In a standard RNN, there is a tendency for the average magnitude of the summed inputs to the recurrent units to either grow or shrink at every time-step, leading to exploding or vanishing gradients. In a layer normalized RNN, the normalization terms make it invariant to re-scaling all of the summed inputs to a layer, which results in much more stable hidden-to-hidden dynamics. 4 Related work Batch normalization has been previously extended to recurrent neural networks [Laurent et al., 2015, Amodei et al., 2015, Cooijmans et al., 2016]. The previous work [Cooijmans et al., 2016] suggests the best performance of recurrent batch normalization is obtained by keeping independent normalization statistics for each time-step. The authors show that initializing the gain parameter in the recurrent batch normalization layer to 0.1 makes significant difference in the final performance of the model. Our work is also related to weight normalization [Salimans and Kingma, 2016]. In weight normalization, instead of the variance, the L2 norm of the incoming weights is used to

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

9

weight normalization, instead of the variance, the L2 norm of the incoming weights is used to normalize the summed inputs to a neuron. Applying either weight normalization or batch normalization using expected statistics is equivalent to have a different parameterization of the original feed-forward neural network. Re-parameterization in the ReLU network was studied in the Pathnormalized SGD [Neyshabur et al., 2015]. Our proposed layer normalization method, however, is not a re-parameterization of the original neural network. The layer normalized model, thus, has different invariance properties than the other methods, that we will study in the following section. 5 Analysis In this section, we investigate the invariance properties of different normalization schemes. 5.1 Invariance under weights and data transformations The proposed layer normalization is related to batch normalization and weight normalization. Although, their normalization scalars are computed differently, these methods can be summarized as normalizing the summed inputs aito a neuron through the two scalars and. They also learn an adaptive bias band gaingfor each neuron after the normalization. hi=f(gi i(aii) +bi) (5) Note that for layer normalization and batch normalization, andis computed according to Eq. 2

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

10

i(aii) +bi) (5) Note that for layer normalization and batch normalization, andis computed according to Eq. 2 and 3. In weight normalization, is 0, and=kwk2. 3 Weight matrix Weight matrix Weight vector Dataset Dataset Single training case re-scaling re-centering re-scaling re-scaling re-centering re-scaling Batch norm Invariant No Invariant Invariant Invariant No Weight norm Invariant No Invariant No No No Layer norm Invariant Invariant No Invariant No Invariant Table 1: Invariance properties under the normalization methods. Table 1 highlights the following invariance results for three normalization methods. Weight re-scaling and re-centering: First, observe that under batch normalization and weight normalization, any re-scaling to the incoming weights wiof a single neuron has no effect on the normalized summed inputs to a neuron. To be precise, under batch and weight normalization, if the weight vector is scaled by , the two scalar andwill also be scaled by . The normalized summed inputs stays the same before and after scaling. So the batch and weight normalization are

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

11

summed inputs stays the same before and after scaling. So the batch and weight normalization are invariant to the re-scaling of the weights. Layer normalization, on the other hand, is not invariant to the individual scaling of the single weight vectors. Instead, layer normalization is invariant to scaling of the entire weight matrix and invariant to a shift to all of the incoming weights in the weight matrix. Let there be two sets of model parameters ,0whose weight matrices WandW0 differ by a scaling factor and all of the incoming weights in W0are also shifted by a constant vector , that isW0=W+1 >. Under layer normalization, the two models effectively compute the same output: h0=f(g 0(W0x0) +b) =f(g 0 (W+1 >)x0
+b) =f(g (Wx) +b) =h: (6) Notice that if normalization is only applied to the input before the weights, the model will not be invariant to re-scaling and re-centering of the weights. Data re-scaling and re-centering: We can show that all the normalization methods are invariant

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

12

Data re-scaling and re-centering: We can show that all the normalization methods are invariant to re-scaling the dataset by verifying that the summed inputs of neurons stays constant under the changes. Furthermore, layer normalization is invariant to re-scaling of individual training cases, because the normalization scalars andin Eq. (3) only depend on the current input data. Let x0 be a new data point obtained by re-scaling xby. Then we have, h0 i=f(gi 0 w> ix00
+bi) =f(gi  w> ix
+bi) =hi: (7) It is easy to see re-scaling individual data points does not change the model’s prediction under layer normalization. Similar to the re-centering of the weight matrix in layer normalization, we can also show that batch normalization is invariant to re-centering of the dataset. 5.2 Geometry of parameter space during learning We have investigated the invariance of the model’s prediction under re-centering and re-scaling of the parameters. Learning, however, can behave very differently under different parameterizations,

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

13

the parameters. Learning, however, can behave very differently under different parameterizations, even though the models express the same underlying function. In this section, we analyze learning behavior through the geometry and the manifold of the parameter space. We show that the normalization scalar can implicitly reduce learning rate and makes learning more stable. 5.2.1 Riemannian metric The learnable parameters in a statistical model form a smooth manifold that consists of all possible input-output relations of the model. For models whose output is a probability distribution, a natural way to measure the separation of two points on this manifold is the Kullback-Leibler divergence between their model output distributions. Under the KL divergence metric, the parameter space is a Riemannian manifold. The curvature of a Riemannian manifold is entirely captured by its Riemannian metric, whose quadratic form is denoted as ds2. That is the infinitesimal distance in the tangent space at a point in the parameter space. Intuitively, it measures the changes in the model output from the parameter space along a tangent direction. The Riemannian metric under KL was previously studied [Amari, 1998] and was shown to be well approximated under second order Taylor expansion using the Fisher 4 information matrix: ds2= D KL

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

14

1998] and was shown to be well approximated under second order Taylor expansion using the Fisher 4 information matrix: ds2= D KL P(yjx;)kP(yjx;+) 1 2>F(); (8) F() = E xP(x);yP(yjx)" @logP(yjx;) @@logP(yjx;) @># ; (9) where,is a small change to the parameters. The Riemannian metric above presents a geometric view of parameter spaces. The following analysis of the Riemannian metric provides some insight into how normalization methods could help in training neural networks. 5.2.2 The geometry of normalized generalized linear models We focus our geometric analysis on the generalized linear model. The results from the following analysis can be easily applied to understand deep neural networks with block-diagonal approximation to the Fisher information matrix, where each block corresponds to the parameters for a single neuron. A generalized linear model (GLM) can be regarded as parameterizing an output distribution from the exponential family using a weight vector wand bias scalar b. To be consistent with the previous

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

15

A generalized linear model (GLM) can be regarded as parameterizing an output distribution from the exponential family using a weight vector wand bias scalar b. To be consistent with the previous sections, the log likelihood of the GLM can be written using the summed inputs aas the following: logP(yjx;w;b) =(a+b)y(a+b) +c(y;); (10) E[yjx] =f(a+b) =f(w>x+b);Var[yjx] =f0(a+b); (11) where,f(
)is the transfer function that is the analog of the non-linearity in neural networks, f0(
) is the derivative of the transfer function, (
)is a real valued function and c(
)is the log partition function. is a constant that scales the output variance. Assume a H-dimensional output vector y= [y1;y2;


;yH]is modeled using Hindependent GLMs and logP(yjx;W;b) =PH i=1logP(yijx;wi;bi). LetWbe the weight matrix whose rows are the weight vectors of the

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

16

i=1logP(yijx;wi;bi). LetWbe the weight matrix whose rows are the weight vectors of the individual GLMs, bdenote the bias vector of length Handvec(
)denote the Kronecker vector operator. The Fisher information matrix for the multi-dimensional GLM with respect to its parameters = [w> 1;b1;


;w> H;bH]>= vec([W;b]>)is simply the expected Kronecker product of the data features and the output covariance matrix: F() = E xP(x)Cov[yjx] 2  xx>x x>1 : (12) We obtain normalized GLMs by applying the normalization methods to the summed inputs ain the original model through and. Without loss of generality, we denote Fas the Fisher information matrix under the normalized multi-dimensional GLM with the additional gain parameters = vec([W;b;g]>): F() =2 664F11


F1H ......... FH1


FHH3 775;Fij=E xP(x)2 664Cov[yi; yjjx]

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

17

FH1


FHH3 775;Fij=E xP(x)2 664Cov[yi; yjjx] 22 664gigj iji> jigi iigi(ajj) ij > jgj j1ajj j > jgj(aii) ijaii i(aii)(ajj) ij3 7753 775 (13) i=x@i @wiaii i@i @wi: (14) Implicit learning rate reduction through the growth of the weight vector: Notice that, comparing to standard GLM, the block Fijalong the weight vector widirection is scaled by the gain parameters and the normalization scalar i. If the norm of the weight vector wigrows twice as large, even though the model’s output remains the same, the Fisher information matrix will be different. The curvature along the widirection will change by a factor of1 2because the iwill also be twice

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

18

The curvature along the widirection will change by a factor of1 2because the iwill also be twice as large. As a result, for the same parameter update in the normalized model, the norm of the weight vector effectively controls the learning rate for the weight vector. During learning, it is harder to change the orientation of the weight vector with large norm. The normalization methods, therefore, 5 050100150200250300 iteration x 30034353637383940414243mean Recall@1Image Retrieval (Validation) Order-Embedding + LN Order-Embedding(a) Recall@1 050100150200250300 iteration x 3007172737475767778mean Recall@5Image Retrieval (Validation) Order-Embedding + LN Order-Embedding (b) Recall@5 050100150200250300 iteration x 30084858687888990mean Recall@10Image Retrieval (Validation) Order-Embedding + LN Order-Embedding (c) Recall@10 Figure 1: Recall@K curves using order-embeddings with and without layer normalization. MSCOCO Caption Retrieval Image Retrieval

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

19

Figure 1: Recall@K curves using order-embeddings with and without layer normalization. MSCOCO Caption Retrieval Image Retrieval Model R@1 R@5 R@10 Mean rR@1 R@5 R@10 Mean r Sym [Vendrov et al., 2016] 45.4 88.7 5.8 36.3 85.8 9.0 OE [Vendrov et al., 2016] 46.7 88.9 5.7 37.9 85.9 8.1 OE (ours) 46.6 79.3 89.1 5.2 37.8 73.6 85.7 7.9 OE + LN 48.5 80.6 89.8 5.1 38.9 74.3 86.3 7.6 Table 2: Average results across 5 test splits for caption and image retrieval. R@K is Recall@K (high is good). Mean ris the mean rank (low is good). Sym corresponds to the symmetric baseline while OE indicates order-embeddings. have an implicit “early stopping” effect on the weight vectors and help to stabilize learning towards convergence. Learning the magnitude of incoming weights: In normalized models, the magnitude of the incoming weights is explicitly parameterized by the gain parameters. We compare how the model output

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

20

convergence. Learning the magnitude of incoming weights: In normalized models, the magnitude of the incoming weights is explicitly parameterized by the gain parameters. We compare how the model output changes between updating the gain parameters in the normalized GLM and updating the magnitude of the equivalent weights under original parameterization during learning. The direction along the gain parameters in Fcaptures the geometry for the magnitude of the incoming weights. We show that Riemannian metric along the magnitude of the incoming weights for the standard GLM is scaled by the norm of its input, whereas learning the gain parameters for the batch normalized and layer normalized models depends only on the magnitude of the prediction error. Learning the magnitude of incoming weights in the normalized model is therefore, more robust to the scaling of the input and its parameters than in the standard model. See Appendix for detailed derivations. 6 Experimental results We perform experiments with layer normalization on 6 tasks, with a focus on recurrent neural networks: image-sentence ranking, question-answering, contextual language modelling, generative modelling, handwriting sequence generation and MNIST classification. Unless otherwise noted, the default initialization of layer normalization is to set the adaptive gains to 1and the biases to 0in the experiments. 6.1 Order embeddings of images and language

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

21

the default initialization of layer normalization is to set the adaptive gains to 1and the biases to 0in the experiments. 6.1 Order embeddings of images and language In this experiment, we apply layer normalization to the recently proposed order-embeddings model of Vendrov et al. [2016] for learning a joint embedding space of images and sentences. We follow the same experimental protocol as Vendrov et al. [2016] and modify their publicly available code to incorporate layer normalization1which utilizes Theano [Team et al., 2016]. Images and sentences from the Microsoft COCO dataset [Lin et al., 2014] are embedded into a common vector space, where a GRU [Cho et al., 2014] is used to encode sentences and the outputs of a pre-trained VGG ConvNet [Simonyan and Zisserman, 2015] (10-crop) are used to encode images. The orderembedding model represents images and sentences as a 2-level partial ordering and replaces the cosine similarity scoring function used in Kiros et al. [2014] with an asymmetric one. 1https://github.com/ivendrov/order-embedding 6 0 100 200 300 400 500 600 700 800 training steps (thousands)0.40.50.60.70.80.91.0validation error rateAttentive reader

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

22

0 100 200 300 400 500 600 700 800 training steps (thousands)0.40.50.60.70.80.91.0validation error rateAttentive reader LSTM BN-LSTM BN-everywhere LN-LSTMFigure 2: Validation curves for the attentive reader model. BN results are taken from [Cooijmans et al., 2016]. We trained two models: the baseline order-embedding model as well as the same model with layer normalization applied to the GRU. After every 300 iterations, we compute Recall@K (R@K) values on a held out validation set and save the model whenever R@K improves. The best performing models are then evaluated on 5 separate test sets, each containing 1000 images and 5000 captions, for which the mean results are reported. Both models use Adam [Kingma and Ba, 2014] with the same initial hyperparameters and both models are trained using the same architectural choices as used in Vendrov et al. [2016]. We refer the reader to the appendix for a description of how layer normalization is applied to GRU. Figure 1illustrates the validation curves of the models, with and without layer normalization. We plot R@1, R@5 and R@10 for the image retrieval task. We observe that layer normalization offers

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

23

plot R@1, R@5 and R@10 for the image retrieval task. We observe that layer normalization offers a per-iteration speedup across all metrics and converges to its best validation model in 60% of the time it takes the baseline model to do so. In Table 2, the test set results are reported from which we observe that layer normalization also results in improved generalization over the original model. The results we report are state-of-the-art for RNN embedding models, with only the structure-preserving model of Wang et al. [2016] reporting better results on this task. However, they evaluate under different conditions (1 test set instead of the mean over 5) and are thus not directly comparable. 6.2 Teaching machines to read and comprehend In order to compare layer normalization to the recently proposed recurrent batch normalization [Cooijmans et al., 2016], we train an unidirectional attentive reader model on the CNN corpus both introduced by Hermann et al. [2015]. This is a question-answering task where a query description about a passage must be answered by filling in a blank. The data is anonymized such that entities

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

24

about a passage must be answered by filling in a blank. The data is anonymized such that entities are given randomized tokens to prevent degenerate solutions, which are consistently permuted during training and evaluation. We follow the same experimental protocol as Cooijmans et al. [2016] and modify their public code to incorporate layer normalization2which uses Theano [Team et al., 2016]. We obtained the pre-processed dataset used by Cooijmans et al. [2016] which differs from the original experiments of Hermann et al. [2015] in that each passage is limited to 4 sentences. In Cooijmans et al. [2016], two variants of recurrent batch normalization are used: one where BN is only applied to the LSTM while the other applies BN everywhere throughout the model. In our experiment, we only apply layer normalization within the LSTM. The results of this experiment are shown in Figure 2. We observe that layer normalization not only trains faster but converges to a better validation result over both the baseline and BN variants. In Cooijmans et al. [2016], it is argued that the scale parameter in BN must be carefully chosen and is set to 0.1 in their experiments. We experimented with layer normalization for both 1.0 and 0.1 scale

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

25

set to 0.1 in their experiments. We experimented with layer normalization for both 1.0 and 0.1 scale initialization and found that the former model performed significantly better. This demonstrates that layer normalization is not sensitive to the initial scale in the same way that recurrent BN is.3 6.3 Skip-thought vectors Skip-thoughts [Kiros et al., 2015] is a generalization of the skip-gram model [Mikolov et al., 2013] for learning unsupervised distributed sentence representations. Given contiguous text, a sentence is 2https://github.com/cooijmanstim/Attentive_reader/tree/bn 3We only produce results on the validation set, as in the case of Cooijmans et al. [2016] 7 5 10 15 20 iteration x 5000082.082.583.083.584.084.585.085.586.0Pearson x 100 Skip-Thoughts + LN Skip-Thoughts Original(a) SICK(r) 5 10 15 20 iteration x 500002728293031323334MSE x 100 Skip-Thoughts + LN Skip-Thoughts Original (b) SICK(MSE) 5 10 15 20

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

26

Skip-Thoughts + LN Skip-Thoughts Original (b) SICK(MSE) 5 10 15 20 iteration x 5000070727476788082Accuracy Skip-Thoughts + LN Skip-Thoughts Original (c) MR 5 10 15 20 iteration x 5000074767880828486Accuracy Skip-Thoughts + LN Skip-Thoughts Original (d) CR 5 10 15 20 iteration x 5000090.090.591.091.592.092.593.093.594.094.5Accuracy Skip-Thoughts + LN Skip-Thoughts Original (e) SUBJ 5 10 15 20 iteration x 50000838485868788899091Accuracy Skip-Thoughts + LN Skip-Thoughts Original (f) MPQA Figure 3: Performance of skip-thought vectors with and without layer normalization on downstream tasks as a function of training iterations. The original lines are the reported results in [Kiros et al., 2015]. Plots with error use 10-fold cross validation. Best seen in color. Method SICK( r) SICK() SICK( MSE ) MR CR SUBJ MPQA

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721

1607.06450

27

Method SICK( r) SICK() SICK( MSE ) MR CR SUBJ MPQA Original [Kiros et al., 2015] 0.848 0.778 0.287 75.5 79.3 92.1 86.9 Ours 0.842 0.767 0.298 77.3 81.8 92.6 87.9 Ours + LN 0.854 0.785 0.277 79.5 82.6 93.4 89.0 Ours + LN y 0.858 0.788 0.270 79.4 83.1 93.7 89.3 Table 3: Skip-thoughts results. The first two evaluation columns indicate Pearson and Spearman correlation, the third is mean squared error and the remaining indicate classification accuracy. Higher is better for all evaluations except MSE. Our models were trained for 1M iterations with the exception of (y) which was trained for 1 month (approximately 1.7M iterations) encoded with a encoder RNN and decoder RNNs are used to predict the surrounding sentences. Kiros et al. [2015] showed that this model could produce generic sentence representations that

1607.06450

Layer Normalization

Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques.

http://arxiv.org/pdf/1607.06450

stat.ML

20160721

20160721