Datasets:

---

spiralworks

/

openreview-structured-forums

like 0

Follow

Spiral Works 3

[ "Daquan Zhou", "Xiaojie Jin", "Qibin Hou", "Kaixin Wang", "Jianchao Yang", "Jiashi Feng" ]

[ "Network Compression", "Classification", "Deep Learning", "Weights Sharing" ]

{ "note_id": [ "8mLSR1giU", "rJxZSF8hsB", "Bkxrjd83jS", "rygGUdInjr", "Skej2PP_jH", "HJe0vwDOsH", "S1lORAAAcB", "rkenIi7GqB", "B1gMrWSXFS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798733110, 1573837112710, 1573836957312, 1573836874013, 1573578675245, 1573578597976, 1572953807781, 1572121428259, 1571143993958 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1816/Authors" ], [ "ICLR.cc/2020/Conference/Paper1816/Authors" ], [ "ICLR.cc/2020/Conference/Paper1816/Authors" ], [ "ICLR.cc/2020/Conference/Paper1816/Authors" ], [ "ICLR.cc/2020/Conference/Paper1816/Authors" ], [ "ICLR.cc/2020/Conference/Paper1816/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1816/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1816/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The paper proposed a novel way to compress arbitrary networks by learning epitiomes and corresponding transformations of them to reconstruct the original weight tensors. The idea is very interesting and the paper presented good experimental validations of the proposed method on state-of-the-art models and showed good MAdd reduction. The authors also put a lot of efforts addressing the concerns of all the reviewers by improving the presentation of the paper, which although can still be further improved, and adding more explanations and validations on the proposed method. Although there's still concerns on whether the reduction of MAdd really transforms to computation reduction, all the reviewers agreed the paper is interesting and useful and further development of such work would be useful too.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to AnonReviewer1 comments and suggestions (1/3)\", \"comment\": \"Thank you for your detailed review! Your questions are valuable, and we are happy to address your concerns.\\n\\nReply to the writing of the paper\\n\\n\\u2022\\tWe have revised the paper and correct the citation errors and revise the words carefully\\n\\u2022\\tWe have re-organized section 3 to improve the presentation of the explanations of the method. We have re-designed figure 2 to give a detailed illustration of how the weight tensor can be generated using the epitome along the spatial dimension. We have also added Figure 3 to illustrate the transformation process along the input channel dimension. The transformation along the filter dimension is the same as the transformation along the input channel dimension.\", \"reply_to_the_potential_extension_to_dynamic_allocation\": \"\\u2022\\tFor each designed scheme, the network needs to be trained from scratch. We do not try the dynamic allocation experiments so far. But it is interesting and useful if a single epitome can be used to construct different networks without re-training. Our method can be extended intuitively to achieve the target by incorporating the concept of switch introduced in the slimmable neural network [1]. More specifically, we can put in four switches where each switch controls the percentage of epitome that can be used for the transformation and we can learn four different transformation rules for each switch. In that case, our epitome in each layer can be used to generate different weight tensors dynamically. However, we put it as a separate work from the current version since it is an extension to the current method and can always be integrated into the current method.\\n\\nReply to the experiments\\n\\nLatency measurement\\n\\u2022\\tInitially, we use Madd, number of parameters and the classification accuracy as metrics for comparisons with other methods. We do not measure the energy due to the hardware limitation. We have added in one more experiments to measure the latency of our method and the results are shown as below:\\n\\nMethod\\t\\t Latency\\tDevice\\t\\tBatch Size\\t Madd(M)\\n----------------------------------------------------------------------------------------------------\\nMBV2-0.5\\t\\t 26 ms\\t CPU\\t\\t 1\\t\\t 100\\nMBV2-0.75\\t\\t 28 ms\\t CPU\\t\\t 1\\t\\t 220\\nMBV2-1.0\\t\\t 34 ms\\t CPU\\t\\t 1\\t\\t 310\\nMBV2-1.4\\t\\t 42 ms\\t CPU\\t\\t 1\\t\\t 590\\nMBV2-1.4\\t\\t 7.79 s\\t CPU\\t\\t 1000\\t\\t 590\\n----------------------------------------------------------------------------------------------------\\nMBV2-1.0 + NES-0.5\\t 27 ms\\t CPU\\t 1 157\\nMBV2-1.4 + NES-0.5\\t 33 ms\\t CPU\\t 1 340\\nMBV2-1.4 + NES-0.5\\t 6.27 s\\t CPU 1000 \\t 340\\n\\nThe server that we have used for the evaluation has 80 Intel(R) Xeon(R) E5-2698 v4 @ 2.20GHz CPUs. As CPU adjusts its power automatically, we put two dry run to change the CPU to maximum power mode. We set the batch size to 1 and run 1000 iterations. We first use MobileNetV2 as a baseline to check the latency measurement. We found that the effect of the speedup is larger when the base model\\u2019s computation is large. Hence, we increase the computation by using a MBV2-1.4 base model with 1000 batchsize. To have more accurate measurement, engineering works on the hardware setup is needed. We will try to add in more latency measurements using mobile devices later.\\n\\nError bar on CIFAR-10 dataset\\n\\n\\u2022\\tWe re-train our methods on CIFAR-10 dataset for 8 times and calculate the statistics as shown in the table below:\\n\\nMethod\\t\\t Params(M)\\t Madd(M)\\t\\t Accuracy(%)\\t \\n-----------------------------------------------------------------------------------------------------------\\nMBV2-1.0\\t\\t 2.2 \\t CPU\\t 94.06\\t\\t \\nMBV2-0.18\\t\\t 0.4 \\t CPU\\t\\t 91.7\\t\\n-----------------------------------------------------------------------------------------------------------\\nMBV2-1.0 + NES-0.2\\t 0.39 \\t CPU\\t 93.19\\u00b1 0.013\\t\\t \\n\\n\\n[1] Yu, Jiahui, et al. \\\"Slimmable neural networks.\\\" arXiv preprint arXiv:1812.08928 (2018).\"}", "{\"title\": \"Response to AnonReviewer3 comments and suggestions (2/3)\", \"comment\": \"Madd calculation\\n\\n\\u2022\\tThere are fours parts that contributes to the computation cost: (1) the convolution between the epitome and the input feature map to construct the product map; (2) the construction of the integral map; (3) channel wrapping along the input channel dimension and (4) transformation along the filter dimension.\\n\\u2022\\tThe computation cost for the product map comes from the convolution between the input feature map and the epitome and hence it is calculated as: 2 * (C_{in} * W^{E} * H^{E} \\u2212 1 )W * H * C^{E}_{out}\\n\\u2022\\tThe computation cost of the integral map is the addition used to construct the map. The number of additions is proportional to the number of elements in the map and hence it is calculated as: W * H * W^{E} * H^{E} * C^{E}_{out}\\n\\u2022\\tThe computation cost for the transformation along the input channel dimension comes from the channel wrapping. It is equal to the number of additions between the feature map and multiplications with the coefficient as introduced in Eqn.(11) in the main text. Thus, the total computation is: 2Rcin * W * H * \\u03b2_{1}\\n\\u2022\\tThe computation cost of the transformation along the filter dimension is similar to the one along the input channel dimensions. The only difference is that channel wrapping is not needed and hence the total computation cost is computed as: 2R_{cout} * \\u03b2_{2}\\n\\u2022\\tThe total computation cost is thus calculated as: 2 * C_{in}W^{E} 8 H^{E} \\u2212 1 )W * H * C^{E}_{out} + W * H * W^{E} * H^{E} * C^{E}_{out} + 2Rcin * W * H * \\u03b2_{1} + 2R_{cout} * \\u03b2_{2}\\n\\u2022\\tThe size of the integral map can be quite small in the implementation by setting the spatial dimension of the epitome to be small.\\n\\u2022\\tWe have added in more details in section 3.4 regarding how the computation can be saved. Madd is calculated theoretically. The actually speed up is not fully proportionally to the reduction in madd. However, we believe those engineering works can be done with optimized coding and driver layer.\\n\\u2022\\tThe calculation of Madd is based on equation 12 and the detailed derivation can be found in Appendix D in the revised version of the paper. \\n\\u2022\\tThe pseudo code for Madd calculation is shown at the end of the reply\\n\\nDifference between NES and Associative Convolutional Layers(ACL):\\n\\nBoth NES and ACL targets to express the convolution layers with a compact set. ACL builds on the basis of tensor low-rank factorization. It first divides a tensor into a set of basises (or the so called slice in the paper) and then uses a convolution slice generator to form the weight tensor used for convolution. The generator and code word used in ACL can be regarded as a fully connected layer. The core idea is similar to decompose a convolution layer into a set of smaller convolution tensors followed by a fully connected layer. Differently, our method uses an index learner to learn the starting index of a sub-tensor and use a bi-linear interpolation based method to generate the weight tensor with much larger dimensions. The advantages of our method over ACL are summarized as below: \\n\\n\\u2022\\tOur method has much less storage and computation overhead than ACL method. The slice generator and the code word used in ACL can be regarded as a fully connected layer to combine all slices. The generator itself consumes the storage memory. Besides, using a fully connected layer to combine all slice basis might consume a large computation memory and the computation cost is not reported in the paper. Differently, our method removes the index learner by implementing a routing map. Both the memory and computation cost can be saved using our method.\\n\\u2022\\tIn ACL, the partition of the weight tensor into slice is a manual work. When applied on different backbone networks, the partition scheme might need to be re-designed. However, in our method, we use an index learner to learn the starting index and then generate the weight tensor based on a non-parametric interpolation function. No manual work is needed for the generation process.\\n\\u2022\\tThe experiments in ACL focus on Densenet and ResNet. ResNet is popular but redundancy is higher than the current SOTA mobile models. Densenet is efficient on memory but the computation cost is expensive and the computation cost is not reported in ACL. Differently, our experiment focuses on current SOTA mobile networks such as MobileNetv2 and EfficientNet and we report both the memory size and computation results.\\n\\u2022\\tThe key idea of this paper is to treat the network architecture and parameters separately and search a transformation function that can match the gap between the architecture and the parameter. This is expected to give a new search space for neural architecture design.\"}", "{\"title\": \"Response to AnonReviewer3 comments and suggestions (3/3)\", \"comment\": \"The pseudo code for the madd calculation is shown as below:\", \"pseudo_code_for_madd_calculation\": \"def conv_flops_counter_hook(conv_module, input, output):\\n# calculation based on equation 21 in the paper\\n batch_size = input.shape[0]\\n output_dims = list(output.shape[2:])\\n\\n kernel_dims = list(conv_module.kernel_size)\\n in_channels = conv_module.in_channels\\n out_channels = conv_module.out_channels\\n epi_in = self.weight.shape[1]\\n epi_out = self.weight.shape[0]\\n epi_w, epi_h = self.weight.shape[2], self.weight.shape[3]\\n\\n # we choose \\\\beta_1 = epi_in and \\\\beta_2 = epi_out empirically\\n rep_dim1 = out_channels // epi_in\\n rep_dim0 = in_channels // epi_out\\n\\n filters_per_channel = out_channels // groups\\n product_map_cost = (2* in_channels * epi_w * epi_h -1) * np.prod(output_dims) * epi_out\\n integral_map_cost = np.prod(output_dims) * epi_w * epi_h * epi_out\\n channel_trans_cost = 2 * rep_dim1 * np.prod(output_dims)\\n filter_trans_cost = 2 * rep_dim0\\n\\n conv_per_position_flops = product_map_cost + integral_map_cost +\\n channel_trans_cost + filter_trans_cost\"}", "{\"title\": \"Response to AnonReviewer2 comments and suggestions\", \"comment\": \"Thank you for your detailed and valuable review! Following your suggestions, we have updated the paper with improved presentation and we're happy to address your concerns. A summary of the modifications is shown as below:\", \"reply_to_major_comments\": \"\", \"computation_savings_with_routing_map\": \"\\u2022\\tThe benefits brought by the routing map is that the index learner can be discarded during the inference phase, and the outputs of the index learner is recorded in the routing map and saved as a look-up table. With the implementation of the routing map, the computation overhead due to the index learner is removed. We have revised section 3.2 to state this benefit clearer in a separate paragraph named as \\u201cRouting map\\u201d\\n\\nRedesign of the weight tensor generation process(figure 2 in original version):\\n\\n\\u2022\\tWe agree that previous figure 2 is not clear on the generation process. We have modified figure 2 in section 3.2 to improve the presentation. We combined figure 4 and figure 2 in the original version to give an example of the transformation process along the spatial dimension with the structure of the index learner. We also added figure 3 to illustrate the transformation along the channel dimensions. Based on figure 3, we put in more details regarding the meaning of R_{cin}. The meaning of R_{cin} is the number of samplings applied along the input channel dimension. The transformation along the filter dimension is same as the transformation along the input channel dimension. The difference is that the computation reuse along the input channel dimension needs special techniques of channel wrapping. We illustrate the channel wrapping process in Figure 4 of the updated version.\\n\\nYes, each layer has an epitome and the epitome is not shared among layers. We have revised the paper to emphasize this point in section 3.2 paragraph \\u201cIndexing function\\u201d.\", \"design_of_the_transformation_function\": \"\\u2022\\tWe use separate transformation learners for each layer. The learner is implemented by two convolution layers. It takes in the input feature tensor of the corresponding layer and outputs the mapped starting index of the sub-tensor in the epitome. The sub-tensor at the new position will be used to construct the weight tensor for convolution. We also show the structure of the learner in the revised figure 2. We rename this transformation learner as index learner in the revised version in section 3.2.\", \"memory_and_computation_compression_calculation_of_the_routing_map\": \"\\u2022\\tThe memory cost of the routing map is 3 x R_{cin} + R_{cout}. The routing map is used to record the index mapping between the sub-tensor in the epitome and the weight tensor. This is implemented as a look-up table manner. Because we take a patch of the tensor in the epitome, only the starting indices are recorded. The position of the rest of the elements can be calculated as starting index + offset where offset is the relative position for those elements in the sub-tensor. \\n\\nExplanation of R_{cin} and R_{cout}:\\n\\n\\u2022\\tFor example, with a learned routing map M, the sub-tensor in the epitome that will be used to fill up the sub-tensor at position (i,j,m) in the weight tensor is E[p:p+w, q:q+h, c\\u2019_{in}:c\\u2019_{in} + \\\\beta_1, :], where (p,q, c\\u2019_{in}) = M(i,j,m). The selected sub-tensors in the epitome are concatenated together to form a larger tensor. Hence, the number of such indices is calculated as C\\u2019_{in} / \\\\beta_1 and this is noted as R_{cin}. Since each recorded index has three values, the size of the routing map is 3 x R_{cin}. Similarly, along the filter dimension, the size is R_{cout} where R_{cout} = C '_{out}/ \\\\beta_2. Here, \\\\beta_1 and \\\\beta_2 denotes the length of the sub-tensor along the input channel dimension and the filter dimension that we select from the epitome. We have put in more details regarding the meaning of R_{cin} and R_{cout} in section 3.2. Since the routing map is implemented as a look-up table, the computation is O(1). The formulation for the computation of the compression ratio has taken the routing map into consideration.\", \"reply_to_minor_comments\": \"\\u2022\\tMeaning of \\u201chardwired\\u201d: this term is supposed to describe the relationship between the neural network architecture and the corresponding parameters. In this work, we consider the network architecture and the parameters as separate components. The word \\u201chardwired\\u201d is used to describe the one-to-one correspondence between the architecture and the parameters. We have updated the abstract to make this clearer.\\n\\u2022\\tWe have modified the experiments summary part in the abstract as pointed out.\"}", "{\"title\": \"Response to AnonReviewer3's comments and suggestions\", \"comment\": \"Thank you for your review! Your questions are valuable, and we are happy to address your concerns. In summary, we explain below how the epitome parameters are updated and why the method is agnostic to the model architecture.\", \"q1\": \"It is not too clear how to learn the epitome and transformation functions.\\n\\n\\u2022\\tEpitome parameter updates: the epitome is defined as a parameter tensor with smaller dimensions. The parameters inside the epitome are updated with standard back-propagation to minimize the training loss, after being transformed to instantiate model parameters. The gradient of each parameter in the epitome is the summation of the gradients of all the elements in the convolution weight tensor where the corresponding epitome element is used. The transformation function is end-to-end differentiable and hence the epitome parameters are updated in an end-to-end manner. The learning procedure for epitome is introduced in detail in Section 3.3. We have revised the section to make it clearer.\\n\\u2022\\tTransformation function: We updated the description in Section 3.2 on transformation functions in the paper. In summary, the transformation function includes three parts: (1) an index learner, (2) a routing map and (3) an interpolation based sampler. The transformation function works as follows: It uses the index learner to learn a group of indices that map the sub-tensor inside the epitome to the convolution weight tensor. The learned indices and the epitome are then fed into an interpolation based sampler. Each pair of the learned starting index will have a corresponding output from the interpolation based sampler. Those outputs are then concatenated together to constitute the weight tensor. The index learner is implemented by a two-layer module. We use separate index learners for each layer and the learner takes the input feature of the corresponding layer as its input and outputs the learned starting indices. Thus, the learner can be optimized via standard back propagation.\", \"q2\": \"Independence between the epitome and the architecture\\n\\n\\u2022\\tIn section 3, we show that the model size and the computation are mainly related to the size of the epitome. With our proposed transformation function, the shape of the epitome is independent of the model architecture. This is because the transformation function is able to transform any shape of the epitome to a tensor matching the shape of weight tensor for the architecture.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper focuses on the problem of neural network compression, and proposes a new scheme, the neural epitome search. It learns to find compact yet expressive epitomes for weight parameters of a specified network architecture. The learned weight tensors are independent of the architecture design. It can be encapsulated as a drop in replacement to the current\\nconvolutional operator. It can incur less performance drop. Experiments are conducted to show the effectiveness of the proposed method. However, there are some concerns to be addressed.\\n-It is not too clear how to learn the epitomes and transformation functions.\\n-Authors stated that the proposed method is independent of the architecture design. From the current statements, it is not explained clearly.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"In this work, the authors describe a technique for compressing neural networks by learning a so-called Epitome (E), and a transformation function (\\\\theta) such that the weights for each layer can be constructed using \\\\theta(E). The epitome and the transformation function can be learnt jointly while optimizing the network for the task specific loss.\\n\\nThe main idea of this paper is really interesting, and the experimental results which compare against other recent techniques also validate the proposed technique. However, while I think the main idea is relatively clear, I personally found the description of the proposed techniques -- particularly 3.2 and 3.4 -- to be somewhat hard to follow, particularly given that some details only appear in the Appendix. I would suggest that the authors try to revise this section by trying to move Figures 4 and 5 from the appendix into the main text. Some additional suggestions also appear below.\\n\\nOverall, I would while I like the ideas in this paper, based on the current presentation I am inclined to rate the paper as a \\u201cweak reject\\u201d, though I would raise my rating if the paper was revised to improve the presentation.\", \"main_comments\": \"1. The authors mention that \\u201cDuring inference, (the) routing map enables the model to reuse computations when the expanded weight tensors are formed based on the same set of elements in the epitomes and therefore effectively reduces the computation cost.\\u201d It would be nice to include some results which indicate what the savings are with the proposed routing map.\\n\\n2. Section 3.2: There were a few aspects of section 3.2 that I think could be improved for clarity.\\nA.) Personally, I found Figure 2 somewhat tricky to follow. I would suggest removing the 3x2 \\u201cEpitome\\u201d and \\u201cGenerated Kernel\\u201d figures on the extreme right of the image, since I\\u2019m assuming the only goal of these is to indicate that \\u201corange\\u201d and \\u201cblue\\u201d correspond to \\u201cEpitome\\u201d and \\u201cKernel\\u201d respectively. I would also suggest mentioning the correspondence of the colors as the first sentence in the caption. Finally, if possible, I would suggest adding a small description alongside the (a), (b), (c) subcaptions: e.g., (a) straightforward but non-differentiable mapping, \\u2026 , (c) Generated Weigh Kernel.\\n\\nB.) I believe that the authors use a separate E for each layer, and that Epitomes are not shared across layers. I may have missed this in the text, but it would be useful to clarify this explicitly again in the section.\\n\\nC.) The \\u201cparameterized transformation layer\\u201d is mentioned before Equation 3. I think it would be useful to mention that this is implemented using neural networks in your work for clarity. E.g.: \\u201cTo handle the above two obstacles, ... three\", \"parts\": \"(1) a parameterized transformation learner \\u03b7 (See Section 3.3) used to learn a set of starting indices for patches in epitome (i.e. all elements in the same epitome patch share identical starting indices); \\u2026 and an interpolation based generator (Eqn. 3).\\u201d\\n\\nD.) The exact structure of the \\u201cparameterized transformation layer\\u201d wasn\\u2019t exactly clear to me. In 3.3, the authors mention that it consists of \\u201cof two convolutional layer, followed by a sigmoid function \\u2026 takes the feature map of the convolutional layer as input\\u201d. Please clarify exactly what is fed in as the input to this network e.g., (input feature map: F, and the indices i,j).\\n\\n3. I personally also found Section 3.4 which discusses the Computation reduction was also somewhat hard to follow. Some clarification questions: Is the memory/computation cost of the storing/creating the routing map included in the compression calculations? I think it is important to factor these costs when computing the savings achieved by the model. Also, I was unclear on what R_{cin} and R_{cout} are, and why they appear in Equation 5. Could the authors please clarify.\", \"minor_comments\": \"1. Abstract: \\u201cTraditional compression methods \\u2026 all assume that network architectures\\nand parameters should be hardwired.\\u201d What does it mean for them to be \\u201chardwired\\u201d in this context?\\n2. Abstract: \\u201cExperiments demonstrate that, \\u2026 with 25% MAdd reduction and AutoML for Model Compression (AMC) by 2.5% with nearly the same compression ratio.\\u201d --> \\u201cExperiments demonstrate that, \\u2026 with 25% MAdd reduction, and a 2.5% Madd reduction for AutoML for Model Compression (AMC) with nearly the same compression ratio.\\u201d\\n\\n------- Update after Author Response --------\\nI would like to thank the authors for their responses and for the updates which strengthen the paper in my view. I have updated my score accordingly.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"In this paper, the authors learn epitomes, which are small weight tensors which can be used with a learnt transform to produce tensors of an appropriate size (e.g. the sizes used in MobileNet v2). This gives a reduction in the number of parameters required, and the number of MAdds in theory.\\n\\nThis paper is badly written, and could do with a rewrite:\\n\\n- Citations are used incorrectly (\\\\cite should be used when the citation is meant to be read as part of the sentence). \\n- \\\"less elements\\\" --> \\\"fewer elements\\\"\\n- \\\"misuse the notion\\\" --> \\\"abuse the notation\\\"? \\n- \\\"for fair comparisons\\\" --> \\\"for a fair comparison\\\"\\n\\nThe method is poorly explained; I have read Section 3.2 several times, and I'm still not entirely certain of what's going on. Figure 1 is helpful, but Figure 2 is not, and could be redesigned. 2(b) makes it look like you are going from a 3x3 epitome to a 2x2 kernel, which is clearly not what is happening. I think it would be helpful to give a detailed pictoral example of an epitome mapping to a weight tensor, with arrows between relevant indices changing. \\n\\nOn initial reading, I thought the method allowed dynamic allocation of your epitomes to lots of different tensor sizes. From what I can tell, the network has to be trained from scratch for each possible size, so it isn't flexible in that respect. From what I can gather, the paper is presenting an alternate approach to *downscale* networks, as opposed to say, reducing width or depth. The comparisons to different widths of MobileNet v2 make more sense under this scenario.\\n\\nThe main sell of the methods appears to be on the basis of MAdd reduction. This makes me nervous, as it doesn't necessarily correspond to actual speed-up or a reduction in energy (see https://arxiv.org/abs/1801.04326). EfficientNet was mainly about Madds too, but they provided some inference times. Would it be possible to add these? The method used with the integral image sounds expensive.\\n\\nThe results look good, but error bars,on the CIFAR experiments at the very least, would be appreciated.\", \"pros\": [\"-------\", \"Good results\", \"Method appears largely novel (although bears some resemblance to https://arxiv.org/abs/1906.04309)\"], \"cons\": [\"--------\", \"Badly written\", \"The method is poorly explained\", \"Uncertainty re: MAdds as a primary comparator\"], \"i_propose_a_weak_reject_for_this_paper_for_two_primary_reasons\": \"1) The standard of writing, and the explanation of the all-important method are not up to scratch for a top tier conference\\n2) I have concerns regarding the MAdd calculations. Perhaps you could provide some pseudo-code in the author response?\\n\\nI am happy to upgrade my score if the authors deal with these issues sufficiently.\"}" ] }

[ "Kuen-Han Tsai", "Hsuan-Tien Lin" ]

[ "learning from label proportions", "consistency regularization", "semi-supervised learning" ]

{ "note_id": [ "zYGganLx2C", "B1xT2IkssB", "r1ef2Q1iiB", "HkxPdCi9oB", "H1ez6JiqsH", "rJx0ZWZNqr", "HJlG4z0RtS", "rJgxABc9tr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798733079, 1573742261368, 1573741482009, 1573727854626, 1573724090376, 1572241670455, 1571901994296, 1571624391666 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1815/Authors" ], [ "ICLR.cc/2020/Conference/Paper1815/Authors" ], [ "ICLR.cc/2020/Conference/Paper1815/Authors" ], [ "ICLR.cc/2020/Conference/Paper1815/Authors" ], [ "ICLR.cc/2020/Conference/Paper1815/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1815/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1815/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"After reading the author's rebuttal, the reviewer still hold that the main contribution is just the simple combination of already known losses. And the paper need to pay more attention on the clarity of the paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Paper Updates\", \"comment\": [\"We would like to thank all reviewers for their valuable suggestions and detailed comments. We have updated the paper with the following changes:\", \"Addressing the typo pointed out by Reviewer #1 in Section 2.\", \"Adding the pseudo algorithm of our proposed method in Section 3.\", \"Replacing all expected losses with regular loss terms in Section 3.\", \"Adding the baseline performance with bag size of one in Table 1.\", \"Adding Figure 3 for the distribution of bag sizes under the K-means bag generation procedure in Section 4.\", \"Providing the evolution of the model performance on all benchmarks in Appendix B.\"]}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"We thank the reviewers for their time and effort in reviewing our paper. We are happy that you find the ideas interesting and grateful for many useful comments.\\n\\n1. Thanks for your valuable suggestions. We present the experimental results for one trail on three benchmarks under the computation constraints. However, we repeated several experiments on CIFAR10 and observed that the performance of our method is stable. Currently, we had run more trails on all benchmarks and we would provide more experimental results in the future.\\n\\n2. Thanks for your comments. We added the performance of the vanilla approach with the bag size of 1 in Table 1. The test accuracy with a bag size of 1 corresponds to the performance of the fully-supervised setting. As shown in Table 1, we can find that the test accuracy gradually decreases as the bag size increases. This finding is coherent to previous LLP works (Yu et al., 2013; Dulac-Arnold et al., 2019). \\n\\n3.4. We provide more information and details for the K-means bag generation in the latest revised version. As shown in Figure 3, we can see that the distribution of bag sizes becomes diverse as the number of clusters K decreases. We find that the model performance with a smaller value of K is superior to or than the model performance with larger bag sizes in some cases. For example, as shown in Figure 3, the median bag size is around 100 when the value of K is 390 on CIFAR100. The condition is similar to that of uniform bag generation with a bag size of 128. However, the test accuracy of the vanilla approach with a bag size of 128 is 5.82%, whereas the model performance with K of 390 gets 15.47%. This result indicates that the performance can be more as we had already captured the underlying structure of data by the K-means bag generation. Nevertheless, we can also find that the performance would decrease and is not well-correlated with the value of K. One possible reason is that we might drop informative bags as we randomly split bags into validation and training. It is challenging and worth-studying to investigate the cross-validation technique for the K-means bag generation. Also, we acknowledge that it is also an interesting direction to study the privacy-preserving issues for LLP with the K-means procedure.\\n\\n5. Thanks for your suggestion. Although we didn't apply the Mixup techniques in our work, we agree that it is an interesting future direction to combining this technique into LLP. It is really interesting to think of generating perturbations by mixing the data examples within a bag. Also, we can generate consistency targets by mixing the proportion labels and optimize the proportion loss with mixed bags. We hope that future LLP works can follow our work and further explore new consistency-based methods for multi-class LLP.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"Thank you for your valuable suggestions and detailed insights. We would like to address your concerns as follows:\\n\\n1. We agree about the simplicity of our proposed method, but we view it as an unexplored direction for the LLP problem. In comparison to ROT (Dulac-Arnold et al., 2019), our work focuses on finding a simpler and computationally cheaper regularization technique that can be easily extended to the LLP scenario. Since the ROT approach needs to optimize the model parameters and estimate the pseudo-labels within a bag alternatively, the procedure of ROT is time-consuming. Although we borrow the concept of consistency regularization from semi-supervised learning, our proposed method for LLP is more efficient than ROT. Also, the empirical results demonstrate not only superior performance but also the stability of our proposed method. For example, the ROT method receives 62.93% test accuracy on CIFAR10 with a bag size of 64, whereas the vanilla approach receives 70.68% test accuracy and our approach obtains a better test accuracy of 72.49%. If we look at Table 1, we can observe that the performance of ROT on CIFAR10 and CIFAR100 with medium bag sizes is unstable and lead us to conclude that the unhelpful pseudo-labels would easily result in a worse classifier. Conversely, our LLP-VAT is more stable and obtains better test accuracy in most cases.\\n\\n2. The reason why we conduct the experiments on SVHN, CIFAR10, and CIFAR100 is that we follow the standard experiment setting in the previous LLP work (Dulac-Arnold et al., 2019). In terms of large scale and multi-class classification, we think those image datasets are suitable and competitive for multi-class LLP. Besides, to our best knowledge, there are no existing LLP works make the same smoothness assumption for the multi-class LLP problem. Our work first explores a new direction by incorporating consistency regularization into LLP and demonstrate it does help learn a better classifier.\\n\\n3. Thank you for your detailed comments and insights. We present the experimental results for one trail on three benchmarks under the computation constraints. However, we repeated several experiments on CIFAR10 and observed that the performance of our method is stable. Currently, we had run more trails on all benchmarks and we would provide more experimental results in the future.\\n\\n4. We checked the learning curve of our method, but we did not see the situation where the accuracy goes up quickly at first and decreases gradually. We plot the test curve over training epochs and the results are shown in Figures 4 and 5. As shown in Figures 4 and 5, the performance of our approach rises steadily and converges in the end. We provide the evolution of test accuracy on all benchmarks with two bag generation procedures in the supplementary material.\\n\\n5. Thanks for pointing that, we removed the expected loss J_cons in the paper. Please check the latest revised version. To avoid confusion, we replace all expected loss with regular loss term computed by each bag. Also, we provide a pseudo algorithm of our approach in the revised paper.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"Thanks for your valuable suggestions and detailed comments.\\n\\n1. We agree about the simplicity of our proposed method, but we view it as an unexplored direction for the LLP problem. In comparison to ROT (Dulac-Arnold et al., 2019), our work focuses on finding a simpler and computationally cheaper regularization technique that can be easily extended to the LLP scenario. Since the ROT approach needs to optimize the model parameters and estimate the pseudo-labels within a bag alternatively, the procedure of ROT is time-consuming. Although we borrow the concept of consistency regularization from semi-supervised learning, our proposed method for LLP is more efficient than ROT. Also, the empirical results demonstrate not only superior performance but also the stability of our proposed method. For example, the ROT method receives 62.93% test accuracy on CIFAR10 with a bag size of 64, whereas the vanilla approach receives 70.68% test accuracy and our approach obtains a better test accuracy of 72.49%. If we look at Table 1, we can observe that the performance of ROT on CIFAR10 and CIFAR100 with medium bag sizes is unstable and lead us to conclude that the unhelpful pseudo-labels would easily result in a worse classifier. Conversely, our LLP-VAT is more stable and obtains better test accuracy in most cases.\\n\\n2. Thanks for pointing that, we corrected the typo mistake of \\\"K\\\" in the paper. We provide the expected loss over bags because we would sample a few bags as a mini-batch and then compute the corresponding mini-batch loss in each epoch. That is, we would not compute the total loss by summation of all bags in the experiments. To avoid any confusion, we replace all expected losses in section 3 with the regular loss terms and provide a pseudo algorithm of our approach in the last revised version.\\n\\n3. The K-means bag generation would cluster data examples into bags by the K-means algorithm. This procedure of bag generation doesn't require image datasets, i.e. we can apply it to any dataset. Therefore, the use of image attributes is not the requirement for the K-means bag generation.\\n\\n4. Thanks for your valuable suggestion. We present the experimental results for one trail on three benchmarks under the computation constraints. However, we repeated several experiments on CIFAR10 with different random seeds and observed that the performance of our method is stable. Currently, we had run more trails on all benchmarks and we would provide more experimental results in the future.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"[Summary]\\nThe paper presents a solution to the problem of learning from label proportion (LLP) by incorporating regularization by data generation motivated by semi-supervised learning. \\nIt is argued that, by consistency assumption, the classification function defined on the data manifold should be a locally consistent mapping (in the sense of bag label proportion) such that the discrepancy between two bags of data points within a small neighborhood should be constraint. This is further combined with the conventional LLP loss (cross entropy between bag label proportion and classification results) to produce the final loss function for learning. For bag data generation, image attributes are leveraged to group similar data points in the feature space. Evaluation is performed on three benchmarks against vanilla solution and ROT loss. \\n\\n[Comments]\\nI\\u2019m not sure if I fully follow the contribution and several technical details.\\n\\n It looks to me that the major contribution claimed is the novel loss function that combines the proportion loss and the consistency loss, but both seem to be from off-the-shelf solutions from literature with slight variation. E.g., J_prop is the standard cross entropy loss (Ardehaly & Culotta (2017) and Dulac-Arnold et al. (2019)), and J_cons is from the vanilla consistency definition. I had a hard time getting the novelty here. \\n\\nThe notion and definition are somehow confusing to me too. What is K in the second line of page 3? Shouldn\\u2019t J_prop in page 4 defined as sum of all per-bag losses? In J_cons of pge 4, shouldn\\u2019t x under the summation be x_\\\\hat as x_\\\\hat is sampled? The equations should be properly numbered for easy reference.\\n\\nThe use of image attributes in 4.4 for K-means bag generation seems a strong requirement? What kinds of image attributes are used and how are they generated? Looks like this strategy only applies to image classification?\\n\\nThe results reported could be more clear. The gap between the proposed method and vanilla or ROT does not seem quite big in many cases (less than 1% in the best cases in table 1). I\\u2019m not sure if these results are convincing or not as statistical significance is unclear. \\n\\nWith all of the above uncertainty, I do not have confidence to have the paper accepted in the current format based on my preliminary assessment.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\nThis paper proposes using Consistency Regularization and a new bag generation technique to better learn classification decision boundaries in a Label Proportion setting. The consistency regularization works to make sure that examples in the local neighbourhood have similar outputs. The authors further use K-means clustering to create a new bagging scenario they use to mimic real-world LLP settings. \\n\\nOverall, this paper sheds light on how we can use the underlying structure to better be able to make predictions on labels even when we don't have original labels but proportional bags. This is a very interesting problem area and will be interesting to the field.\\n\\n# For the experiments, a few notes and comments?\\n1. It's a bit harder to appreciate some of the performance gains here without understanding the error rates around the accuracy. This would be especially good to know for the larger bag sizes where one expects more uncertainty.\\n2. For each of the datasets, it would have been also enlightening to provide what the normal labelled performance would be as it would give an indication of the limit when the bag sizes get smaller.\\n3. I am a bit confused by the k-means scenario. With bag sizes similar to the uniform bagging scenario, would the performance not be more as we are already capturing latent structure with the k-means algorithm? \\n4. Following up on 3, Does this not have implications for actually preserving some privacy?\\n\\n\\n# Other Comments\", \"5\": \"It would be interesting to see an approach like mixup can be combined with the consistency concept in this case. So what do you expect when you now combine examples and train the underlying mixed bag objective.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary of the paper: Learning from label proportions (LLP) is an area in machine learning that tries to learn a classifier that predicts labels of instances, with only bag-level aggregated labels given at the training stage. Instead of proposing a loss specialized for this problem, this paper proposes a regularization term for the LLP problem. The core contribution of this paper is to use the idea of consistency regularization, which has become very popular in semi-supervised learning in the recent years. The regularization term takes a perturbation of an input sample, and then force the output of the original and perturbed sample to be similar by minimizing a KL divergence of the two output distributions. Experiments show the performance of the proposed method under two bag generation settings. The paper also finds empirically that the hard L_1 has high correlation with the test error rate, which makes it an ideal candidate when the user splits the validation data from training data (meaning there are no ground truth labels for each instances).\", \"reasons_for_the_decision_of_the_paper\": \"The proposed consistency regularization term seems to be borrowed directly from semi-supervised learning (SSL) research, and is not really specialized for the LLP problem. Although the discovery that adding this regularization term for the LLP problem increases generalization performance is novel, using this term for data without ground truth labels in image datasets such as CIFAR10/100 and SVHN itself has low novelty, since these datasets satisfy the smoothness assumption used by consistency regularization methods. Another issue is that there are only one trial for every experiment, making it hard for the reader to see if the results are statistically significant or not. For these two reasons, it was hard to give a high score decision for this paper. However, the paper\\u2019s motivation is interesting, and the direction to explore regularization techniques for the label proportion learning problem is important and seems novel for this area. If the regularization method can be extended in a way that relates to the LLP problem, it would make the paper much stronger.\", \"other_minor_comments\": \"In the experiments, the average test accuracy of the last 10 epochs are reported. I was curious if the last epochs have better models than earlier epochs. In many weakly-supervised areas such as noisy labels, it is common that the accuracy goes up very quickly but then gradually decreases. Does this also happen for the LLP problem?\\n\\nIn J_cons, does the expectation need to be over p?\\n\\n\\n***After author response:\\nThank you for answering my questions! Although I have a better understanding of the paper now, I still have the same concerns and would like to keep my score.\"}" ] }

[ "Kacper Piotr Kielak" ]

[ "deep learning", "reinforcement learning", "data efficiency", "DQN", "Rainbow", "SimPLe" ]

{ "note_id": [ "H-ehLtvS7", "SklzXhN3sS", "H1xRhf42sr", "r1ll5A73sB", "BygXZnm2sr", "H1eNYZKRtS", "Sygk5-d0Kr", "ByechtJ6tB", "HJevJKu2_r", "H1xyE8kNdH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment", "comment" ], "note_created": [ 1576798733050, 1573829657665, 1573827253523, 1573826183948, 1573825530694, 1571881340174, 1571877255338, 1571776946004, 1570699486759, 1570137639027 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1814/Authors" ], [ "ICLR.cc/2020/Conference/Paper1814/Authors" ], [ "ICLR.cc/2020/Conference/Paper1814/Authors" ], [ "ICLR.cc/2020/Conference/Paper1814/Authors" ], [ "ICLR.cc/2020/Conference/Paper1814/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1814/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1814/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1814/Authors" ], [ "~Ankesh_Anand1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper makes broad claims, but the depth of the experiments is very limited to a narrow combination of algorithms.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Paper has not been revised\", \"comment\": \"Unfortunately, we were not able to finish revising the paper in the given time and include additional evaluations. Thus, we'll proceed with the initial version of the manuscript.\"}", "{\"title\": \"Response to Official Blind Review #2\", \"comment\": \"Thank you for your thorough review and useful feedback. We identified two points raised in your comment.\", \"point_1\": \"The paper is limited to image-space discrete RL tasks, and does not cover the wide range of other RL.\\n\\nWe partially addressed the second issue in the first point of our response to all reviewers. Please find the details there. \\n\\nIn short, the problem was in the framing of the paper. It aimed to underline the importance of using fair baselines when proposing improvements in data efficiency of reinforcement learning algorithms and not necessarily compare MFRL to MBRL. We mainly found this issue in the studies that focus on image-space discrete RL.\", \"point_2\": \"The proposed algorithm is not fundamentally different from Rainbow.\\n\\nIndeed, that is the case. We did not intend to introduce a novel algorithm. Instead, our goal was to underline the importance of using appropriate baselines for data-efficient reinforcement learning by showing that already existing DQN-like algorithms can be much more data-efficient than it is often portrayed.\"}", "{\"title\": \"Response to Official Blind Review #1\", \"comment\": \"Thank you for your thorough review and useful feedback. Please find our response to the two issues raised below:\", \"issue_1\": \"In addition to comparing OTRainbow(100k) with SimPLe(100k), Figure 2 serves the analysis of the long term effects for each of the versions of Rainbow we introduced, similarly to what (Kaiser et al.) did in section 7.3. We do not aim to compare, e.g., OTRainbow(500k) with SimPLe(100k) because clearly, as you pointed out, it is an unfair comparison.\\n\\nHowever, given that we are currently extending our study to cover another recent advancement in the data-efficient reinforcement learning, analysis backed by this figure becomes less relevant. Therefore, as per your suggestion, we will delete it from the paper.\", \"issue_2\": \"We addressed the second issue in the first point of our response to all reviewers. Please find the details there. \\n\\nIn short, as you have pointed out, the problem was in the framing of the paper. It aimed to focus on advancement in data efficiency for Atari (including MFRL for Atari), and not on MBRL as a whole.\"}", "{\"title\": \"Response to the most common points raised by the reviewers\", \"comment\": \"We want to thank all reviewers for all the time spent on analyzing our papers and for the constructive feedback. It is very much appreciated. We'll try to summarize and address all the concerns below.\\n\\n\\n1. The paper is framed as an analysis of MBRL methods but only compares to the dyna-style algorithms.\\n\\nWe aimed to show that recent advancements in sample efficiency of DRL (as benchmarked by the Atari domain) are solely due to using invalid baselines. I.e., they tend to use Rainbow/DQN in it's original or hypertuned form from Hessel et al. (2018) or Mnih et al. (2015). This form, however, did not focus on data efficiency but can be very easily modified to do so (as presented by OTRainbow). We chose the model-based dyna-like SimPLe as an example because it showed the most impressive results in that area. Nevertheless, we did not want to focus on MBRL but on sample-efficient RL as a whole. We agree that both the title and parts of the body are misleading. Simply when writing this paper, the MBRL algorithms were the main studies we were aware of that reported meaningful improvements over the Rainbow DQN. However, the same point would apply to MFRL for Atari. As an example, Lee et al. (2019) proposed a novel way of using experience replay for DQN that improves DQN's data efficiency. However, while doing so, it also increases the ratio r, without doing the same for the baseline to ensure a fair comparison.\\n\\nWe're in the process of experimentally comparing EBU and Overtrained version of DQN, and we will revise our paper to make our aims clearer. \\n\\n\\n2. Other papers (Holland et al., 2018; van Hasselt et al., 2019) have already introduced similar ideas. \\n\\nWhile we agree it is often the case, there are some missing points that our work tries to cover:\\n\\na) Holland et al. (2018) have already shown that more training updates with the agent's existing experience results in higher sample efficiency of DQN. Nevertheless, Kaiser et al. (2019) mention that SimPLe outperforms the DynaDQN (and thus it outperforms replay-based methods). We clearly show that it is not the case.\\n\\nb) van Hasselt et al. (2019) is a concurrent study that focuses on a comparison of model-based and replay-based algorithms. While the conclusion is similar, our goal, first and foremost, is to underline the importance of using appropriate baselines when introducing more data-efficient algorithms, not to analyze differences between different types of methods. To adhere to our aim of establishing a fair evaluation of data efficiency, we will add the comparison between EBU and Overtrained DQN showing that the problem of wrong baselines is not only relevant to MBRL.\\n\\nd) Given the points above, we believe our study adds an essential aspect to the discourse in data-efficient RL. Especially, given that novel algorithms, although significant in other areas, are praised for their substantial improvements of data-efficiency, even though well established, existing methods are not worse. EBU and SimPLe being an example.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper presents an emprical study of how a properly tuned implementation of a model-free RL method can achieve data-efficiency similar to a state-of-the-art model-based method for the Atari domain.\\n\\nThe paper defines r as ratio of network updates to environment interactions to describe model-free and model-based methods, and hypothesizes that model-based methods are more data-efficient because of a higher ratio r. To test this hypothesis, the authors take Rainbow DQN (model-free) and modify it to increase its ratio r to be closer to that SiMPLe (model-based). Using the modified verison of Rainbow (OTRainbow), the authors replicate an experimental comparison with SiMPLe (Kaiser et al, 2019), showing that Rainbow DQN can be a harder baseline to beat than previously reported (Figure 1). This paper raises an important point about empirical claims without properly tuned baselines, when comparing model-based to model-free methods, identifying the amount of computation as a hyperparameter to tune for fairer comparisons.\\n\\nI recommend this paper to be accepted only if the following issues are addressed. The first is the presentation of the empirical results. In Figure 1, OTRainbow is compared against the reported results in (Kaiser et al, 2019), along with other baselines, when limiting the experience to 100k interactions. Then, in Figure 2, human normalized scores are reported for varying amounts of experience for the variants of Rainbow, and compared against SiMPLe with 100k interactions, with the claim that the authors couldn't run the method for longer experiences. Unless a comparison can be made with the same amounts of experience, I don't see how Figure 2 can be interpreted objectively. In any case, the results in Figure 1 and the appendix are useful for showing that the baselines used in prior works were not as strong as they could be.\\n\\nThe second has to do with the interpretation of the results. The paper chooses a single method class of model-based methods to do this comparison, namely dyna-style algorithms that use the model to generate new data. But models can also be used for value function estimation (Model Based Value Expansion) and reducing gradient variance(using pathwise derivatives). The paper is written as if the conclusions could be extended to model-based methods in general. Can we get the same conclusions on a different domain where other model-based methods have been successful; e.g. continuous control tasks? A way to improve the paper would be to make it clear from the beginning that these results are about Dyna-style algorithms in the Atari domain.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper presents a data-efficient version of the Rainbow DQN by Hessel et al. (2018) that matches the performance of a recent state of the art model-based method by Kaiser et al. (2019) on several Atari games. Particularly, the paper empirically shows that a simple hyper-parameter tuning, in this case increasing the ratio of number of training steps to the environment interactions as well as decreasing epsilon-decay period, can result in significant improvements in sample efficiency of the Rainbow DQN agent. They show that their method (which requires significantly less computation) can outperform the model-based variant on half of the games tested, while performing worse on the rest.\\n\\nOverall, I believe that this paper is below the acceptance threshold due to lack of 1) novelty, 2) significance and 3) depth of analysis.\\n\\nThe observation that more training updates with the agent\\u2019s existing experience results in sample efficiency in DQN method has already been shown empirically by Holland et al. (2018) which has also been cited by this paper. Additionally, more recent work by Hasselt et al. (2019) which is not currently cited, explicitly addresses the same problem as this work by tuning the hyper-parameters of the Rainbow DQN to achieve significant sample efficiency, outperforming Kaiser et al. (2019) in 17 out of 26 Atari games tested. In addition, their work gave a rather detailed motivation and analysis of their findings, proposing and testing several hypotheses for how and when model-based methods could outperform replay-based model-free variants. \\n\\nIn comparison to prior work, the current paper has a more limited scope and significance. Hence, I believe more work would be needed to warrant acceptance.\\n\\nHessel et al., 2018: Rainbow: Combining Improvements in Deep Reinforcement Learning https://arxiv.org/abs/1710.02298\\nKaiser et al., 2019: Model-Based Reinforcement Learning for Atari https://arxiv.org/abs/1903.00374\\nHolland et al., 2018: The Effect of Planning Shape on Dyna-style Planning in High-dimensional State Spaces https://arxiv.org/abs/1806.01825\\nHasselt et al., 2019: When to use parametric models in reinforcement learning? https://arxiv.org/abs/1906.05243\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Do recent advancements in model-based deep reinforcement learning really improve data efficiency?\\n\\nIn this paper, the authors revisit the model-free baselines used in recent model-based reinforcement learning algorithms. And after more careful tuning of the hyperparameters, the model-free baselines can obtain comparable and much better performance using the same number of samples.\\n\\nThe paper is well-written, and the experiments are well-designed to support the claim.\\nHowever, the research contribution of the project is limited to image-space discrete RL tasks, and does not cover the wide-range other RL. In terms of the novelty, the proposed algorithm is not fundamentally different from Rainbow. \\nTherefore I tend to vote for borderline for this paper and am willing to increase the scores if more improvement is updated.\\n\\nBesides, I would like to thank Mr. Ankesh Anand for mentioning the Hasselt et al. (2019) [1] paper, which is indeed very similar in terms of the topic discussed and the methods used to evaluate the algorithms. \\nI have also read Hasselt et al. (2019) before this submission, but I think it would be fair to say the two papers are relatively concurrent.\", \"potential_improvement\": \"- It will be great if the authors also extend the discussion to current reinforcement learning algorithms that are applied in continuous tasks from states. In [2], similar conclusion is observed in continuous control tasks, where SAC [3] / TD3 [4] perform substantially better than many of the state-of-the-art model-based baselines.\\n\\n[1] van Hasselt, Hado, Matteo Hessel, and John Aslanides. \\\"When to use parametric models in reinforcement learning?.\\\" arXiv preprint arXiv:1906.05243 (2019).\\n[2] Wang, Tingwu & Bao, Xuchan & Clavera, Ignasi & Hoang, Jerrick & Wen, Yeming & Langlois, Eric & Zhang, Shunshi & Zhang, Guodong & Abbeel, Pieter & Ba, Jimmy. (2019). Benchmarking Model-Based Reinforcement Learning.\\n[3] Haarnoja, Tuomas, Aurick Zhou, Pieter Abbeel, and Sergey Levine. \\\"Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor.\\\" arXiv preprint arXiv:1801.01290 (2018).\\n[4] Fujimoto, Scott, Herke van Hoof, and David Meger. \\\"Addressing function approximation error in actor-critic methods.\\\" arXiv preprint arXiv:1802.09477 (2018).\"}", "{\"comment\": \"Thank you very much for your comment. Hasselt et. al (2019) indeed reaches a similar conclusion, thank you for pointing it out. We were not aware of this preprint before and the work was done fully concurrently.\\n\\nAlthough both studies achieve similar experimental outcomes, there are some important distinctions:\\n\\n1. The high-level contrast between Hasselt et al. (2019) and our paper is that the former discusses differences between replay and model-based methods and the latter, first and foremost, emphasizes the importance of appropriate baselines in any future work focused on sample efficient deep reinforcement learning (not only model-based). We can agree that the title and related work section can be slightly misleading in that regard. It was motivated by the fact that the model-based approaches were producing the best results at the time of writing. One example of non-model-based work where arguments from this paper would be applicable could be Su Young et al. (2019) that proposed a novel model-free algorithm. It is not included as related work because we were not aware of any existing preprint at the moment of submission and did not want to base the analysis solely on abstract. We plan to include these points once the revision period opens.\\n\\n2. Also, approaches to data-efficient hyperparameters of Rainbow DQN differ. Hasselt et. al (2019) increases the length of the multi-step update, whereas our paper increases the number of training steps per each data sample. Given that both studies employ simple and intuitive but different hyperparameter changes and do not exhaustively tune the algorithm, we believe that together they provide much stronger case behind using appropriate baselines showing that it does not take much to drastically improve the efficiency of existing methods.\\n\\nLee, Su Young, Sungik Choi, and Sae-Young Chung. \\\"Sample-efficient deep reinforcement learning via episodic backward update.\\\" To appear at NeurIPS '19 https://arxiv.org/abs/1805.12375\\n\\nvan Hasselt, Hado, Matteo Hessel, and John Aslanides. \\\"When to use parametric models in reinforcement learning?.\\\" To appear at NeurIPS '19. https://arxiv.org/abs/1906.05243\", \"title\": \"Thank you for sharing the preprint!\"}", "{\"comment\": \"Thanks for the thorough investigation of sample-efficiency of model-free methods in the low-data regime. It seems Hasselt et. al (2019) [1] also reach the same conclusion: a well tuned Rainbow outperforms SimPLE in the low-data regime. They go on to add scenarios where MBRL could be useful.\\n\\nIt's nice that multiple groups reached the same conclusion as it adds credibility to the baseline. I am not sure about ICLR rules regarding this, but there could be an argument to consider this as concurrent work. \\n\\n[1] van Hasselt, Hado, Matteo Hessel, and John Aslanides. \\\"When to use parametric models in reinforcement learning?.\\\" To appear at NeurIPS '19. https://arxiv.org/abs/1906.05243\", \"title\": \"Findings very similar to Hasselt et. al (2019)\"}" ] }

[ "Krzysztof Maziarz", "Mingxing Tan", "Andrey Khorlin", "Kuang-Yu Samuel Chang", "Andrea Gesmundo" ]

[ "hybrid agent", "agents", "architecture search", "potential", "design", "neural networks", "deep reinforcement", "complex architectural patterns", "explore" ]

{ "note_id": [ "mIeQVxPRl_", "HJxgtV33ir", "Bke0AninsB", "BJxpRqs3oB", "HyeRIIo3iH", "Bylvkp8jiS", "HJgAFLz7oS", "S1gae4gzsB", "BklitU6Zor", "BklcQEo-oH", "S1ej9ZoWor", "SJehrtaloS", "BJe6-Up1iS", "HkgQaG51iH", "HyemFkD7cH", "BkxX0Fu0YS", "Byg4gXUTKB", "SylXPYgV_B", "BJgPzQU5PH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "comment", "official_review", "official_review", "official_review", "official_comment", "comment" ], "note_created": [ 1576798733022, 1573860472406, 1573858518483, 1573858005470, 1573856853641, 1573772511450, 1573230213977, 1573155829379, 1573144195266, 1573135393575, 1573134738551, 1573079364224, 1573012997106, 1572999867126, 1572200314664, 1571879370805, 1571803884081, 1570142555146, 1569510158820 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1813/Authors" ], [ "ICLR.cc/2020/Conference/Paper1813/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1813/Authors" ], [ "ICLR.cc/2020/Conference/Paper1813/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1813/Authors" ], [ "ICLR.cc/2020/Conference/Paper1813/Authors" ], [ "ICLR.cc/2020/Conference/Paper1813/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1813/Authors" ], [ "ICLR.cc/2020/Conference/Paper1813/Authors" ], [ "ICLR.cc/2020/Conference/Paper1813/Authors" ], [ "ICLR.cc/2020/Conference/Paper1813/Authors" ], [ "~Qingquan_Song1" ], [ "~Linnan_Wang1" ], [ "ICLR.cc/2020/Conference/Paper1813/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1813/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1813/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1813/Authors" ], [ "~Linnan_Wang1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"Thanks to the authors for the revision and discussion. This paper provides a neural architecture search (NAS) method, called Evolutionary-Neural hybrid agents (Evo-NAS), which combines NN-based NAS and Aging EVO. While the authors' response addressed some of the reviewers' comments, during discussion period there is a new concern that the idea proposed here highly overlaps with the method of RENAS, which stands for Reinforced Evolutionary Neural Architecture Search. Reviewers acknowledge that this might discount the novelty of the paper. Overall, there is not sufficient support for acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Evo-NAS contribution\", \"comment\": \"Thanks for the swift response. RENAS is the only Evolutionary-RL architecture search hybrid that we are aware of, and the approach it took does lead to good performance. However, we believe that the differences between RENAS and Evo-NAS are fundamental.\\n\\nAs we wrote in point 3 of our \\\"Comparison with RENAS\\\" comment, RENAS uses an additional mutation grammar (so, a mutation search space), and the task for RL is to navigate that space. Because of that, we don't see a straightforward way to create an ablation of RENAS containing just the RL part alone (and such an ablation is not covered in the RENAS paper).\\n\\nOn the other hand, Evo-NAS lies on a continuous spectrum. Two extreme points of that spectrum are Evo-NAS's ablations: either having just the evolutionary part, or just the RL part. We believe that, as far as advocating for further exploring this spectrum goes, our paper makes a much stronger point for it than the RENAS paper.\"}", "{\"title\": \"thanks\", \"comment\": \"My overall concern is that the combination of Evolutionary NAS algorithm and RL (or other model to guide the mutation) existed before this work. The contribution of a more detailed study on the combination and/or a slight variant of the combination is not significant enough, in my opinion. Let me know if I misunderstood anything, thanks.\"}", "{\"title\": \"Thanks for the response\", \"comment\": \"Thank you for your comment! Could you elaborate on what you mean by \\\"limited by the scope of the paper\\\" and \\\"better NAS frameworks\\\"? Our understanding is that Evo-NAS is fairly general, so in principle it can be applied to any architecture search space.\"}", "{\"title\": \"Thanks for the follow ups\", \"comment\": \"I would like to thank the authors for the detailed followup comment. Especially the clarification on the difference with RENAS. The goal on studying how RL and evolution work together is great and should be encouraged. However, in my opinion, the totality of the contribution is limited by the scope of the paper.\\n\\nTo improve upon the current version, one possible way is to use the understanding to build better NAS frameworks (on top of the domain specific adaption as mentioned in the comment).\"}", "{\"title\": \"Paper revision after all the reviews and comments\", \"comment\": \"We would like to thank all reviewers and commenters again for their valuable feedback!\\n\\nWe have just uploaded an updated version of our paper, which includes two changes:\\n- We updated the NAS-Bench figure in response to comments from Linnan Wang. We also switched to reporting regret instead of accuracy, and using log scales on both axis - this matches \\u201cNAS-Bench-101: Towards Reproducible Neural Architecture Search\\u201d), making Figure 7 in that paper directly comparable with our new Figure 3.\\n- We expanded the \\u201cRelated Works\\u201d section to describe differences between our work and RENAS.\\n\\nFor a more detailed comparison of Evo-NAS with RENAS, see our comment \\u201cComparison with RENAS\\u201d below. Note that we have recently updated that comment (after reviewing RENAS more closely) to include another important difference (number 3).\"}", "{\"title\": \"Comparison with RENAS\", \"comment\": \"We agree that the parallel work of RENAS is definitely related to Evo-NAS. Below we review some key differences and reasons why we think the main contributions of our work still stand. While both works hybridize RL with evolution, we believe the goals are very different, and the insights complementary. Please let us know your thoughts on this comparison.\\n\\n1) We believe that the purpose of RENAS was achieving SotA accuracy on CIFAR10 and ImageNet with low computational cost. To that end, they proposed a framework which does use both RL and Evolution to produce child models. However, RENAS is also augmented with several domain-specific optimizations and improvements that could be complementary to many possible search algorithms, including:\\n- a complex way of proposing new mutations by using attention and a bidirectional LSMT encoder ran on the architecture description\\n- reloading weights from the parent model into the child model (related to the approach in ENAS)\\n- a refined reward function: instead of using validation accuracy directly, it is rescaled and passed through the tan() function\\n- improved search space (not reused from any prior work)\\n\\nWe think that likely all of these different ideas, together with combining RL with Evolution, contribute to RENAS achieving good results. In contrast, our aim was not to necessarily push Evo-NAS to its limits in terms of accuracy or computational savings - in fact, the improvements listed above could also be applied to Evo-NAS. We explicitly chose not to do that (for example, not to reload child weights or tune the search space), because our goal was to study how RL and evolution work together. The RL and Evolutionary baselines that we compare with are simply ablations of Evo-NAS, so the fact that Evo-NAS outperforms them both really shows that there is something to be gained from both worlds.\\n\\n2) Since our aim was to show that combining learning and evolution leads to gains across domains, we experiment with many: image (ImageNet, NAS-Bench), NLP, and synthetic. While we focus on architecture search, we also believe that our results hint that Evo-NAS would show strong performance in different (although similar) discrete optimization problems - and we would be happy to see follow-up works of this kind. In contrast, RENAS approach strongly focuses on the image domain, and applying it to other domains is a future research opportunity.\\n\\n3) Evo-NAS learns a policy directly on the architecture search-space, while RENAS learns it over the space of possible mutations applicable to a given parent model. Thus, the application of RENAS requires to define an additional mutation grammar and a more complex agent (e.g. composed of 4 distinct mutation generation sub-steps).\"}", "{\"title\": \"additional comparison should be provided\", \"comment\": \"Thanks for the comment.\\n\\nAfter reading others' comment, it seems that similar ideas have been published, e.g., RENAS. I think it is important to compare against these newer baselines with similar ideas to fully justify the contribution, which in my opinion is the consistent good empirical result.\"}", "{\"title\": \"Thank you for your review\", \"comment\": \"Thank you for your review! We address each of your concerns below. Please let us know if you have some further comments - we are happy to make changes to the paper to improve it.\\n\\n1. While the gains are not enormous, we agree with Reviewer #2 in saying that they are consistent across a range of different tasks: both in real-life AS scenarios (NAS-Bench, NLP, ImageNet), and in the synthetic task. However, making sure that differences between results are statistically significant is certainly an important point.\\n\\nFor example, note that for NLP (Table 1), we ran 10 runs for each (approach, dataset) pair. We claim that Evo-NAS matches the better of the baselines in all cases, and in 3 out of 7 it outperforms the best baseline with statistical significance. We made this claim for the three datasets where the difference exceeded twice the standard-error-of-the-mean. This can also be easily translated into p-values: for example, for the ConsumerComplaints dataset, the p-value for Evo-NAS\\u2019s gains over either baseline being statistically significant is < 0.02 (so we reject the null hypothesis that there is no gain).\\n\\nWe used simple statistical measures such as the ones described above to assess significance. However, we are open to performing additional nonparametric tests, such as the Wilcoxon Signed-Rank Test you proposed. Please let us know what would be the exact test that you had in mind (i.e. what would be the null hypothesis). For example, we could consider pairs of (average result of Evolution, average result of Evo-NAS) over all benchmarks considered in the paper (there are 10 in total since we should count each NLP dataset separately), and perform the test on these pairs. In that case, the Signed-Rank test does show significance, as in all cases the mean result of Evo-NAS is strictly better than the mean result of Evolution.\\n\\n2. You are right in saying that there are many smart evolutionary algorithms out there. However, to our knowledge, the NAS domain is dominated by \\u201cnon-adaptive\\u201d evolution (i.e. with uniform prior over mutations), which also seems to perform very well across different AS domains. Since here we focus on architecture search, we compare our method to two strong, common and most recent baselines (the combination of which gives Evo-NAS).\\n\\nNote that our hybrid may not necessarily be the best algorithm of this kind; our intention was to show that combining these two worlds (learning-based and evolution-based) in a relatively simple way we can get consistent gains across multiple domains.\", \"one_kind_of_follow_up_works_that_we_wish_to_inspire_is_exactly_what_you_are_proposing\": \"bringing different \\u201csmart\\u201d genetic algorithms from the evolution community to the NAS community, and investigating their behavior on architecture search. It is surely an exciting direction, although one that is not the focus of our current work.\"}", "{\"title\": \"Thank you for your review\", \"comment\": \"Thank you for your comments and suggestions! Please let us know if you have any more feedback about our work, since we would be happy to use it to make the paper better. Below we specifically address each of the points that you made.\\n\\n1. We agree that the synthetic task is far from real NAS applications. In our paper we do focus on NAS, and the other lines of experiments (NAS-Bench, NLP datasets, ImageNet) are all real-life architecture search scenarios. However, as we hint in the \\u201cConclusion\\u201d section, it is possible to apply Evo-NAS to tasks outside of architecture search. Discrete optimization problems such as our toy task could be one example, and we see that Evo-NAS also performs very well in this domain. We think that makes the approach even more promising in terms of inspiring future work.\\n\\n2. As we understand the proposal, you would consider running the NN learning component not from the beginning of the experiment, but start learning later on. This would save some computational cost; are we missing any other advantages of this approach? Our initial intuition is that the savings might be limited due to low overhead of running the NN compared to the training of candidate architectures. However, it would be interesting to experiment further in this direction in the future.\\n\\n3. You made a good point that we could use something simpler than a Neural Network for the learning part of Evo-NAS. Since we wanted our approach to be comparable to existing methods, we used a NN, since it\\u2019s somewhat standard in the NAS community. Also, from a theoretical standpoint, a NN has a useful property of being a universal approximator. On the other hand, it does seem likely that guiding the evolutionary search (i.e. producing a good prior over mutations) might be an easier problem than learning the structure of all solutions (i.e. producing a good prior over the entire search space). \\n\\nTherefore, it\\u2019s certainly possible that using something simpler as the learning part of Evo-NAS would be sufficient to get good results. However, again note that in terms of computational cost the NN part is negligible - so any gains would not be in terms of compute, but perhaps in enhanced explainability of the learning component.\"}", "{\"title\": \"Thank you for your review\", \"comment\": \"Thank you for you comments! Below we specifically address both of the points that you raised. Let us know if you have any more concerns - we are happy to make improvements to the paper.\\n\\n1. To make the assessment that the methods have converged, we looked at the best reward attained (right plots), not reward running average (left plots). By \\u201creward running average\\u201d, we mean sliding a constant size window over the reward, and computing the average reward in each such window. This measure is not a good metric to compare the algorithms, and we use it just to show how much the reward varies from sample to sample for different methods.\\n\\nNote that learning-based methods (like Neural PQT which you mentioned) do see small improvements in running average reward even long after they stop exploring new parts of the search space - this is simply because they keep making their distributions over actions \\u201csharper\\u201d. Therefore, it is more informative to look at best reward attained over time.\\n\\nWith that in mind, we review below the experiments that you pointed to - that is, Figures 2 and 3. You also mentioned Figure 1, which does not correspond to any experiment, so we assumed this to be a typo - but let us know if you meant something else there, and we can discuss further.\\n\\n- For NAS-Bench (Figure 3), the methods see only marginal improvements at the end of the experiment.\\n\\n- For the synthetic task (Figure 2), all methods see small improvements of the reward if ran longer, but the overall pattern remains the same (Evo-NAS (PQT) attaining the best results, matched by Neural (PQT) from around 4500 trials onward).\\n\\n2. Our understanding is that PQT provides a stronger training signal than Reinforce. This comes at a cost, since Reinforce is known to be unbiased, and therefore it produces gradients that (in expectation) match the gradient of the expected reward. On the other hand, PQT is a heuristic, and it gives no theoretical guarantees.\\n\\nWe wouldn\\u2019t say that PQT is always better than Reinforce - this is likely very domain dependent. Our focus is architecture search, where the number of trials is typically low given the complexity of the search spaces. In this domain, as seen in our experiments, trading off theoretical guarantees and lack of bias for a stronger and more greedy training signal leads to better results.\\n\\nNevertheless, comparing PQT to classical RL algorithms (like Reinforce, but also many others) is a very interesting research direction. Since our main goal was hybridizing RL-based and Evolution-based approaches to architecture search, we did not explore this direction in our work.\"}", "{\"title\": \"Thank you for pointing out a related paper\", \"comment\": \"Thanks for directing us to this work, it does look related to our approach. After reviewing the paper, it looks like it also hybridizes learning and evolution, although in a somewhat complicated way, making it harder to reason about the source of accuracy gains. Nevertheless, we are glad to see that these kinds of ideas are an active area of research. When we update our paper, we will mention RENAS in the \\\"Related Works\\\" section.\"}", "{\"title\": \"Potential Baseline to Compare with?\", \"comment\": \"Hello there,\\n\\n Nice paper! It would be great to provide some comparison or introduction of a related work:\", \"renas\": \"Reinforced Evolutionary Neural Architecture Search (CVPR 2019)\\n\\nsince it also proposed a mixture approach to conjoint the advantage of reinforcement learning and evolutionary method.\"}", "{\"title\": \"Thank you\", \"comment\": \"Thank you for your response. It will be great if the authors can compare their algorithms to prior works (LaNAS [1], AlphaX[2] and several peer submissions [3][4]) on NASBench following a similar standard, i.e. using the metric of #samples to global optimum to fairly evaluate the search efficiency.\\n\\n[1] Wang, Linnan, et al. \\\"Alphax: exploring neural architectures with deep neural networks and monte carlo tree search.\\\" arXiv preprint arXiv:1805.07440 (2018).\\n\\n[2] Wang, Linnan, et al. \\\"Sample-Efficient Neural Architecture Search by Learning Action Space.\\\" arXiv preprint arXiv:1906.06832 (2019).\\n\\n[3] https://openreview.net/forum?id=rJgffkSFPS\\n\\n[4] https://openreview.net/forum?id=B1lxV6NFPH\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"In this paper, the authors proposed to combine both NN-based NAS and Aging EVO to get the benefit of both world: good global and local sample efficiency. The main idea is to use Aging EVO algorithm to guide the overall search process, and a NN predicting the final performance is used to guide the mutation process. The combined EVO-NAS has showed consistent good performance over a range of tasks.\\n\\nOverall, while the novelty of the paper is not exceptional, since it is a rather straightforward combination of two existing approaches, the end-results is promising. I would like to see more in-depth analysis on the combined algorithm to validate authors' hypothesis on why EVO-NAS works better. More detailed comments can be found below.\\n\\n1. The experiment on the synthetic task is not very helpful, since the domain can be far apart from the real NAS applications. One evidence is that the NN significantly outperform EVO in this task but not in the other tasks.\\n\\n2. In difficult tasks, the proposed EVO-NAS and the original Aging EVO are very close in the first few hundreds of trials, however, later the gap remains the same. It would be interesting to see if we can eliminate the gap by adding the NN component in the middle of the Aging EVO experiment (e.g., at the point of 2000 trials).\\n\\n3. I am also curious about the difference between the NN learned from Neural agent vs. those learned from EVO-NAS agent. Moreover, do we need a NN for the EVO-NAS? Or something simpler would be sufficient to guide the search.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper is well organized. The applied methods are introduced in detail. But it lacks some more detailed analysis.\\n\\nMy concerns are as follows.\\n1. The performance differences between Evolutionary agent and EVO-NAS agent seem not significant. Please conduct additional statistical tests such as the Wilcoxon signed-rank test to verify the performance improvements are significant.\\n2. Many studies have been conducted to automatically adjust control parameters such as crossover and mutation probabilities in evolutionary algorithm literature. It would be better to compare one of these approaches in the experiments.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"It is a nice paper that combines the deep reinforcement learning and evolutionary learning techniques to neural architecture search problem. Experimental results are promising. However, I still have some concerns on the current submission.\\n1.In Fig 1,2 &3, it seems that the performances of Neural (PQT) keeps increasing. For better compassion, we recommend the authors reports the performances of compared algorithms until they are convergent.\\n2.The different training algorithms (Reinforce and PQT) have difference performances whether because different training algorithms converge to difference local minima or stationary points.\"}", "{\"comment\": \"Thank you for your comment!\\n\\nFigure 3 (right) actually shows the best moving average of reward over time (i.e. first the moving average was applied, and then maximum over time was computed). This lowered the results for all agents (especially for Random Search) and made Figure 3 not directly comparable to the results from the NasBench paper. Note that RS is not stuck, it's just progressing very slowly.\\n\\nAll the other \\\"best reward\\\" figures in our paper correctly show the best reward over time.\\n\\nThank you for pointing out this discrepancy. As soon as it is possible to update the paper, we will upload a version with Figure 3 regenerated. Apart from showing best reward, we will also convert it into the log regret format used in the original NasBench paper.\", \"title\": \"Good catch\"}", "{\"comment\": \"Hello there,\\n\\n In your figure Fig.3(right), random search seems stuck in a local optimum, reflected by the plateau. I suspect this figure is not from sufficient runs. In theory, random search should keep consistently increasing until converging with Evo-NAS; and it will find the best in expected N/2 trails (N is the size of NASBench). It is questionable for random search to reach plateau in only 5000 trails on a dataset that has 420K samples. Thank you.\", \"title\": \"Fig 3 is misleading\"}" ] }

[ "Yongheng Zhao", "Tolga Birdal", "Jan Eric Lenssen", "Emanuele Menegatti", "Leonidas Guibas", "Federico Tombari" ]

[ "3d", "capsule networks", "pointnet", "quaternion", "equivariant networks", "rotations", "local reference frame" ]

{ "note_id": [ "boqsHBeQww", "SJlKIDS2sr", "rygxhMlYjH", "rJg8bGeYiH", "SyeRw-lYsH", "rJgTqxck5S", "SklKk4uy9B", "rkgmo9psKS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732995, 1573832528677, 1573614248360, 1573614078482, 1573613925831, 1571950740566, 1571943392939, 1571703450924 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1812/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1812/Authors" ], [ "ICLR.cc/2020/Conference/Paper1812/Authors" ], [ "ICLR.cc/2020/Conference/Paper1812/Authors" ], [ "ICLR.cc/2020/Conference/Paper1812/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1812/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1812/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper presents a capsule network to handle 3d point clouds which is equivariant to SO(3) rotations. It also provides the theoretical analysis to connect the dynamic routing approach to the Generalized Weiszfeld Iterations. The equivariant property of the method is demonstrated on classification and orientation estimation tasks of 3D shapes.\\nWhile the technical contribution of the method is sound, the main concern raised by the reviewers was the lack of details in the presentation of methodology and results. Although the authors have made substantial efforts to update the paper, some reviewers were still not convinced and thus the scores remained the same. The paper was on the very borderline, but because of the limited capacity, I regret that I have to recommend rejection. \\nInvariances and equivariances are indeed important topics in representation learning, for which the capsule network is known as one of the promising approaches but still not well investigated compared to other standard architectures. I encourage authors to resubmit the paper taking in the reviewers' comments.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thanks\", \"comment\": \"Thank you for your detailed comments and the responses all make sense to me - very happy to remain supportive of the paper.\"}", "{\"title\": \"Our response to Official Blind Review #3\", \"comment\": \"We thank the reviewer for the specific comments and acknowledging the improved results we achieve in the paper. Please find our responses below.\\n\\n1. \\u201cThe activation\\u201d: It is true that the update of alpha is only included in the algorithm as it is a standard result from previous works [Sabour et al. 2017a, Lenssen et al., 2018]. The input to Alg 1. are the alphas of the previous layer and we set the initial activations to 1. We have now clearly indicated that in the last paragraph of Sec.3.2. \\n\\n2. \\u201cThe transform t()\\u201d: We have significantly increased our explanation and reworked the formalization of the transform network in the paper. The t() network does not directly compute the vote. It is merely producing the transformation, on which the input pose is applied to produce the vote: v = q t(). It is the continuous analogue to the approach in the original capsule networks by Sabour et al. There, those transformations lie in a discrete kernel window and are directly optimized like weights for convolution. Here, we train a continuous kernel function instead, given that input capsules are attached to points that lie in continuous space and not on a fixed grid. The function t() is shared for all input points and produces transformations for all combinations of input capsules from the respective point and output capsules, which we clarified in the first paragraph of Sec.3.2. Additionally, we added specific details of the architecture with pseudocode in Alg.3. \\n\\n3.\\u201dLoop in Alg.1\\u201d: Thank you for making us aware of this error. We fixed the algorithm in the updated version.\\n\\n4.\\u201dLimited space for explanation\\u201d: We have now shortened the background section to give room for further explanations. The paper is also slightly longer. As mentioned above, we have added more explanation of the network architecture in Sec.3.2 (added alpha and t), Sec.4 and appendix(Alg.3). In Alg.3, we explain the details of the QE-net with pseudocode, included the \\u201ct(.)\\u201d and how the DR is used after we got the transformations.\\n\\n5.\\u201d more detailed comments\\u201d: Thanks for those corrections. We have now incorporated them.\"}", "{\"title\": \"Our response to Official Blind Review #1\", \"comment\": \"We thank the reviewer for finding our problem important and our approach novel. Please find our responses below.\\n\\n1. \\u201cLimited space for architecture details\\u201d: We have now slightly shortened the background section and rewrote the descriptions about the architecture, the t-network and QE-network in Sec.3.2. The specific details of the architecture are also presented by the pseudocode in Alg.3 in order to make the work easier to replicate. We also increased the implementation details in Sec. 4. \\n\\n2. \\u201cHow to obtain Q_i from X_i\\u201d : This is done by computing local reference frames. We have now mentioned this explicitly.\\n\\n3. \\u201cHow to create LRF with point sampling\\u201c: In section 4 we have indicated that we use uniformly sampled points and cite the related literature that we utilize. While our work has theoretical equivariance properties (we do not discretize), it is affected by the choice of sampling. This is the reason why we included the resampling (point density) experiment. The LRFs are calculated for each patch using a center point and its neighbors. This is done before we perform uniform sampling to get the pooling centers hierarchically. The uniform sampling is not rotation invariant and hence might result in different pooling centers and this, as we have mentioned, is the primary source of error in our network. We have evaluated how changes in sampling affect our network in the last paragraph of the experiments section (Table 3). Note that during training, we feed the network with multitudes of different samplings and as a result, the network is trained to deal with the variance caused by this kind of data discretization. We also added Fig.8 in the appendix to show the robustness of our proposed algorithm by computing the LRF with different point densities.\\n\\n4. \\u201cObject centroid for LRF computation\\u201d: For computing the LRFs we do not need a known centroid. These computations are done only locally. However, to transform them into the same coordinate frame we assume a common origin. Note that this does not have to be the center of the object, even though we use that piece of information in the paper.\\n\\n5. \\u201cEquivariance towards non-linearities\\u201d: Note that all the non-linearities are captured in the weights. Our proof is independent of the weighting scheme and we prove equivariance for an arbitrary (up to the provided conditions) choice of real weights. Also chaining k-consecutive equivariant operators would still preserve equivariance. This is a standard result and we have mentioned this in Dfn. 2. Nevertheless, we have mentioned this again when we speak of the architecture. \\n\\n6. \\u201cLRF after shape transformation\\u201d: There might be a misunderstanding here. LRFs do not need to be the same. In fact, they are expected to transform equivariantly, e.g. the surface normal of a point rotates with R when the object is rotated by R. We do not assume that LRFs are given to us. We compute them by FLARE as we have aptly cited. It is worth mentioning again here is that we now report more results in Fig.8 of the appendix with the LRFs calculated in different point cloud densities to show the robustness. \\n\\n7. \\u201cClassification performance\\u201d: In the paper, we already reported one such work \\u201cPoint2Seq\\u201d that achieves 92.6% performance. However, most of those methods are not truly equivariant to rotations. This is why they suffer under the NR/AR case. However, our equivariant network can classify shapes in random rotations even only trained with shapes of such categories in single orientation(aligned). Note that our contributions can also be applied to more complex networks. As we use PointNet-like layers, we believe that it is best to evaluate against the methods of a similar nature. On another aspect, note that we also estimate canonical object poses, and not only the pairwise ones.\"}", "{\"title\": \"Our response to Official Blind Review #2\", \"comment\": \"We thank the reviewer positive feedback, going through the theoretical analysis and constructive criticism. We have revised the manuscript according to her/his concerns.\\n\\n1. \\u201cOn the use of quaternions\\u201d: It is true that multiple representations for rotations do exist and seem potentially suitable for the task at hand. While we do not theoretically know whether the quaternions are indeed the \\u201cbest\\u201d choice (or if any such choice ought to exist), we do have multiple strong motivations for why quaternions (being one of the long-standing standards) are one reasonable choice: 1. Single redundancy, 2. No singularities, 3. Closed form mean that is differentiable and batch-friendly (see source code). 4. As acknowledged by the reviewer, the employed mean also allows us to draw connections to existing IRLS methods (Weiszfeld) enabling us to provide theoretical guarantees. \\n\\nThanks for bringing this important point up. We have now included this discussion in Sec. 2.2 of the main paper.\\n\\n2. \\u201cSensitivity to point density\\u201d: We agree that this is an important issue. We have now included an ablation study in the main paper that analyzes the sensitivity of the local reference frame we employed. The results are tabulated in Tab. 3, where we gradually decrease the number of points used in the LRF computation besides the point set size used in the actual network. This way, we show that even when the number of points is reduced to 2K and 1K respectively, we are still able to get acceptable results.\\n\\n3. \\u201cPresentation of results\\u201d: We have now included error bars and more details of the classification result as suggested in Fig.6 and Fig, 7. ModelNet is a standard dataset that is used by many works to measure the performance. We have chosen that dataset for fairer comparison to the existing works. As others, we believe this is adequate. Although our experimental evaluation follows the standard protocols of Qi et. al. 2017, we now included more details on how we prepare the data in the appendix and main paper. To summarize:\\n 1> \\\"Training and test data set-up\\\": We used the official split for training and testing.[Qi et. al. 2017b] \\n 2> \\\"Objects used in training and test?\\\" : The objects in training and testing dataset are different, but they are from the same categories so that they can be oriented meaningfully. \\n 3> \\\"Protocol for the AR test\\\": During testing, we randomly generate arbitrary SO(3) rotations for each shape in the test dataset and do the classification evaluation. \\n\\n4. \\u201cMore on pose estimation\\u201d: The literature is mainly dominated by three types of works: 1. Using meshes/multiviews as input representations, which might not be available under noisy real scenarios, 2. Using equivariance to yield invariant representations but not explicitly evaluating the rotation errors (such as Tensorfield Networks), 3. Proof-of-concept equivariant networks that are quite challenging to run on the large datasets we have. Nevertheless, in order to address the concern of the reviewer, we have now included an additional evaluation against the recent work of Iterative Transformer networks (IT-Net) that can simultaneously classify and estimate the pose of an object. The results can be found in Table 2 as well as in Figure 6 of the appendix. Note that we do not use data augmentation during training. The relevant discussions are added to the manuscript. \\n\\n5. \\u201cRegarding limitations\\u201d: Other than symmetries, we do not observe any particular pattern in the way the errors are made in our network at a first glance. We leave it as future work to thoroughly analyze the characteristics of the errors and hope that our evaluations are enough indicators of the performance of our network. \\n\\n6. \\u201cFurther notes\\u201d: We fixed C_i to {q_i} now and made the notation in the figure consistent with the text. The illustrations represent architecture. As our network is quite different from a typical one, we tried to clarify these in the text. We also moved the figure to page 4 where its details are given and included Figure 2 as the initial figure, which summarizes the idea of aggregating LRFs. We also double-checked the consistency of the figure, algorithm and text. Thanks for mentioning these. As suggested by the reviewer, we now included the algorithm corresponding to the QE-network in the appendix (Alg. 3). We also added more details in the architecture section(Section 3.2). \\n\\nWe have now removed the definition of d_{riemann} as it was practically not used. We keep the geodesic distance expressed in terms of quaternions. We now consistently use the circle operator to refer to the Hadamard product and no-operator implies a dot product. Thank you for bringing this up.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This work builds a capsule network, for use with point cloud data, that has units that are equivariant to SO(3) 3D rotations. I found that the authors made the case for rotation equivariance well and I liked the analysis of the dynamic routing approach and its mapping to the Generalized Weiszfeld Iterations.\\n\\nAs a caveat to the review I should point out that while I am familiar with capsule networks, they are not my main area of expertise (I do work on geometry however) so this should be taken into consideration - apologies if there is literature I have missed. \\n\\nMy main concern is that group equivariance for Capsule networks has been studied before - I acknowledge that SO(3) can present additional challenges but would have expected a work that specifically targeted SO(3) to investigate the relative benefits of the different approaches to encoding the lie algebra and not just quaternions. How do we know that quaternions are the best fit for the Capsule framework? Could the authors show empirical results that indicate that other representations for the equivariance do not work as well? My experience from optimization is that the best representation can often be application specific and it would be very helpful to understand their relative merits for this sort of architecture.\\n\\nThe preparation of the local patches and the estimation of the rotations seems very important to these applications to me and would be worthy of more discussion and an empirical presentation of their different merits. I know the authors discussed the issue of the second axis but it would seem worth including an experiment to detect the sensitivity in a similar manner to the point cloud density.\\n\\nMy biggest concern about the paper is the presentation of the results. Please could the authors provide appropriate error bars for the tables - this is the least action necessary to at least start to estimate significance. Ideally something like histograms or violin plots would be more useful. The use of a single dataset makes it hard to identify the efficacy of the approach - especially when it is not clear how the training and test data are set-up. Are the same objects used in training and test but with different rotations? Or different objects from the same categories. What is the protocol for the AR test set?\\n\\nI agree that table 1 confirms the equivariance so that is great but it is less clear that the network is suitable for the pose task since the baselines seem very simple for the pose estimation experiment (and the related work covers a lot of prior work in this area) and again, without error bars it is hard to judge significance.\\n\\nIt is nice that the authors discuss limitations and I agree with the comment on symmetries - would it be possible to include an illustrative example to confirm these suspicions?\", \"other_notes\": \"I have no issue with the proofs - they all seemed to make sense to me.\\n\\nI found Figure 1 (and to a lesser extent Figure 3) to be unclear and uninformative - the notation doesn't match the main part of the paper (also where does C come from in the caption?) and its not clear what the illustrations are meant to represent. Would it not be much more clear to include something like Algorithm 1 instead of Figure 1 which is much more clear and easy to replicate the precise process.\\n\\nI think there is a lot of material in the appendix that should really be in the main paper - I'm happy for the proofs to be in the appendix but I think it is a bit wrong to essentially violate the page restrictions by moving important related work into the appendix. The whole of section 2 is standard text-book information about quaternions - would it not be more appropriate for that to be in the appendix and the related work to be in the main paper?\\n\\nCould the authors indicate the use of the Frobenius norm in eqn (5)? There also seems to be a mixing of notations for inner products through-out the paper and it might be helpful to standardize this?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": [\"This paper proposes a quaternion equivariant network for 3d point clouds. It builds on the capsule networks and dynamic routing, but instead of using arbitrary 4x4 transforms, they propose to use quaternions to represent 3D rotations.\", \"Invariances and equivariances are very important in DNN for classification/detection, so extending these properties to 3d point clouds based applications is interesting and important. The proposed approach is novel and seems to be equivariant to SO(3) rotations, translations, and set permutations.\", \"However, there are a couple of major negatives in the paper. The first is that the details of the architecture are very unclear. A large amount of space is allocated to discussing equivariant properties but not enough is dedicated to the actual implementations of the network. In particular, only on page 5 in section 3.2 is the QE network detailed discussed.\", \"How does one obtain Q_i from points X_i ?\", \"How are the points subsampled to create LRFs? is there a clustering first step? If so, how would that clustering be invariant to global rotation?\", \"Does the method require a known centroid of the object point clouds in order to compute the quaternions of the LRFs?\", \"Theorem 1 has shown that equation 6 of Alg. 1 is equivariant, but are the multiple interactions of DR equivariant? In particular, after k iterations of DR as presented in algorithm 1, with nonlinear sigmoid activations, are the output capsule representations equivariant?\", \"How do you guarantee that the LRF\\u2019s are the same under any global rotation+translations of the point cloud? Or do you assume the same LRFs are known?\", \"The main selling point for equivarince and learning the separation of pose and object class representation is in the hope of more accurate classification. However, empirical classification results are far from the state-of-the-art.\", \"As a side question, there are other methods listed on the ModelNet40 website with performances in the high 90\\u2019s, are there reasons those results are not included as a part of Table 1?\"]}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": [\"This paper presents a rotation-equivariant model for use with point-clouds. Capsule networks with activation and rotation pose estimate in the form of a quaternion are applied to local patches of 3D points, and combined up through progressively more coarser receptive fields. Rotation equivariance is obtained by looking for agreement between input point frames (or capsules), by iteratively clustering using the quaternion mean. The method is evaluated on ModelNet40, obtaining much improved results in the case where rotation orientation is arbitrary.\", \"Unfortunately, I found too many important parts of the method difficult to understand, enumerated below, and am also not very clear on the details of how they fit together. At a high level, the approach of aggregating pose agreement with quaternion averages makes intuitive sense, the quaternion average step itself is described well, and the experimental results seem to corroborate this (results on NR/AR condition are very good, and also near identical to NR/NR condition, a major verification of the method). However, the rest of the pieces of the method still leave me guessing too much.\", \"It isn't clear what computes the activations alpha. I don't see a description for how to compute activations alpha that are the input to Alg. 1, unless these are simply the activations of the previous layer's capsules, and constant for the initial point frames?\", \"Likewise, the transform t() could use more description. Sec 3.2 says it is R3->R4, indicating that the input to t() is a single point. Does this mean that the mere presence of a point in a location R3 (relative to a mostly-canonicalized pose) counts as a vote for a rotation pose? And if so, how does does this work? Or does the regressor actually take multiple points as input? According to the picture in Fig. 1, it looks like it takes all points in the local patch as input.\", \"Algorithm 1 outer for loop for i doesn't seem quite right. It suggests that the body is performed either independently or in sequence for each i. I believe what may actually be intended, is that all i are used at once in parallel, so that the v_ij votes assignment is found for *all* i in order to use in finding the cluster mean for each j. That is, \\\"v_i,j = Q_i * t_i,j for all i\\\", rather than having the outermost loop over i. Is this correct?\", \"Much space is dedicated to background explanations of quaternions, geodescic distance, etc. Given my difficulty in understanding much of the system, it seems like some of could be shortened or put into the appendix, and more of the main text devoted to more detailed explanations of each component in the system (e.g. alpha and t), and how the capsule agreement step (QE DR) makes use of these.\", \"Also, some more detailed comments/suggestions:\", \"Alg.1: output alpha^hat_K, seems K should be M here?\", \"Table 1: \\\"Right hand side denotes symmetric objects\\\" seems it actually shoudl be in the caption for Table 2.\"]}" ] }

[ "Sang Michael Xie*", "Aditi Raghunathan*", "Fanny Yang", "John C. Duchi", "Percy Liang" ]

[ "data augmentation", "adversarial training", "interpolation", "overparameterized" ]

{ "note_id": [ "hEZEkuBTlX", "rJxfPtsnsH", "Bkxjoce2jr", "HygoNcl2ir", "B1xv0Fx3sr", "rJexjFlhor", "BkeSIRjNqr", "Sken8X2g5r", "r1eQCCNUYH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732966, 1573857625852, 1573812898823, 1573812786913, 1573812687232, 1573812631549, 1572286029294, 1572025172214, 1571339979300 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1809/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1809/Authors" ], [ "ICLR.cc/2020/Conference/Paper1809/Authors" ], [ "ICLR.cc/2020/Conference/Paper1809/Authors" ], [ "ICLR.cc/2020/Conference/Paper1809/Authors" ], [ "ICLR.cc/2020/Conference/Paper1809/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1809/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1809/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper describes situations whereby data augmentation (particularly drawn from a true distribution) can lead to increased generalization error even when the model being optimized is appropriately formulated. The authors propose \\\"X-regularization\\\" which requires that models trained on standard and augmented data produce similar predictions on unlabeled data. The paper includes a few experiments on a toy staircase regression problem as well as some ResNet experiments on CIFAR-10. This paper received 2 recommendations for rejection, and one weak accept recommendation. After the rebuttal phase, the author who recommended weak acceptance indicated their willingness to let the paper be rejected in light of the other reviews. The reviewer highlighted: \\\"I think the authors could still to better to relate their theory to practice, and expand on the discussion/presentation of X-regularization.\\\" The main open issue is that the theoretical contributions of the paper are not sufficiently linked to the proposed algorithm.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response\", \"comment\": \"Thanks for your response.\"}", "{\"title\": \"Author Response to R3\", \"comment\": [\"We thank R3 for their feedback, questions and suggestions. We have significantly improved our presentation of both the theoretical results and the contributions and positioning. We request R3 to refer to the general comment above and evaluate the revised paper.\", \"Addressing the major comments,\", \"\\u201cSample covariance of augmented dataset\\u201d: In the previous version, we used \\\\Sigma_std to denote empirical covariance of standard points and \\\\Sigma_ext for the empirical covariance of additional training points. \\\\Sigma_aug did not correspond to the empirical covariance of augmented data. We defined \\\\Sigma_aug to have a scaling of n in the denominator because we wanted to analytically compare (as a function of n) what happens with n points from the standard distribution with and without the extra \\\\alpha n points. We note that these are just notational issues and do not affect the correctness of the results. However, as R3 points out a possible confusion in notation, we changed the presentation in our revision to work with only the unnormalized quantities, thereby resolving the confusion about what is the right normalization quantity. In our revision, \\\\Sigma_std = X_std^\\\\top X_std; \\\\Sigma_ext = X_ext^\\\\top X_ext; \\\\Sigma_aug = \\\\Sigma_std + \\\\Sigma_ext.\", \"\\u201cx_ext = [e_1; e_2] and hence bias upon augmentation in the 3D example\\u201d: Unfortunately, we had a typographical error in our description. In our simple 3D example, the augmented point should be the single point x_ext = e1 + e2, such that the nullspace of the original+extra data is nonzero and the augmented estimator is still biased. The picture and following text reflect the correct augmentation (e1+e2). We apologize for this typographical error and thank R3 for pointing it out. We have fixed this in our revision.\", \"\\u201cSigma_aug\\u201d: We have changed notation to make this clearer. We note that all our results only depend on the column spaces and null spaces of \\\\Sigma_aug and \\\\Sigma_std. The scaling does not change anything. However, we do acknowledge the source of confusion and have revised accordingly.\", \"\\u201cClarifications about Fig. 1\\u201d: We have now moved the plot in Figure 1 (right) to Figure 4. The spline simulation details are present in Appendix G. Each point in the spline subsampling experiment is an average of 25 trials. In the middle pane of Fig. 1, every training point has an accompanying green point that is a local transformation. In general, we allow arbitrary augmentations in our framework, including local perturbations, and we clarify the augmentations which fall in the framework in the revision.\", \"\\u201cFocus on adversarial training from the intro\\u201d: We have changed our introduction and presentation more generally to portray adversarial training as the motivation of our study\", \"\\u201cBaseline comparisons\\u201d: We have added more baseline comparisons and different perturbations - see the general comment (d) for more details.\", \"We also thank R3 for carefully pointing out minor comments. We note that we have fixed all these in the revision.\"]}", "{\"title\": \"Author Response to R1\", \"comment\": [\"We thank R1 for the feedback and their thoughtful summary. We have incorporated the suggestions for improvement in our revision. We request R1 to refer to the general comment above and evaluate the revised paper. In particular, we have clarified our setting of data augmentation and added more experiments and baselines for X-regularization.\", \"Regarding specific comments,\", \"\\u201cCrosses unclear in figure\\u201d: Addressed in Fig 1\", \"\\u201cStairs word wrongly mentioned\\u201d: Fixed the typo in setting\", \"\\u201cExamples to clarify setting\\u201d: Added examples of covariate-shifted data augmentation under setting\", \"\\u201cConnection between linear regression setting and neural nets\\u201d: We added a new section, Section 5 in our revised paper that describes this clearly. As a quick summary, the theory suggests that as the size of the original dataset increases, the harmful effect of data augmentation decreases. We see this holds in practice.\", \"\\u201cEmphasize that AT perturbations are label-preserving, and thus falls in our setting\\u201d We make this clear in the setting in the revision\", \"\\u201c\\\\citep missing\\u201d: fixed\", \"\\u201cAssumption on \\\\Sigma\\u201d: We make no assumption on \\\\Sigma, the population covariance. Rather, we assume we have access to \\\\Sigma via unlabeled data\", \"\\u201cComparison to semisupervised methods\\u201d: We added a brief comparison to standard semisupervised learning in Discussion (Section 7) and a more detailed comparison in Appendix H. To summarize, consistency regularization does not use pseudo-labels. X-regularization requires pseudo-labels from a model that is trained without augmentations\", \"\\u201cX-reg vs. Robust Self Training (RST)\\u201d: Table 1 reports all relevant numbers. We would like to note that we realized that the methodological difference between what we proposed as X-reg and what was studied as RST in Carmon et al. 2019 is quite small. Hence, we revised the presentation to not claim X-reg as a new method, but rather view it as a generalization of RST. We view and present our contribution to be a theoretical justification for why RST improves standard accuracy. Empirically, we additionally evaluate RST on more forms of adversarial training and perturbations such as rotations.\"]}", "{\"title\": \"Author Response to R2\", \"comment\": \"We thank R2 for their feedback. The concerns raised stem from a confusion regarding the setting. R2 says that we assume that the augmented points are from the true distribution. We would like to clarify that we make *no* assumption on the augmented inputs. We only assume that the targets are from the true predictive/target distribution. As R2 points out, we typically augment with transformations. Since these transformations are label-preserving, our setting indeed covers these cases. We have revised the paper to make the setting more clear. We alco clarify that we differ from the standard semisupervised setting because we study both robustness (via augmentations) and standard accuracy. We request R2 to reconsider all our contributions with these clarifications. We have also expanded the discussions/contextualization of our work, experiments with more kinds of augmentation (rotations) and empirical comparisons to other relevant work. We request R2 to see the general comment above and also the revised paper.\"}", "{\"title\": \"General Comment\", \"comment\": [\"We thank the reviewers for their feedback. We have made substantial revisions to the paper that reflect the main concerns raised by the reviewers. The main changes are as follows:\", \"Modified the introduction: we now clearly state that our main motivation is to understand and prevent the well-documented drop in standard accuracy with adversarial augmentations via adversarial training.\", \"Elaborated on what we mean by \\u201ccovariate-shifted\\u201d data augmentations: We allow arbitrary augmented inputs and only constrain their targets to come from the true target distribution. In particular, this includes augmenting with \\u2018label-preserving\\u2019 transformations such as small translations, flips, crops, etc. or imperceptible l-p perturbations.\", \"Contributions and positioning: The main contribution is to theoretically understand why training with augmented data can lead to an increase in standard error. This theoretical study motivates X-regularization, a general self-training-like procedure that leverages unlabeled data to fix the increase in test error. X-reg for robustness has been studied in recent work (Carmon et al. 2019, Najafi et al. 2019, Uesato et al. 2019) as Robust Self-Training (RST). Our work thus provides theoretical justification for why RST should also improve standard error. We note that standard semisupervised learning studies standard accuracy, while we study standard accuracy alongside robustness.\", \"Additional baselines and perturbation types: We added augmentations with random and adversarial rotations/translations. We additionally compare with standard semi-supervised self-training, Interpolated Adversarial Training and Neural Architecture Search. X-reg improves robust and standard error over vanilla counterparts, and the gains are comparable to IAT and NAS (Table 1).\"]}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper describes situations whereby data augmentation (particularly drawn from a true distribution) can lead to increased generalization error even when the model being optimized is appropriately formulated. The authors describe a somewhat intuitive example of how this can happen, and proceed to support this example and generalize with mathematical rigor.\\n\\nThe machinery of the analysis and the positive suggestion of X-regularization strike me as technically sophisticated and effective.\\n\\nThis is largely a theoretical paper. The central example and description assume the availability of a true distribution to draw an arbitrary amount of examples from. While CIFAR-10 is used to demonstrate the effectiveness of X-regularization, it is not compared to any other augmenatation / semi-supervised / regularization approach. This makes contextualizing the contribution of this work more challenging. General data augmentation techniques frequently do not draw examples form a true distribution (here the test distribution or a held-out development set) but rather manipulate real examples through some transformation function (like noise addition, rotation, scaling, warping, etc.). This work does not address these conditions. Rather the augmentation investigated here is more like self-training, where information from unlabeled examples are used in the training (i.e. regularization) of a model. If i understand the technique correctly, assuming the eigenvector decomposition is relatively consistent under different samples of augmentation data, the regularization should be consistent whether regularized using the test set or some other unlabeled set. While I think this is theoretically true, it would be useful to see this reflected on the CIFAR data, where the test data is unused during training at all. \\n\\nThe contextualization of this work to related work -- particularly self-training and semi-supervised training -- is quite thin. While acknowledged, the technical rigor that is brought to the main result is not used to contrast with other techniques. Neither are there empirical comparisons to existing techniques in this space.\\n\\nOn balance, I find the theoretical investigation and explanation for \\\"how does this happen\\\" to be more compelling than the presentation of X-regularization.\\n\\nNote on \\\"Experience Assessment\\\": While I've done research on data augmentation and published in this space, the formulation and analytical tools employed in this paper I am less familiar with.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Overall I think this paper provides an interesting perspective on when and why augmented data can increase test error. Most of the theory deals with a simple convex problem, but some work is done to measure whether the insights carry over to more realistic settings. I think the main ways the paper could be improved are a) better explanation of connections between the theoretical assumptions and practice and b) better treatment of \\\"X-regularization\\\". For a) in particular, more discussion of which exiting data augmentation schemes fall under the theoretical framework would be useful. For b), X-regularization seems tacked on to the end of the paper, despite the potential strength of the contribution. It is not compared directly to existing semi-supervised or robust learning methods. As it currently stands I think the paper spends a bit too much time in the main text on the toy problem and theory, and could be bolstered by additional discussion of x-regularization. But, I think it's acceptable in its current form.\", \"specific_comments\": [\"In Figure 1 the green crosses use such wide lines that they look like dots at first glance. When the text referred to \\\"crosses\\\" I was confused. I'd suggest making them more obviously crosses.\", \"\\\"the distribution of Px of the augmented dataset is potentially different than Stairs on the original dataset\\\" Up until this point, your setting was not specific to the \\\"Stairs\\\" problem so I'm not sure why you're mentioning it here. This sentence would make more sense to me if you removed the words \\\"Stairs on the\\\".\", \"I think it would be useful if you provided some examples of data augmentation schemes (that are used in practice) that satisfy and do not satisfy the assumptions of your theoretical analysis.\", \"I would also suggest that you provide some additional connections or discussion of the relation between the behavior of your setting (section 2) to the behavior of models we actually use in practice (neural nets). How much of the behavior do we expect to carry over? Why?\", \"\\\"AT augments with imperceptible perturbations of training images with the corresponding correct target and thereby falls in our framework of covariate-shifted data augmentation.\\\" You should mention the assumptions implicit in this definition, namely that perturbing an image with a perturbation whose l_inf norm is less than some threshold will not change the target class. This is what connects it to your theoretical setting and so is an important detail.\", \"Ma et al. (2018); Belkin et al. (2018) need \\\\citep\", \"\\\"The assumption that we have \\u03a3\\\" I think you are missing an \\\"on\\\" before \\\"\\u03a3\\\".\", \"IIUC General X-regularization essentially enforces that the distance between a model trained on augmented + non-augmented data produces the same predictions as a model trained on only the non-augmented data when both models are fed additional unlabeled data. If this is the case I would suggest clarifying this a bit. This bears some similarity to consistency regularization/the \\\"Pi-Model\\\" (Laine & Aila 2017, Sajjadi et al. 2016), except in those cases consistency is enforced between a model fed with augmented and non-augmented versions of the same image. In your case IIUC you train two separate models and use unlabeled data to compare them; in the consistency regularization a single model is trained on labeled and unlabeled data together.\", \"\\\"Hence empirically, our X-regularization has lower standard error but higher robust error than RSL.\\\" I don't think you have empirical evidence that supports this claim.\"]}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper studies data augmentation in the regime where labels for the augmented datapoints are known. Special emphasis is put on the study of overparametrised linear models with minimum Euclidean norm of the regression weights (a.k.a. ridgeless regression). The results of this study are then used to motivate their \\u201cX-regularization\\u201d method, a semi-supervised learning algorithm which they test on the task of improving accuracy of adversarially trained models. The authors report improvement in accuracy of adversarially trained classifiers on CIFAR-10 when X-regularization is applied.\\n\\nWhile I think the paper could be potentially interesting to the ICLR community, I am currently leaning towards rejection for the following two reasons: (i) I am confused by significant chunks of the theoretical derivations; (ii) the paper is hard follow at several places, seems to be hastily written, and not well placed in the existing literature. Finally, I would like to state that since the paper is longer than 8 pages, I am holding it to a higher standard than an 8 page one as instructed by the guidelines.\", \"major_comments\": [\"I am very confused by the way you compute the (uncentred) sample covariance for the augmented dataset. On the penultimate line before eq.5, you state that Sigma_data = 1 / n X^T X if I interpret n as the total number of rows of X (which is IMHO correct). But then a couple of lines below you say that Sigma_aug = 1 / n (X_std^T X_std + alpha X_ext^T X_ext) which by the above logic should have been Sigma_aug = 1 / ((1 + alpha) n) (X_std^T X_std + X_ext^T X_ext) which is different from your definition. Inspecting appendix A.1, you also use Sigma_data = Sigma_std + alpha Sigma_ext whereas the above would suggest Sigma_data = 1 / ((1 + alpha)n) (n Sigma_data + alpha n Sigma_ext) = 1 / (1 + alpha) (Sigma_data + alpha Sigma_ext)?! Can you please clarify? Note that I did not go through the rest of the proof in the appendix.\", \"On a related note, I am also rather confused by your example in sect.3.1.1. In particular, if you observe e_1, e_2, and e_3, and you are in the scenario where you can get the true labels for the X_ext data points + there is no noise on y (sigma = 0), then you will get theta_aug = (X^T X)^dagger X^T X theta* = Id^(-1) Id theta* = theta* (where Id stands for the identity matrix), i.e., theta_aug = theta* and so there will be zero generalisation error since y = x^T theta* is the correct data generating process in this scenario. However, in the first display on p.4, you claim the bias (and thus generalisation error) of theta_aug is non-zero. The rest of the section similarly seems to contradict the theta_aug = theta*. Can you please clarify?\", \"Since the rest of sect.3 also uses the Sigma_aug = Sigma_data + alpha Sigma_ext instead of scaling this expression by Sigma_aug = 1 / (1 + alpha) (Sigma_data + alpha Sigma_ext), I will withhold my judgement until you could please clarify this confusion.\", \"Can you please point me to where I can find more detail about fig.1? In particular, how was the right plot produced (incl. how many repetitions of the experiment were executed)? Also, is it true that the green points in the middle plot were not chosen at random? If so, this should be clarified in the caption and the text.\", \"Since all of your non-toy experiments are focused on adversarial learning, it might benefit the paper to advertise this focus more clearly from page one. Furthermore, have you tested X-regularization with non-adversarial data augmentation (on CIFAR-10 or other standard benchmark) please?\", \"While you briefly mention the existence of alternatives to your algorithm, you don\\u2019t include any of these in your experiments section. I would suggest including some benchmark results (even if this only means restating numbers from other papers with citation).\"], \"minor_comments\": [\"In the abstract, you claim about X-regularization: \\u201cWe prove that our new estimator never increases test error.\\u201d while sect.4 you say that this guarantee only holds for linear models. I think this should be clarified in the abstract.\", \"p.1 in par.1, you say \\u201cbut in practice, we often work with models which can fit the augmented training data perfectly\\u201d (similar statements also appear throughout the paper, e.g., the last sentence of par.3 on p.1). I am confused by this statement because many augmentation schemes have full support in the input space, i.e., you will eventually have to be able to classify correctly any point in the input space. Are you assuming that the conditional distribution of y given x is almost surely deterministic, and using some universal approximation theorem for neural networks (even those tend to have some restrictions on the y = f(x) function!)? Please clarify.\", \"p.1 in par.3, \\u201c... may increase increase error \\u2026\\u201d -> \\u201cmay increase error\\u201d\", \"Second to last paragraph on p.2, \\u201cY = X theta + sigma N(0, I)\\u201d is mixing random variables and distributions. Please consider replacing \\u201cN(0, I)\\u201d with \\u201cepsilon, epsilon ~ N(0, I)\\u201c. Also, comparing with eq.1, did you mean \\u201csigma^2\\u201d in eq.1?\", \"Last paragraph on p.2, \\u201ccompare the performance two estimators\\u201d -> \\u201c compare the performance of two estimators\\u201d.\", \"Top of p.2 in eq.3: the definition of the augmented estimator is somewhat confusing. If the augmentation causes the total number of points (i.e., n + \\\\alpha n) to be higher than d (the number of parameters), then the set on the right hand side can be empty (and will be unless there is some perfect multicollinearity). Can you please clarify how is the augmented estimator defined in that case?\", \"Please reconsider use of contractions (e.g., \\u201cdoesn\\u2019t\\u201d -> \\u201cdoes not\\u201d in par.2 of sect.3) throughout the paper.\", \"In eq.5, in the expression for bias, I think you are missing star superscript for the 2nd theta.\", \"Under eq.5, \\u201c... we say that augmentation is safe if the predictive risk or bias does not increase, and hurtful if it does.\\u201d Can you please clarify whether the use of \\u201cor\\u201d is in the strictly logical sense, i.e., arbitrarily large increase in variance is fine as long as bias does not increase (in which case the second part of the statement and thus the full statement is still true)?\", \"Beginning of \\u201cLarge Sample Regime\\u201d paragraph of sect.3, you claim \\u201cCommon intuition from a statistical standpoint would suggest that more data is always good.\\u201d I am not sure I agree: there is a plethora of literature on data poisoning, robustness, differential privacy, adversarial examples, etc., which shows that adding data can significantly worsen performance of models. Even without reading these papers, it seems quite natural that adding **arbitrary** data into the dataset may not be the best idea (even in the large sample regime).\", \"Unresolved equation reference on p.11, 2nd line. Also, in the very next sentence, \\u201cyields zero and variance \\u2026\\u201d -> \\u201cyields zero bias and variance ...\\u201c?! Broken references also on p.14 (twice).\"]}" ] }

[ "Fanhua Shang", "Lin Kong", "Yuanyuan Liu", "Hua Huang", "Hongying Liu" ]

[ "non-smooth optimization", "SVRG", "proximal operator", "extragradient descent", "momentum acceleration" ]

{ "note_id": [ "uKtHXmb2jX", "B1xtsw5isr", "HklJuIqjsB", "ryxr33Yijr", "Sye1yhFojS", "r1eA4OKsoB", "B1eqF1Ykqr", "B1x_ANL1cS", "S1l6oaWstr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732936, 1573787553353, 1573787239354, 1573784748966, 1573784534838, 1573783605664, 1571946370114, 1571935440053, 1571655076652 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1808/Authors" ], [ "ICLR.cc/2020/Conference/Paper1808/Authors" ], [ "ICLR.cc/2020/Conference/Paper1808/Authors" ], [ "ICLR.cc/2020/Conference/Paper1808/Authors" ], [ "ICLR.cc/2020/Conference/Paper1808/Authors" ], [ "ICLR.cc/2020/Conference/Paper1808/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1808/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1808/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes a stochastic variance reduced extragradient algorithm. The reviewers had a number of concerns which I feel have been adequately addressed by the authors.\\n\\nThat being said, the field of optimizers is crowded and I could not be convinced that the proposed method would be used. In particular, (almost) hyperparameter-free methods are usually preferred (see Adam), which is not the case here.\\n\\nTo be honest, this work is borderline and could have gone either way but was rated lower than other borderline submissions.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Responses to \\\"Detailed comments\\\" and \\\"Additional question\\\" of \\\"Official Blind Review #3\\\"\", \"comment\": \"Response for \\u201cDetailed comments\\u201d:\\nThank you for checking the manuscript so carefully. According to your \\u201cDetailed comments\\u201d, we have revised and improved our paper. Besides, here are the explanations of some main questions:\\nResponse for \\u201c1.\\u201d: Our algorithm is applicable to both sparse and dense data sets. But in practical problems, the data is usually large-scale and sparse, and thus our main purpose of this algorithm is to solve the problem of large-scale sparse data sets. Besides, our sparse asynchronous variant can take advantage of the sparsity of data, and thus we have moved the asynchronous algorithm to the main body of the paper according to your suggestion.\\nResponse for \\u201c6\\u201d: Here \\u201cthe process\\u201d refers to the whole process of solving the problem to be solved. In the optimization problem, it refers to the process of solving the optimal value of the objective function. And \\u201cit\\u201d refers to the idea of extragradient.\\nResponse for \\u201c10\\u201d: Here \\u201cthe function\\u201d means the objective function to be optimized. We have also revised it in the revised manuscript.\\nResponse for \\u201c11\\u201d: We have given the citations where \\u201cAPG\\u201d and \\u201cAcc-Prox-SVRG\\u201d appear in the revised manuscript. Due to the limited space, we do not give the citations where they reappear.\\nResponse for \\u201c14\\u201d: The equations of Section 3.2 solve the minimum point of the function, and thus adding and reducing a constant term which is independent of the variable in the function will not change the final minimum point. In addition, the purpose of this equation is to explain that the proximal operator will introduce a gap of the optimal value. In particular, we can add the statement in the revised manuscript.\", \"additional_question\": \"[R4] update the extragradient step by sampling a new stochastic gradient while in this paper the same sample is used twice. How you compare these two approaches in terms of their performance and convergence?\", \"response\": \"We know that the algorithm without extragradient, such as Prox-SVRG, will sample one stochastic gradient to update in each inner iteration. Therefore, sampling a new stochastic sample to update the extragradient is equivalent to using the algorithm without extragradient to update twice. In this way, we cannot take advantage of the advantage of extragradient. However, we update twice on the same sample point in our algorithm. It is different from the algorithms without extragradient, and can get better results, which can be seen from the experimental comparison between VR-SExtraGD and Prox-SVRG.\\n\\n[R4] R. Rockafella, Convex Analysis, 1970.\"}", "{\"title\": \"Responses to \\\"Official Blind Review #3\\\"\", \"comment\": \"Thank you for your valuable comments. Through your comments, we get a lot of inspiration and directions to improve the manuscript better.\", \"question_1\": \"This paper claims that extragradient reduces the gap of the optimal value, which is confusing.\", \"response\": \"We have carefully revised the manuscript according to your suggestions. We have moved all the lemmas to Appendix, fixed typos and incorrect grammar, and added the definitions of some undefined symbols.\\n\\n[R1] T. Nguyen et al., Extragradient method in optimization: Convergence and complexity, 2017.\\n[R2] Zhou et al., A simple stochastic variance reduced algorithm with fast convergence rates, 2018.\\n[R3] Allen-Zhu Zeyuan. Katyusha: the first direct acceleration of stochastic gradient methods. 2017.\", \"question_2\": \"Besides, how this claim is reflected in the convergence result is not discussed. Section 1.2: the claim extragradient examines the geometry and curvature of the problem is confusing.\", \"question_3\": \"How does extragradient affect the complexity and choice of hyperparameters such as $K$, $\\\\eta_1$, $\\\\eta_2$, and $\\\\beta$?\", \"question_4\": \"A more careful experimental design is required to better demonstrate the performance:\\n1.\\tThe choice of inner iterations.\\n2.\\tThe comparisons on number of iterations.\\n3.\\tThe comparison with MiG [R2].\\n4.\\tAt least solve two different optimization programs (e.g. logistic regression, neural network).\", \"question_5\": \"The presentation and structure of this paper need to be improved.\"}", "{\"title\": \"Responses to \\\"Official Blind Review #1\", \"comment\": \"Question 1: It would be good and well understood to explain intuitively the steps of the proposed algorithm.\", \"response\": \"In our experiments, we used the once extragradient update every 25 inner-iterations, which have also been explained in Section 5 in the manuscript. In practical problems, the frequency of using the extragradient update step can be adjusted according to specific problems and specific requirements. In this way, $K$ can be determined. In addition, we have added more experimental results for choosing the parameter $K$, in the revised manuscript, and also provided a discussion in choosing the parameter.\", \"question_2\": \"How should \\u201c$K$\\u201d be chosen in practice?\"}", "{\"title\": \"Revision uploaded\", \"comment\": \"We thank all the reviewers for their invaluable comments. We answered all the reviewers' concerns and questions, respectively, and uploaded a revised version of our paper.\"}", "{\"title\": \"Responses to \\\"Official Blind Review #2\\\"\", \"comment\": \"Question 1: How does it compare to SVRG++ and Proximal Proximal Gradient?\", \"response\": \"In this paper, we first introduced the idea of extragradient descent into Prox-SVRG and propose a new algorithm, called VR-SExtraGD. In particular, we also proposed a new momentum accelerated VR-SExtraGD algorithm, called AVR-SExtraGD. Thus, we not only combine momentum acceleration with the proposed VR-SExtraGD algorithm, but also introduce an innovative idea into the algorithm. Another main contribution of this paper is the convergence results of the proposed algorithms including VR-SExtraGD and AVR-SExtraGD. Due to the introduction of extragtadient descent, our convergence analysis for the proposed algorithms needs more improvement and innovation, which is also our main novelty and overcoming technical difficulty.\", \"question_2\": \"Combining momentum with an existing algorithm is not extremely novel.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": [\"This is an optimization algorithm paper, using the idea of \\\"extragradient\\\" and proposing to combine acceleration with proximal gradient descent-type algorithms (Prox-SVRG). Their proposed algorithm, i.e., accelerated variance reduced stochastic extra gradient descent, combines the advantages of Prox-SVRG and momentum acceleration techniques. The authors prove the convergence rate and oracle complexity of their algorithm for strongly convex and non-strongly convex problems. Their experiments on face recognition show improvement on top of Prox-SVRG as well Katyusha. They also propose an asynchronous variant of their algorithm and show that it outperforms other asynchronous baselines.\", \"technically sound, seems like a nice addition to the variance reduced gradient-type methods. Combines the nice properties of proximal methods with variance reduced gradient-descent.\", \"Nice summary of recent progress in this research area.\", \"How does it compare to SVRG++? How about the Proximal Proximal Gradient?\", \"algorithm suitable for non-smooth optimization problems\", \"their experimental results look convincing.\", \"Having said that, it seems to me that combining momentum with an existing algorithm is not extremely novel -- I would defer to reviewers who are experts in the optimization area to fully assess the novelty and technical difficulty of the proposed solution.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposed a new stochastic algorithm: AVR-ExtraGD. AVR-ExtraGD combines the extended extragradient method proposed in [3] and the accelerated SVRG method in [1][2]. In their experiments, AVR-ExtraGD outperforms [2] in running time for sparse linear regression.\\n\\nThis paper presents their convergence analysis using results in [1][2]. They showed that the proposed algorithm can achieve O(sqrt{kappa n} log (1 / \\\\epsilon)) complexity for strongly convex problem and O(1/sqrt{epsilon}) for convex problem, which are the best results for both cases.\\n\\nThe idea of an accelerated version of variance reduced stochastic extragradient method is novel. However, there are some issues that the authors should address in order for the paper to match the quality of ICLR.\\n\\nEach step of extragradient approximates the proximal operator x_{k+1} = argmin_x P(x) + 1/(2 eta_k)\\\\|x \\u2013 x_k\\\\|_2^2, therefore we would expect a faster and more stable convergence from this method. This paper claims that extragradient reduces the gap between the obtained optimal value and the real optimal value, which is confusing. The update of extragradient is actually biased towards x_k. The claim is then discussed in section 3.1 and 3.2 but is not clearly explained. Besides, how this claim is reflected in the convergence result is not discussed. I encourage the authors to clearly elaborate this claim and make relevant remarks after the main theorems.\\n\\nTo better understand the convergence result, it is important to know how extragradient affects the complexity and choice of hyperparameters such as K, eta_1, eta_2, and beta. Such discussion is not in this paper. I suggest the authors to make these aspects clear.\\n\\nThe experiments compare the proposed algorithm with other algorithms by their running time for lasso and elastic-net. The comparisons show the efficiency of the proposed algorithm. However, a more careful experimental design is required to better demonstrate the performance:\\n1.\\tFor the choice of inner iterations, choosing m=2n for Katyusha actually requires calculating 5n stochastic gradient because each iteration of Katyusha does gradient updates twice.\\n2.\\tThis paper only presents comparisons of running time. I encourage the authors to also plots comparisons on number of iterations, which will help revealing where the speed up of AVR-ExtraG comes from.\\n3.\\tIt is also preferable that the author compare with MiG [1], since the proposed algorithm is an extragradient version of [1].\\n4.\\tPlease at least solve two different optimization programs (e.g. logistic regression, neural network) so any conclusions are not specific to the oddities of a particular program.\\n\\nThe presentation and structure of this paper need to be improved. Here are some suggestions:\\n1.\\t In Section 1, only provide a high-level literature review and then motivate the work. A comprehensive review can come after the introduction.\\n2.\\tIn Section 4, put all the lemmas into the appendix while giving more intuitions and remarks. \\n3.\\tIssues including notions without pre-definition or reference, typos, and incorrect gramma need to be fixed.\", \"detailed_comments\": \"1.\\tFrom the title, the main application of this work is sparse learning problem. However, how the proposed algorithm benefits sparsity is not discussed. Besides, I suggest the authors to move the asynchronous algorithm in the appendix to the main paper.\\n2.\\tThe paragraph before section 1.1: lasso and elastic-net are used without citation.\\n3.\\tSection 1.1: PGD and SGD are used without citation.\\n4.\\tBeginning of page 2: \\u201cAnd\\u201d should be \\u201cBesides\\u201d\\n5.\\t\\u201cBesides, for accelerating the algorithm and \\u2026\\u201d: \\u201cfor accelerating\\u201d should be \\u201cto accelerate\\u201d\\n6.\\tSection 1.2: \\u201cNguyen et al. (2017) proposed the idea of extragradient which can be seen as a guide during the process, and introduced it into the optimization problems.\\u201d What does \\u201cthe process\\u201d and \\u201cit\\u201d refers to is unclear.\\n7.\\tSection 1.2: the claim extragradient examines the geometry and curvature of the problem is confusing. The geometry of the problem is inspected through a line search step in [3]. However, line search is not discussed in this paper.\\n8.\\tSection 1.2: \\u201creduce the gap between the optimal value we get and the real optimal value\\u201d, these two kinds optimal values are important notions of this paper but they are not defined.\\n9.\\tIn Assumption 2, you can refer to Part 2, Section 7 of [5] for the definition of semi-continuity.\\n10.\\t\\u201cdw is the gradient of the function at w\\u201d, what does \\u201cthe function\\u201d refers to?\\n11.\\t\\u201cAPG and Acc-Prox-SVRG\\u201d needs citation.\\n12.\\t\\u201cwas proposed to simply the structure of Katyusha\\u201d, \\u201csimply\\u201d should be \\u201csimplify\\u201d\\n13.\\tSection 3.1: \\u201cupdated with the update rules of MiG\\u201d: \\u201cwith\\u201d should be \\u201cby\\u201d\\n14.\\tIn the equations of Section 3.2, the equivalent of gradient norm square and function f is incorrect, and the purpose of this equation is unclear.\\n15.\\tSection 4.1 Theorem 1: The inequality in theorem 1 is not intuitively related to the convergence rate. I suggest the author to simplify the inequality (For example, Theorem 2.1 in [2]).\\n16.\\tThe references are not in a uniform format. Conference/Journal names are missing for some references.\\n17.\\tOne useful reference for this paper is [4], it discussed extragradient for online convex learning.\", \"additional_question\": \"[5] update the extragradient step by sampling a new stochastic gradient while in this paper the same sample is used twice. How you compare these two approaches in terms of their performance and convergence?\\n\\n[1] A simple stochastic variance reduced algorithm with fast convergence rates, Zhou et al., 2018.\\n[2] Katyusha: the first direct acceleration of stochastic gradient methods, Z. Allen-Zhu, 2017\\n[3] Extragradient method in optimization: Convergence and complexity, T. Nguyen et al., 2017\\n[4] Online Optimization with Gradual Variations, Chiang et al., 2012\\n[5] Convex Analysis, R. Rockafella, 1970\\n[6] Reducing Noise in GAN Training with Variance Reduced Extragradient, Chavdarova et al. 2019\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I made a quick assessment of this paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposes an optimization method for solving unconstrained convex optimization problems where the objective function consists of a sum of several smooth components f_i and a (not necessarily smooth) convex function R. The proposed method AVR-SExtraGD is a stochastic descent method building on the previous algorithms Prox-SVRG (Lin 2014) and Katyusha (Zeyuan 2017). The previous Prox-SVRG method using a proximal operator is explained to converge fast but leads to inaccurate final solutions, while the Katyusha method is an algorithm based on momentum acceleration. The current paper builds on these two approaches and applies the momentum acceleration technique in a stochastic extragradient descent framework to achieve fast convergence.\\n\\nI am not working in the field of optimization, therefore, unfortunately I am not in a position to give detailed technical comments for the authors. However, as far as I could follow the paper, it seemed sound and well-written to me in general. I hope the following minor comments may be useful for improving the paper:\\n\\n- The paper gives detailed explanations about previous work. However, the proposed AVR-SExtraGD algorithm is only presented in the form of a pseudocode in Algorithm 1 and it is not explained in much detail. It would be good to explain and discuss intuitively the steps of the proposed algorithm in the main body of the paper as well, so that it is well understood.\\n\\n- Algorithm 1 has a set K as input, according to which the solution is updated. How should this set be chosen in practice?\"}" ] }

[ "Vinenzo Crescimanna", "Bruce Graham" ]

[ "autoencoder", "information theory", "infomax", "vae" ]

{ "note_id": [ "PT5Y5E4AQi", "S1xGmqNnjB", "HygxY7V3oH", "r1gwWRm3iB", "rkgKQMmd9r", "SkxE4Bxr9r", "r1lWnRs7YH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732908, 1573829145980, 1573827448093, 1573826046547, 1572512289436, 1572304172404, 1571172009138 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1806/Authors" ], [ "ICLR.cc/2020/Conference/Paper1806/Authors" ], [ "ICLR.cc/2020/Conference/Paper1806/Authors" ], [ "ICLR.cc/2020/Conference/Paper1806/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1806/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1806/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper describes a new generative model based on the information theoretic principles for better representation learning. The approach is theoretically related to the InfoVAE and beta-VAE work, and is contrasted to vanilla VAEs. The reviewers have expressed strong concerns about the novelty of this work. Some of the very closely related baselines (e.g. Zhao et al., Chen et al., Alemi et a) are not compared against, and the contributions of this work over the baselines are not clearly discussed. Furthermore, the experimental section could be made stronger with more quantitative metrics. For these reasons I recommend rejection.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We thank the reviewer for the detailed review and for the thoughtful comments.\\n\\n> Figs 2 and 3 are only few examples manually chosen from rather simple tasks. \\nIn order to evaluate the quality of the generated samples we used the Frechet Inception Distance for the CIFAR10 and CelebA dataset (see the appendix) and the Negative Log Likelihood for the Omniglot, in all these cases the quality of the samples of VIMAE models is higher. In order to evaluate the reconstruction we consider the L_2 norm, and as suggested by (Alemi et al., 2017) each model is denoted by its reconstruction loss and rate term, helping us to observe that the VIMAE models are learning the most informative generator.\\n\\n> I think the paper unnecessarily complicates the presentation of the idea.\\nFollowing your suggestion in the revised version we moved the unnecessary details to the story in other sections, in order to do distract from the main story line\\n\\nWe compare the VIMAE and the model proposed by (Alemi et al., 2017), observing that both are controlling the encoding information, and that the two coincide in the case the bound of the rate term is equal to the network capacity. We observed also that the main difference between the two approaches is in the ratio decoding information/capacity, in VIMAE is one for any p(z), in the Alemi model is in general less than one, that means, according to theory of separability described in (Mathieu et al., 2019) and confirmed by the experiments, that the learned representation is not in general robust and clustered. \\nWe observe moreover that the model as described in (Alemi et al, 2017) is not in general feasible to compute and it is necessary to consider a $\\\\beta$-VAE with $\\\\beta<1$.\\n\\nIn the related work section we underlined that in order to avoid the uninformative issue is possible to consider a family of models learning a flexible prior $p(z)$. But we do not consider them, because are quite expensive to compute and they are not suitable for tasks like the lifelong where different types of data are considered\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"We thank the reviewer for the detailed review and for the thoughtful comments.\\n\\nIn the revised version we reworded the critical parts and we underlined that the VIMAE, in the case the prior $p(z)$ is normally distributed coincides with the InfoVAE (Zhao et al., 2017) and WAE (Tolstikhin et al., 2017). \\n\\nIn the conclusion we suggest a future application to lifelong learning; indeed,the ability to learn the relevant features of the data in a robust way is suitable and particularly important in a lifelong task, where it is necessary to learn only few specific features for each task (dataset) in order to avoid the catastrophic forgetting issue.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We thank the reviewer for the detailed review and for the thoughtful comments.\\n\\nIn the revised version we underline that the main contribution of the paper is not the definition of a model, that was previously defined in (Tolstikhin et al., 2017) and (Zhao et al., 2017), but the information analysis of such model. Thanks to this description we was able to describe the relationship between the decoding information and the network capacity, and then the introduction of the Capacity Constrained InfoMax.\\n\\nIn this new version we cite both (Alemi, et al., 2017) and (Grover and Ermon, 2019). \\nWe observed that the model introduced in (Grover and Ermon, 2019) belongs in the (unbounded) InfoMax family, but we do not compare it with the other models since the sampling is quite expensive, and then out of our interest. \\nInstead we compare both theoretically and computationally the VIMAE model with the one proposed in (Alemi, et al., 2017), observing that if in some special cases the two approaches are identical (minimal bound = Network capacity), computationally the VIM objective is easier to compute, and it is learning more robust and clustered representation.\\n\\nSince the word disentangled was not clear and incorrect in this context, in the revised version we try to describe what, in this paper, is a good representation: a robust and clustered in order to be used as input for simple classifier. This definition of good is similar to what in (Mathieu, et al., 2019) is defined decomposed.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"I went over this work multiple times and had a really hard time judging the novelty of this work. The paper seems to be a summary of existing work reinterpreting variational autoencoding objectives from an information theoretic standpoint. In particular, the paper seems to follow the same analysis as in Wasserstein Autoencoders (Tolstikhin et al., 2017) and InfoVAE (Zhao et al., 2017). It is unfair to say that the objectives were \\\"derived independently\\\" since these works are from a couple of years ago.\", \"the_paper_also_lacks_discussion_on_two_crucial_works_in_this_space\": \"1. https://arxiv.org/abs/1711.00464 shows how to trade off rate-distortion in VAEs.\\n2. https://arxiv.org/abs/1812.10539 shows how to learn informative representations by eliminating the KL divergence term and at the same time specifying an implicit generative model (Theorem 1).\", \"re\": \"disentanglement. Unsupervised disentanglement has been shown to be theoretically impossible and several key challenges have been highlight w.r.t. prior work. Again, the relevant paper: https://arxiv.org/abs/1811.12359 has not even been cited. More importantly, the claims around \\\"more disentangled representation\\\" are imprecise in light of this work.\\n\\nA proper discussion on the contributions of this work as well as discussion on the above related works would be desirable on the author's end.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Overview: This paper describes the Variational InfoMax AutoEncoder (VIMAE), which is based on the learning principle of the Capacity Constrained InfoMax. The core idea behind VIMAE is that the encoding information is not bounded while network capacity is. The issue that VIMAE can handle, and where VAE fails, is that representations are not informative of input data, due to the information bottleneck idea that VAE is built upon. The authors describe InfoVAE and \\u03b2-VAE, which both attempt to solve this problem. The theory behind VIMAE is then described and tested against VAE and \\u03b2-VAE, in their abilities to evaluate the entropy of Z, in reconstruction and generative performance, and in robustness to noise and generalization.\", \"contributions\": \"The authors clearly state the contributions of the paper themselves, but in a summary: a derivation of a variational lower bound for the max mutual info of a generative model, definitions and bounds estimation for a VAE, associations for generative quality and disentanglement representation to mutual information and network capacity, and finally proposing the Capacity-Constrained InfoMax.\", \"comments\": \"\", \"page_1\": \"\\u201cOur derivation allows us to define\\u2026\\u201d -> this sentence is a bit long and took me a few reads to understand, reword this please.\\n\\u201cDerivation of a variational lower bound for ...a geneaive* model belonging..\\u201d -> \\u201cgenerative\\u201d\", \"page_2\": \"\\u201cWe conclude the paper with experimental results and conclusions.\\u201d You already said conclude twice in this sentence, I feel like this could be better worded to avoid that.\", \"page_5\": \"\\u201cIn order to test the assumption that it is sufficient..\\u201d -> This wording is also hard to wrap my head around. Too many commas.\", \"page_6\": \"\\u201cIn figure 1 are plotted the 2d\\u2026\\u201d -> This sentence could be reworded\\n\\nMy main complaint with the paper is that there are quite a few places that could be reworded better in addition to these. Please fix this.\\n\\nYou mention both InfoVAE and \\u03b2-VAE yet only test your models against a \\u03b2-VAE model. What was your reasoning for this?\\n\\nThe semi-supervised learning experiment is interesting (the one based on Zhao et al.) VMAE models are able to classify better regardless of noise or sampling procedure, especially the smaller capacity model VMAE-1. I\\u2019d like to hear a discussion in the paper about future work and how the authors believe this could be applied elsewhere.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper develops an information-theoretic training scheme for Variational Auto-Encoders (VAEs). This scheme is tailored for addressing the well-known disentanglement problem of VAEs where an over-capacity encoder sometimes manages to both maximize data fit and shrink the KL divergence between the approximate posterior and prior to zero. Consequently, the latent representations of the observations become independent, making them unusable for any downstream task.\\n\\nThe method developed in Section 3 and proposed explicitly in Eqs 7 to 11 is novel per se, though not groundbreaking.\\n\\nFigs 2 and 3 are only few examples manually chosen from rather simple tasks. In the absence of a quantitative evaluation metric, they are not informative. In the outcomes of the same runs, there might exist counterexamples where the vanilla VAE generate perceptively more appealing reconstructions than VIMAE.\\n\\nAs a minor comment, I think the paper unnecessarily complicates the presentation of the idea. The development of notions such as f-Divergence, links to MMD, the constrained optimization setting in Eq 5 etc., do not serve to the main story line, but only distracts a mind. I would prefer a brief intro on VAEs and directly jumping into the proposed method. Last but not least, the statement of technical novelty in Section 3 is extremely encrypted. What exactly is \\\"absolutely\\\" novel there and what is prior art? Can the authors give a to-the-point answer to this question during the rebuttal?\", \"the_paper_has_two_fundamental_weaknesses\": \"i) The paper misses a key reference, which addresses the same representation disentanglement problem using the same information-theoretic approach, only with some minor technical divergences:\\n\\nAlemi et al., Fixing a Broken ELBO, ICML, 2018. \\n\\nThis paper is a must-cite, plus a key baseline. This submission cannot be treated as a contribution without showing an improvement on top of this extremely closely related work. The problem is the same, the solution is almost the same, and the theoretical implications of the solution are also the same.\\n\\nii) The paper builds the entire story line around the statement that the representation disentanglement problem is caused by the properties of the ELBO formula and attempts to fix it by developing a new inference technique. This is attitude largely overlooks simpler explanations. For instance, the vanilla VAE assumes a mean field q(z|x) across the z dimensions. This makes q(z|x) fall largely apart from p(x) which definitely does not factorize that way. Earlier work has shown substantial improvements on the quality of q(z|x) when structured variational inference techniques are used, such as normalizing flows. Furthermore, there has also emerged techniques that can very tightly approximate p(x) in closed-form using the change of variables formula without performing any variational inference at all. No need to say, a tight approximation on p(x) means the same for p(z|x) simply using the Bayes rule. This leaves no room for further improvements by information-theoretic inference alternatives. For instance see:\\n\\nDinh et al., Density Estimation Using Real NVP, ICLR, 2017\\n\\nThe presence of such two strong alternative remedies to the problem addressed by this work makes its fundamentals shaky. The only way to get over this situation is to provide a thorough comparison against these methods, which is obviously missing.\\n\\n---\", \"post_rebuttal\": \"Thanks to authors for all the effort they put on the updated version. I appreciate that the paper now provides a quantitative comparison. However, my point (ii) remains fully unaddressed and I fully share the theoretical concerns raised by Reviewer 1. All in all, I keep my position on the below the threshold side with a weak reject.\"}" ] }

[ "Vitchyr H. Pong", "Murtaza Dalal", "Steven Lin", "Ashvin Nair", "Shikhar Bahl", "Sergey Levine" ]

[ "deep reinforcement learning", "goal space", "goal conditioned reinforcement learning", "self-supervised reinforcement learning", "goal sampling", "reinforcement learning" ]

{ "note_id": [ "wteBshrWy", "HJlTYAfnoS", "rJlYx0G3jB", "Bkl-PnTtir", "SJgBwsTYsr", "S1lF1vROjS", "BygXShuZsS", "rkgaGhdWjr", "rJgUssdWjr", "rJgogsu-jB", "BylJ3YZntB", "SJxzk8sitS", "HyeleOROFS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732880, 1573822085255, 1573821937487, 1573669977370, 1573669724830, 1573607136960, 1573125179064, 1573125140964, 1573125022233, 1573124851375, 1571719591300, 1571694042381, 1571510247802 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1805/Authors" ], [ "ICLR.cc/2020/Conference/Paper1805/Authors" ], [ "ICLR.cc/2020/Conference/Paper1805/Authors" ], [ "ICLR.cc/2020/Conference/Paper1805/Authors" ], [ "ICLR.cc/2020/Conference/Paper1805/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1805/Authors" ], [ "ICLR.cc/2020/Conference/Paper1805/Authors" ], [ "ICLR.cc/2020/Conference/Paper1805/Authors" ], [ "ICLR.cc/2020/Conference/Paper1805/Authors" ], [ "ICLR.cc/2020/Conference/Paper1805/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1805/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1805/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper tackles the problem of exploration in RL. In order to maximize coverage of the state space, the authors introduce an approach where the agent attempts to reach some self-set goals. The empirically show that agents using this method uniformly visit all valid states under certain conditions. They also show that these agents are able to learn behaviours without providing a manually-defined reward function.\\n\\nThe drawback of this work is the combined lack of theoretical justification and limited (marginal) algorithmic novelty given other existing goal-directed techniques. Although they highlight the performance of the proposed approach, the current experiments do not convey a good enough understanding of why this approach works where other existing goal-directed techniques do not, which would be expected from a purely empirical paper. This dampers the contribution, hence I recommend to reject this paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Additional Experiments Added\", \"comment\": \"As suggested, we have added experiments that study the importance of the reinforcement learning algorithm used with Skew-Fit. Specifically, we replaced soft actor critic (SAC) with twin delayed deep deterministic policy gradient (TD3) and reran the simulated, vision-based experiments with this new combination. The results show that Skew-Fit performs well with both TD3 and SAC, with both versions achieving approximately the same final error. These results suggest that the benefits of Skew-Fit are not specific to SAC, and can instead be combined with other reinforcement learning algorithms.\\n\\nWe have also added an experiment that explicit tests for exploration, by using a maze environment with long corridors and measuring the state coverage. These experiments are shown in Section B.1, and show that Skew-Fit significantly accelerates exploration. We have also added a simulated robot quadruped experiment that requires a robot to explore a narrow box-shaped corridor. In this experiment, we again see that Skew-Fit results in faster exploration than prior methods (see Section B.1).\"}", "{\"title\": \"Additional Experiments Added\", \"comment\": \"As suggested, we have also added additional exploration experiments on a simple maze setup in Section B.1. The maze and action spaces are designed so that random actions are unlikely to result in fast exploration and instead require goal-directed exploration. In these new experiments, we see that Skew-Fit significantly accelerates exploration.\\n\\nWe have also added a simulated robot quadruped experiment that requires a robot to explore a narrow box-shaped corridor. In this experiment, we against see that Skew-Fit results in faster exploration than prior methods.\\n\\nWe note that prior work in the field have similarly tested their algorithms on 3 simulated domains [1,2]. We believe that with our real-world robot experiments, as well as the additional experiments described above, our evaluation provides a similar level of rigor.\\n\\n[1] Andrychowicz, Marcin, et al. \\\"Hindsight experience replay.\\\" Neural Information Processing Systems. 2017.\\n[2] Nair, Ashvin, et al. \\u201cVisual Reinforcement Learning with Imagined Goals. Neural Information Processing Systems. 2019.\"}", "{\"title\": \"Re: Clarifications, further comments... from AnonReviewer1 (2/2)\", \"comment\": \"(continued from previous comment)\\n\\n> What do you exactly mean by a policy to be reusable?\\nBy reusable, we mean that a policy can reach user-defined goals after performing exploration. In other words, it is a goal-conditioned policy rather than a policy that only performs exploration. For example, imagine you left a robot in a large office. If it performs exploration well, then the robot will autonomously visit every room in the building, regardless of whether it is goal-oriented or not. The difference between a goal-conditioned and non-goal-conditioned exploration policy is based on what happens next: After exploration, you can tell a goal-conditioned policy, \\u201cPlease go to room A.\\u201d and it will know how to go to room A since it has already practiced reaching every possible goal. In other words, we can *reuse* the exploration policy to achieve these user-specified goals. However, if the exploration policy is not goal-conditioned but instead trained with an exploration-reward bonuses, then there is no way to control the policy. It may have visited every room during exploration, but at any given time, it only knows how to reach one location. In particular, it will always go to the last state that was deemed novel by the exploration-reward bonus.\\n\\n\\n> What do you exactly mean by a distribution over terminal states? Why not the normalized discounted weighting of states?\\nBy terminal state, we mean the last state of each episode. So, the distribution over terminal states is the distribution of states where the policy will be located at the end of each episode. For example, will the policy always end at position X? Or will its final position have a Gaussian distribution? Uniform distribution?\\n\\nFor our analysis, we found it more natural to study the distribution over terminal states, since in the goal-conditioned setting, we would like our goal-conditioned policy to *end* at a goal state when the episode is complete. However, we believe that our analysis could be applied to normalized discounted weighting of states by allowing an agent to constantly set new goals rather than waiting until the beginning of a new episode to set the next goal.\\n\\n\\n> Does \\u201cno user-specified rewards\\u201d mean a sparse reward?\\nNo, and we understand that this may have been a source of confusion. By \\u201cno user-specified rewards\\u201d we mean that the user does not need to manually engineer a reward for each task. Instead, as described in Section 4, we use the same reward as the one used in RIG, which is an approximation of the log probability of the goal given the current state. Overall, this means that the same code and generic reward is used for the real-world door task and all of the simulated tasks. For example, there is no reward that specifically tells the robot to open the door, nor is there a reward that specifically tells the robot to pick up or move the objects. Instead, the same Skew-Fit objective encourages the robot to learn to manipulate these objects into as many configurations as possible (by setting diverse goals and then reaching those goals) regardless of the environment.\"}", "{\"title\": \"Re: Clarifications, further comments... from AnonReviewer1 (1/2)\", \"comment\": \"Thank you for the additional response. We answer your remaining questions, and are happy to continue discussing if there are still points of confusion. In particular, we explain the large empirical and conceptual differences between Skew-Fit and prior methods, both of which we believe would be of interest to the ICLR community.\\n\\n\\n> How does Skew-Fit relate to hindsight experience replay (HER) and similar approaches? How is Skew-Fit different?\\nWe note that there are significant empirical differences between Skew-Fit and HER. In Figure 5, we see that HER does not perform well without access to an oracle uniform goal distribution (more on this below). Specifically, HER has a final distance that is 400%, 200%, and 150% higher than the final distance when using of Skew-Fit, on the door, pickup, and pushing tasks, respectively. Similarly, Skew-Fit outperforms other prior methods across all tasks.\\n\\nA major conceptual difference between Skew-Fit and many prior methods, such as HER, is that we *learn* the goal distribution, whereas many prior methods *assume* that a uniform goal distribution is provided. When using images as observations, this amounts to assuming that the agent knows the distribution of natural images -- an unreasonable assumption in most cases. For example, in the object-pushing task of the hindsight experience replay (HER) paper, an XY goal-position for the object is sampled uniformly from within the workspace of the robot and given to the robot for exploration. While this procedure is simple to implement in a simulated domain and when the goal corresponds to an XY-Cartesian position in the plane, it is much more challenging when goals correspond to images, as in the environments that we tested. Randomly sampling an image will result in an image of static noise. So, Skew-Fit learns a goal distribution that corresponds to the uniform distribution over the set of valid states, instead of assuming that we are given access to this goal distribution.\\n\\nBecause most prior goal-conditioned RL methods assume access to an oracle uniform goal distribution, most of the methods have only been applied to simulated domains, where defining such a goal distribution by hand is easy. As our experiments show, without access to a uniform distribution, prior methods such as HER perform poorly. Moreover, many of these methods (AutoGoal GAN, Rank-Based Prioritization, HER) assume access to ground-truth state information for computing the reward, which is readily available only in simulation. Enabling goal-conditioned RL to be applied to domains where the goal-space and reward function are unknown a priori, such as image-based domains, is important if we want to use these methods outside of simulation and for real-world applications. We believe that Skew-Fit is a useful step, both empirically and conceptually, towards this objective. Note that we evaluate Skew-Fit on a real-world image-based robotic manipulation task to demonstrate this (see Figure 1 and 7).\\n\\nAs far as we know, the only other goal-conditioned methods that have been developed for goal images are RIG and DISCERN, but neither of these methods address the important question of how goals should be sampled for exploration. Applying Skew-Fit results in considerable performance gains over these prior methods: Figure 5 shows that RIG and DISCERN have final distances that are about 100% higher than that of Skew-Fit. Lastly, the only method that has been applied to real-world robot domains from images is RIG, and we found that Skew-Fit outperformed it not only in simulation, but also on the real-world door task.\\n\\n\\n> How are these goals determined in the first place?\\nWe first describe how we generate goals chosen for evaluation. For the simulated tasks, we use an \\u201coracle\\u201d sampling procedure that exploits the fact that the task is simulated. Note that this procedure is only used for evaluation and never used by the algorithms. To generate an evaluation goal, we sample a ground-truth state uniformly distribution from the entire state space, set the environment to this state, capture an image corresponding to this state, and use the resulting image as the goal. We then return the environment state back to its original state and instruct the policy to reach the captured goal image. This procedure based on ground-truth state information is only used for evaluation and in simulation. For the real-world door task, we took images of the door at 5 different angles, evenly spaced from 0 to 45 degrees. Like before, these goals are only used for evaluation and not be the algorithms.\\n\\nWe now describe how goals for exploration are generated. At the very beginning of exploration, the agent takes random actions in the environment to collect a set of states. From thereon, the exploration goals are generated by the agent by sampling from the learned goal distribution, which is learned with Skew-Fit.\\n\\n(continued in next comment)\"}", "{\"title\": \"Clarifications, further comments and doubts on technical novelty. It seems like the contribution is somewhat marginal. Not fully convinced about the usefulness of this approach.\", \"comment\": \"Thank you for your detailed response. It is certainly very helpful. However, I have few other comments and questions :\\n\\n- It seems that this approach is useful in the case where there are no user-specified rewards (my understanding is you mean a sparse reward setting). However, as you mention, that p(s) in your case represents the distribution over terminal states for a finite horizing setting - I find it difficult to understand what this distribution actually means in an episodic setting. \\n\\nWhat do you exactly mean by a distribution over terminal states? My understanding is this would be the normalized occupancy measure, where the occupancy is only over the terminal states? Given this, why should we even consider this type of p(s) and why not the normalized discounted weighting of states - and maximize this distribution accordingly? I do agree that simply maximizing this might not be sufficient, which is why you include the H(s|g) term - but that raises further concerns in my opinion, as follows : \\n\\nWith the current formulation, we would now require distributing goals across the state space - this is similar to Hindsight Experience Replay and related approaches? How are these goals determined in the first place - is it more like a random sampling of states that you consider as goals - and then encouraging the agent to reach these goals?\\n\\nIsn't this method then a very slight modification of HER and other related papers? To me, it seems like that makes the contribution very marginal - since instead of HER, now we have a clever way of maximizing a MI term, but it is indeed doing something very similar to HER from a technical contribution?\", \"further_comment\": \"I don't necessarily agree, or fully understand this from a technical perspective - \\\"We instead would like for this policy to be reusable, by, e.g., being able to control what state it reaches\\\".\\n\\nWhat do you exactly mean by a policy to be reusable? Is it more like a transfer learning setting? If so, are there any experiments that justify that using the trained policy in a new task is useful? \\n\\nOverall, as you mention \\\"first uniformly set goals over the state space (maximize H(g)) and then separately learn to reach those goals (minimize H(g | s))\\\" - this seems to me like a very minor modification to several other tons of papers related to HER - where we randomly sample goal states and encourage the agent to reach these goal states. This paper seems like a very marginal contribution compared to that.\"}", "{\"title\": \"Citations for: Re: Official Blind Review #1\", \"comment\": \"Due to space constraints, the citations are included in this separate comment:\\n\\n[1] Hazan, Elad, et al. \\\"Provably Efficient Maximum Entropy Exploration.\\\" International Conference on Machine Learning. 2019.\\n[2] Bellemare, M, et al.. Unifying count-based exploration and intrinsic motivation. NeurIPS. 2016.\\n[3] Tang, H., et al. #Exploration: A Study of Count-Based Exploration for Deep Reinforcement Learning. NeurIPS, 2017.\\n[4] Burda, Y., et. al. Large-scale study of curiosity-driven learning. ICLR. 2019.\\n[5] Burda, Y., et. al. \\\"Exploration by random network distillation.\\\" ICLR. 2019.\\n[6] Nair, Ashvin, et al. \\u201cVisual Reinforcement Learning with Imagined Goals. Neural Information Processing Systems. 2019.\"}", "{\"title\": \"Re: Official Blind Review #1\", \"comment\": \"Thank you for the suggestion and detailed review. As suggested, we have modified the introduction to expand our discussion around H(s). We also answer the questions about the use of H(s) and H(s|g) below, and describe how the experimental results do in fact show that Skew-Fit substantially outperforms prior methods. We believe that these clarifications address the major criticisms raised in your review, but we would be happy to address any other points or discuss this further.\", \"q\": \"What\\u2019s the intuition behind the new objective MI(S;G)?\", \"the_mutual_information_provides_an_equivalent_interpretation_of_our_new_objective\": \"the new objective changes the exploration objective from \\u201cuniformly visit all the states,\\u201d as prior work has advocated, to a two stage process: first uniformly set goals over the state space (maximize H(g)) and then separately learn to reach those goals (minimize H(g | s)). At the optimum, the exploration policy will uniformly visits all states, and has the added benefit that we obtain a goal-conditioned policy that can be reused to reach goals.\\n\\n\\nExperiments\\nWe understand that there were concerns over the significance of the results. We find this concern surprising, as there is a clear difference between Skew-Fit and the next best prior work in Figure 5. Specifically, for the pickup task, Skew-Fit is the only method that makes significant progress: no prior method consistently picks up the object (Figure 6), and Skew-Fit\\u2019s final distance is approximately half that of the next best method. For the pushing tasks, the next best method results in a final distance that is 1.5 times worse than that of Skew-Fit, with an average score that is 3-4 standard deviations away from the average score of Skew-Fit. On the door task, some prior methods perform only slightly worse than Skew-Fit. However, we note that this task is much easier than the other tasks (the x-axis more than 4x shorter than the other tasks), as prior work [6] using these environments has also observed. Lastly, the difference on the real-robot experiments are particularly pronounced, with a final success rate double that of the prior method. While we acknowledge that the presentation of the results in the plots could be improved, the results themselves show that Skew-Fit is substantially better than all prior methods that we compared with.\\n\\nWe agree that it is informative to include a simplified experiment that does not directly jump to using goal-conditioned policies nor images. Therefore, Figure 3 of Section 6 analyzes a simplified 2D navigation task. While we did not have room to include in the main paper, Figure 9 of the appendix provides an \\u201cin between\\u201d experiment that does not contain images, but does include goal-conditioned policies.\"}", "{\"title\": \"Re: Official Blind Review #4\", \"comment\": \"Thank you for the review and suggestions. We have adjusted the experimental discussion to clarify a few points of confusion and to avoid possibly overstating the results.\\n\\n\\n> \\\"We also assume that the entropy of the resulting state distribution H(p(S | p\\u03c6)) is no less than the entropy of the goal distribution H(p\\u03c6(S)). Without this assumption, a policy could ignore the goal and stay in a single state, no matter how diverse and realistic the goals are.\\\" How do you ensure this in practice?\\nWe found that using RIG performed quite well. In particular, we found that using hindsight experience replay with the dense latent-distance ensured that the goal-conditioned policies consistently paid attention to the goal, and eventually learned to reach them.\\n\\n\\n> In the second paragraph of 2.2, it is written \\\"Note that this assumption does not require that the entropy of p(S | p\\u03c6) is strictly larger than the entropy of the goal distribution, p\\u03c6.\\\" Could you please clarify?\\nWe mean that the entropy of p(S | p\\u03c6) and p(\\u03c6) can be equal. It is unnecessary for the entropy to increase during exploration, since we increase it by changing the goal-distribution.\\n\\n\\n> The last sentences seems to imply that only this technique ends up optimizing the entropy of the state coverage, while I believe that the claim (given the experiment) should only be about the fact it does so faster.\\nWe agree that other methods can also eventually maximize the entropy. We have modified the sentence to clarify that we mean that Skew-Fit results in higher entropy faster.\\n\\n\\n> I'm unsure about the interpretation of this sentence given Figure 6 because other methods do not seem to fail entirely when given enough time.\\nThank you for pointing out this unclear phrasing. Since we are plotting the cumulative pickups, the success rate is given by the slope of the curves. While some prior methods do perform better than others, most curves have constant slopes after the first 40k iterations, meaning that their success rate does not increase over time. We have modified the text to clarify this.\\n\\n\\n> it is thus not clear where this \\\"5.5 hours\\\" comes from.\\nWe have corrected the text to say 6 hours. Thank you!\"}", "{\"title\": \"Re: Official Blind Review #2\", \"comment\": \"Thank you for the review and suggestions. Below, we address a number of questions asked and are happy to continue the discussion.\\n\\n> I would also like to see how Skew-Fit works with different goal-conditioned RL algorithms\\n\\nWe are currently running experiments that replace SAC with TD3. The only image-based, goal-conditioned RL algorithm other than RIG that we are aware of is DISCERN, which we found never learned. We are happy to take suggestions for alternate image-based, goal-conditioned RLs algorithm to try.\\n\\n\\n> More elaboration on this point is necessary.\\n\\nThank you for the suggestion. We have updated Section E to clarify, and include the new text here for convenience:\\n\\u201c...one can see that goal-conditioned RL generally minimizes H(G | S) by noting that the optimal goal-conditioned policy will deterministically reach the goal. The corresponding conditional entropy of the goal given the state, H(G | S) would be zero, since given the current state, there would be no uncertainty over the goal (the goal must have been the current state since the policy is optimal). So, the objective of goal-conditioned RL can be interpreted as finding a policy such that H(G | S) = 0. Since zero is the minimum value of H(G | S), then goal-conditioned RL can be interpreted as minimizing H(G | S).\\u201d\\n\\n\\n> Appendix has several broken references.\\n\\nThank you. We have fixed the references.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper introduced a very interesting idea to facilitate exploration in goal-conditioned reinforcement learning. The key idea is to learn a generative model of goal distribution to match the weighted empirical distribution, where the rare states receive larger weights. This encourages the model to generate more diverse and novel goals for goal-conditioned RL policies to reach.\", \"pros\": \"The Skew-Fit exploration technique is independent of the goal-conditioned reinforcement learning algorithm and can be plugged in with any goal-conditioned methods. The experiments offer a comparison to several prior exploration techniques and demonstrate a clear advantage of the proposed Skew-Fit method. It is evaluated in a variety of continuous control tasks in simulation and a door opening task on a real robot. A formal analysis of the algorithm is provided under certain assumptions.\", \"cons\": \"The weakest part of this work is the task setup. The method has only been evaluated on simplistic short-horizon control tasks. It\\u2019d be interesting to see how this method is applied to longer-horizon multi-stage control tasks, where exploration is a more severe challenge. It is especially when the agent has no access to task reward and only explores the environment to maximize state coverage. It is unclear to me how many constraints are enforced in the task design in order for the robot to actually complete the full tasks through such exploration.\\n\\nI would also like to see how Skew-Fit works with different goal-conditioned RL algorithms, and how the performances of the RL policy in reaching the goals would affect the effectiveness of this method in exploring a larger set of states.\", \"section_e\": \"it seems that there\\u2019s a logic jump before the conclusion \\u201cgoal-conditioned RL methods effectively minimize H(G|S)\\u201d. More elaboration on this point is necessary.\", \"minor\": \"Appendix has several broken references.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper introduces SKEW-FIT, an exploration approach that maximizes the entropy of a distribution of goals such that the agent maximizes state coverage.\\n\\nThe paper is well-written and provides an interesting combination of reinforcement learning with imagined goals (RIG) and entropy maximization. The approach is well motivated and simulations are performed on several simulated and real robotics tasks.\", \"some_elements_were_unclear_to_me\": [\"\\\"We also assume that the entropy of the resulting state distribution H(p(S | p\\u03c6)) is no less than the entropy of the goal distribution H(p\\u03c6(S)). Without this assumption, a policy could ignore the goal and stay in a single state, no matter how diverse and realistic the goals are.\\\" How do you ensure this in practice?\", \"In the second paragraph of 2.2, it is written \\\"Note that this assumption does not require that the entropy of p(S | p\\u03c6) is strictly larger than the entropy of the goal distribution, p\\u03c6.\\\" Could you please clarify?\", \"The experiments are interesting, yet some interpretations might be too strong (see below):\", \"In the first experiment, \\\"Does Skew-Fit Maximize Entropy?\\\", it is empirically illustrated that the method does result in a high-entropy state exploration. However, it is only compared to one very naive way of exploring and it is not discussed whether other techniques also achieve the same entropy maximization. The last sentences seems to imply that only this technique ends up optimizing the entropy of the state coverage, while I believe that the claim (given the experiment) should only be about the fact it does so faster.\", \"On the comments of Figure 6, the paper mentions that \\\"The other methods only rely on the randomness of the initial policy to occasionally pick up the object, resulting in a near-constant rate of object lifts.\\\" I'm unsure about the interpretation of this sentence given Figure 6 because other methods do not seem to fail entirely when given enough time.\", \"In the experiment \\\"Real-World Vision-Based Robotic Manipulation\\\", It is written that \\\"a near-perfect success rate [is reached] after five and a half hours of interaction time\\\", while on the plot it is written 60% cumulative success after 5.5 hours and it is thus not clear where this \\\"5.5 hours\\\" comes from.\"]}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary :\\n\\n\\nThe paper proposes an exploratory objective that can maximize state coverage in RL. They show that a formal objective for maximizing state coverage is equivalent to maximizing the entropy of a goal distribution. The core idea is to propose a method to maximize entropy of a goal distribution, or a state distribution since goals are full states. They show that the proposed method to maximize the state or goal distribution can lead to diverse exploration behaviour sufficient for solving complex image based manipulation tasks.\", \"comments_and_questions\": [\"The core idea is to maximize the entropy of the state visitation frequency H(s). It is not clear from the paper whether the authors talk about the normalized discounted weighting of states (a distribution) or the stationary distribution? The entropy of the state visitation distribution only deals with valid states - but I am not sure what it means to maximize the entropy of this term exactly in terms of exploration, since it is neither the discounted weighting of states or the stationary distribution for an infinite horizon task?\", \"The authors do mention that maximizing the entropy of H(s) is not sufficient - so instead suggests for maxmizing entropy of H(s|g). But why is this even sufficient for exploration - if I do not consider new tasks at test time but only the training task? How is this a sufficient exploration objective? Furthermore, since it is the conditional entropy given goal states, the fundamental idea of this is not clear from the paper.\", \"Overall, I am not convinced that an objective based on H(s|g) is equivalent to an maximizing H(s), and why is this even a good objective for exploration? The meaning of H(s) to me is a bit vague from the text (due to reasons above) and therefore H(s|g) does not convince to be a good exploration objective either?\", \"The paper then talks about the MI(S;G) to be maximized for exploration - what does this MI formally mean? I understand the breakdown from equation 1, but why is this a sufficient exploration objective? There are multiple ideas introduced at the same time - the MI(s;g) and talking about test time and training time exploration - but the idea itself is not convincing for a sufficient exploration objective. In light of this, I am not sure whether the core idea of the paper is convincing enough to me.\", \"I think the paper needs more theoretical insights and details to show why this form of objective based on the MI(s;g) is good enough for exploration. Theoretically, there are a lot of details missing from the paper, and the paper simply proposes the idea of MI(s;g) and talks about formal or computationally tractable ways of computing this term. While the proposed solutuon to compute MI(s;g) seems reasonable, I don't think there is enough contribution or details as to why is maximizing H(s) good for exploration in the first place.\", \"Experimentally, few tasks are proposed comparing skew-fit with other baselines like HER and AutoGoal GAN - but the differences in all the results seem negligible (example : Figure 5).\", \"I am not sure why the discussion of goal conditioned policies is introduced rightaway. To me, a more convincing approach would have been to first discuss why H(s) and the entropy of this is good for exploration (discounted weighting or stationary state distribution and considering episodic and infinite horizon tasks). If H(s) is indeed a difficult or not sufficient term to maximize the entropy for, then it might make sense to introduce goal conditioned policies? Following then, it might be convincing to discuss why goal conditioned policies are indeed required, and then tractable ways of computing MI(s;g).\", \"Experimentally, I think the paper needs significantly more work - especially considering hard exploration tasks (it might be simple setups too like mazes to begin with), and then to propose a set of new experimental results, without jumping directly to image based tasks as discussed here and then comparing to all the goal conditioned policy baselines.\", \"Overall, I would recommend to reject this paper, as I am not convinced by the proposed solution, and there are lot of theoretical details missing from the paper. It skips a lot of theoretical insights required to propose a new exploration based objective, and the paper proposes a very specific solution for a set a very specific set of experimental setups.\"]}" ] }

[ "Yan Wu", "Jeff Donahue", "David Balduzzi", "Karen Simonyan", "Timothy Lillicrap" ]

[ "GAN", "adversarial training", "generative model", "game theory" ]

{ "note_id": [ "gsT7X4hKdV", "2loZWAILWk", "iSaJsBR4fb", "tDgK7z-ny8", "rkef_7JQjH", "ByembX1QjB", "BJxDhRyCtS", "HJx7zgxTFr" ], "note_type": [ "official_comment", "comment", "official_comment", "decision", "official_comment", "official_comment", "official_review", "official_review" ], "note_created": [ 1578656695387, 1578575249006, 1577105563378, 1576798732851, 1573217129974, 1573217018554, 1571843758630, 1571778570745 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Paper1804/Authors" ], [ "~Michal_Sustr1" ], [ "ICLR.cc/2020/Conference/Paper1804/Authors" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1804/Authors" ], [ "ICLR.cc/2020/Conference/Paper1804/Authors" ], [ "ICLR.cc/2020/Conference/Paper1804/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1804/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"reply to questions\", \"comment\": \"Dear Michal,\\n\\nThank you for your questions.\\n\\nActually the data-dependent term has already been dropped from eq. 2. This is explained in the text above eq. 2: \\\"We focus on the second term in eq. 1 which is at the heart of the min-max game.\\\" Note that the data-dependent term does not participate in SGA (thus does not affect the following analysis), since its derivative with respect to generator parameters is 0.\\n\\nSeeing \\\\Delta z as the third player provides an intuitive interpretation of the algorithm. Although the first and last players share the same loss, they are not identical: the parameters of the third player are *only* \\\\Delta z. So the derivation following eq.13 simply treats \\\\Delta z as equal of \\\\theta_D and \\\\theta_G.\"}", "{\"title\": \"Question about derivation of SGA dynamics\", \"comment\": \"The dynamics g from eq. (3) seem wrong. The loss of the discriminator also depend on the training data, but this term is dropped without justification. All of the subsequent derivation then cannot be correct.\\n\\nIn appendix B.1 I don't understand how you can justify the computation of latent step $\\\\Delta z$ as being the third player. The first and last player in eq. (13) come from the same loss function, so their weights are identical, as in the latent optimization. SGA update will however shift those weights differently, so the weights for the first and third player will not be identical anymore.\\n\\nThank you for clarification!\"}", "{\"title\": \"Response to Meta-review\", \"comment\": \"The meta-review ignored the positive (two weak accepts) reviews, while raising new criticism that is tangential to our claims. The AC requires additional experimental comparison to the SGA algorithm. However, we only use SGA to provide a theoretical insight into our method, and don\\u2019t claim that our method is an improvement. Our main contribution is empirical, and consists in improving the state of the art by a large margin (>30%) -- this was recognised by the reviewers, but ignored by the AC. As we point out, and reviewer #1 agrees, our method can be viewed as an approximation of SGA which only requires minimal modification of standard GAN training. We mentioned the less straightforward Hessian-vector implementation suggested by the AC in Appendix B.1, and can discuss this in more detail. Further comparison might be an interesting subject for future work, but it\\u2019s not a subject of this paper. This point was not raised by either of the two reviewers.\", \"the_ac_also_mentioned_two_papers_which_we_missed\": \"[1] Sch\\u00e4fer & Anandkumar, 2019, Competitive Gradient Descent; \\n[2] Sch\\u00e4fer, Zheng & Anandkumar, 2019, Implicit competitive regularization in GANs. \\n\\nThe first was just published at NeurIPS 2019 (uploaded to arXiv in May), while the second is a *concurrent* submission to ICLR 2020, and was uploaded to arXiv in October *after* the ICLR deadline. We are happy to cite these papers, but we were not previously aware of them, and could not possibly be aware of the second one as it was not public at the moment of submission. We also note that neither of the original two reviewers mentioned them in their reviews.\"}", "{\"decision\": \"Reject\", \"comment\": \"The authors propose to overcome challenges in GAN training through latent optimization, i.e. updating the latent code, motivated by natural gradients. The authors show improvement over previous methods. The work is well-motivated, but in my opinion, further experiments and comparisons need to be made before the work can be ready for publication.\\n\\nThe authors write that \\\"Unfortunately, SGA is expensive to scale because computing the second-order derivatives with respect to all parameters is expensive\\\" and further \\\"Crucially, latent optimization approximates SGA using only second-order derivatives with respect to the latent z and parameters of the discriminator and generator separately. The second-order terms involving parameters of both the discriminator and the generator \\u2013 which are extremely expensive to compute \\u2013 are not used. For latent z\\u2019s with dimensions typically used in GANs (e.g., 128\\u2013256, orders of magnitude less than the number of parameters), these can be computed efficiently. In short, latent optimization efficiently couples the gradients of the discriminator and generator, as prescribed by SGA, but using the much lower-dimensional latent source z which makes the adjustment scalable.\\\"\\n\\nHowever, this is not true. Computing the Hessian vector product is not that expensive. In fact, it can be computed at a cost comparable to gradient evaluations using automatic differentiation (Pearlmutter (1994)). In frameworks such as PyTorch, this can be done efficiently using double backpropagation, so only twice the cost. Based on the above, one of the main claims of improvement over existing methods, which is furthermore not investigated experimentally, is false.\", \"it_is_unacceptable_that_the_authors_do_not_compare_with_sga\": \"both in terms of quality and computational cost since that is the premise of the paper. The authors also miss recent works that successfully ran methods with Hessian-vector products: https://arxiv.org/abs/1905.12103 https://arxiv.org/abs/1910.05852\", \"title\": \"Paper Decision\"}", "{\"title\": \"Revision and Clarification\", \"comment\": \"Thank you for your comments.\\n\\nWe have included your suggested changes in the current draft. Here we directly answer some main concerns:\\n\\n1. Symplectic Gradient Adjustment (SGA)\\n\\nWe have rewritten the section on how LOGAN relates to SGA to make it more clear. The main idea is that the latent step in LOGAN introduces a 3rd player into the game who is on the same side as the generator. \\n\\nNote that the \\u201cHessian\\u201d of a game differs significantly from the Hessian of a function: it involves second-order derivatives of multiple functions, corresponding to the losses of the players. The game Hessian thus decomposes into blocks labeled by the \\u201ctype\\u201d of interaction: G+G, G+D, G+Z, and so on. SGA makes use of the anti-symmetric component A of the game Hessian, where there are three kinds of blocks corresponding to second-order interactions between G+D, G+z and D+z respectively.\\n\\nSecond-order interactions between G+D are extremely expensive to compute. Thus, LOGAN only adjusts the gradients based on interactions between G+z and D+z. The partial approximation to SGA is well-founded game-theoretically in that it adjusts the dynamics between 2 of 3 possible pairings of the players \\u2014 which suffice for a substantial performance gain. \\n\\nIt is currently unclear how approximations like the Gauss-Newton generalize to game Hessians. \\n\\n\\n2. Natural Gradients (NG) and SGA: \\n\\nOur main motivation is indeed approximating SGA, and NG is the best way we found to implement the gradient updates in SGA. In other words, SGA does not specify the size of gradient steps, and NG provides an efficient way to adapt the step size under our approximation. We realised that the structure of our submission obscured this storyline, and re-organised it in our current draft.\\n\\n3. About the likelihood interpretation of the hinge loss (Appendix C): This section is relatively independent from the rest of the paper. NG has long been used as an approximate optimisation method without considering the probabilistic meaning of the loss. Similarly, SGA does not require the loss to be a likelihood function. We nevertheless provided this section for readers interested in the probabilistic view and acknowledged the limit of this interpretation in the text.\\n\\n4. Your comment about the typo is correct, thanks for pointing it out. It should read: \\u201can ideal generator can perfectly fool the discriminator\\u201d. This assumption is actually unnecessary, and we removed it. Yes, the likelihood function p(t|z) has the same form as the p in eq. 37. We rewrote this part to make the correspondence more clear.\\n\\n5. Thank you for pointing out this typo. The clipping should be after updating z. We have not experimented with many other algorithms for optimising z, and would like to investigate other methods including mirror descent in the future. We will also experiment with ELU activation.\"}", "{\"title\": \"Open Source\", \"comment\": \"Thank you for your comments.\\n\\nWe will open source the CIFAR10 model and training code for the camera-ready version. In the meantime, we would be glad to answer your questions about the implementation, as well as to add any modelling details that might be missing. We appreciate your effort in attempting to replicate the results!\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\n\\nLOGAN optimizes the sampled latent generative vector z in conjunction with the generator and discriminator. By exploiting second order updates, z is optimized to allow for better training and performance of the generator and discriminator.\", \"pros\": [\"A relatively efficient way of exploiting second order dynamics in GAN training via latent space optimization.\", \"A good set of experiments demonstrating the superior performance of the proposed method on both large and small scale models and datasets.\"], \"cons\": [\"Lack of code\"], \"comments\": \"All in all, this appears to be a solid contribution to the GAN literature, addressing some of the limitations of CS-GAN [1]. The lack of open source code accompanying the paper (in this day and age) does it a serious disservice. I have already tried and failed to replicate the cifar10 results. There appears to be some nuance in implementation that would probably clear up if the authors release their code along with the paper.\\n\\n[1] - http://proceedings.mlr.press/v97/wu19d.html\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposes a novel training scheme for GANs, which leads to improved scores w.r.t. state-of-the-art. The idea is to update the sampled latent code in a direction improving the inner maximization in the min-max problem. The work considers both, the gradient direction and a direction motivated by the natural gradient which is shown to yield excellent performance. The overall scheme is motivated as an approximation to the (prohibitively expensive) symplectic gradient adjustment method.\\n\\nWhile the work contains some typos, it was easy to read and overall very well written. The experimental results are impressive and clearly validate the usefulness of the approach. Therefore, I recommend acceptance of the paper. \\n\\nOn the theoretical side, I feel some parts can be improved. I'm willing to further increase my rating if some of the following points are addressed:\\n\\n1. The connection to SGA is a bit hand-wavy, as terms are dropped or approximated with the only reasoning that they are difficult to compute. In some approximations (e.g. Gauss-Newton approximation of the Hessian) one can argue quite well that certain second-order terms can be dropped under some assumptions (e.g. mild nonlinearity, or vanishing near the optimum).\\n\\n2. I had some troubles Sec 4 related to the natural gradient.\\n(a) What is p(t | z)? Is it the same p as in (37), Appendix C? I find the argument that the hinge loss pushing D(G(z)) < 0 during training hand-wavy, as for natural gradient p(t | z) should always be a valid distribution. There are also some typos which made it hard to follow the arguments there. It probably should read \\\"an ideal generator can perfectly fool the discriminator\\\" (not perfectly fool the generator). \\n\\n(b) Maybe it would be easier to directly argue that one wishes to approximate SGA as well as possible, rather than taking a detour through the natural gradient.\\n\\n3. Why is the increment \\\\Delta z clipped in Algorithm 1? Is there a theoretical justification? If the goal of the clipping is to stay inside the support of the uniform distribution, shouldn't it rather be z' = [z + \\\\Delta z]? A soft clipping (e.g. performing a mirror descent step) might give better gradients.\\n\\nTypos, minor comments (no influence on my rating):\\n- In Eq. (5), it should be \\\\partial^2 in the last \\\"Hessian-block\\\" multiplication.\\n- presribed -> prescribed\\n- Eqs. (7) and (8) might be easier to parse if one uses a different notation to distinguish between total derivative and partial derivative, i.e., write on the left side df(z') / d\\\\theta. Also, I think it is clearer to write in the last terms in (7) and (8) \\\\partial f(z') / \\\\partial \\\\delta z instead of \\\\partial f(z') / \\\\partial z'.\\n- In Appendix B.1, shouldn't it be \\\\gamma=(1+\\\\eta)/2 instead of \\\\gamma=\\\\eta (1+\\\\eta)/2?\\n- I've found ELU activations to work well in GAN models which involve the Jacobian w.r.t. z. Maybe it can stabilize things here as well.\"}" ] }

[ "Ruochi Zhang", "Yuesong Zou", "Jian Ma" ]

[ "graph neural network", "hypergraph", "representation learning" ]

{ "note_id": [ "-LPdBvq-TPN", "FQ09y0xlXn", "8sZznAaBz", "ByxHQc2ojr", "S1eSbb3isH", "rkld5J3oor", "SkgvBi60YB", "HJlYRqICtH" ], "note_type": [ "official_comment", "comment", "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review" ], "note_created": [ 1587074164918, 1576864359799, 1576798732823, 1573796381071, 1573794045143, 1573793680035, 1571900222935, 1571871440525 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Paper1803/Authors" ], [ "~Ted_Liu1" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1803/Authors" ], [ "ICLR.cc/2020/Conference/Paper1803/Authors" ], [ "ICLR.cc/2020/Conference/Paper1803/Authors" ], [ "ICLR.cc/2020/Conference/Paper1803/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1803/AnonReviewer3" ] ], "structured_content_str": [ "{\"title\": \"Code for Hyper-SAGNN\", \"comment\": \"The code for Hyper-SAGNN is now maintained under: https://github.com/ma-compbio/Hyper-SAGNN\"}", "{\"title\": \"Some recent work on hypergraph embedding not mentioned in the paper\", \"comment\": \"Heterogeneous Hyper-Network Embedding, ICDM'18\\nModeling Multi-way Relations with Hypergraph Embedding, CIKM'18\\nHyper-Path-Based Representation Learning for Hyper-Networks, CIKM'19\"}", "{\"decision\": \"Accept (Poster)\", \"comment\": \"This work introduces a new neural network model that can represent hyperedges of variable size, which is experimentally shown to improve or match the state of the art on several problems.\\n\\nBoth reviewers were in favor of acceptance given the method's strong performance, and had their concerns resolved by the rebuttals and the discussion. I am therefore recommending acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Summary of revisions made to the paper\", \"comment\": [\"We would like to thank the reviewers for all the comments and suggestions. We have made revisions to the paper to address all the constructive comments. Our major changes include:\", \"In section 2, we now mentioned more related works as suggested by the reviewer and highlighted the differences and novelty in our work.\", \"In section 4.4, we have added a sentence to clarify the difference between the hyperedges of the scHi-C datasets with the other hypergraph datasets used in this work.\", \"In the Appendix section A.3, we have added more hyperparameters related to the training of our method.\", \"We have added a section in the Appendix (A.4) to describe the evaluation tasks and the metrics used in this work.\", \"We have added a section in the Appendix (A.5) to describe the analysis of the dynamic embeddings to explain our rationale of the static/dynamic embedding design more clearly.\"]}", "{\"title\": \"Response to Reviewer 2\", \"comment\": \"Response: We would like to thank the reviewer for the important suggestions. We have thoroughly addressed the questions and provided point-by-point responses to the reviewer\\u2019s comments.\\n\\n>>> A possible weakness of this paper is that the evaluation data sets are all k-uniform hypergraphs with k=3. This is perhaps the minimal case which their method can address. For all the discussion of generality to heterogeneous hyperedges, it would have been better to include some dataset (even if synthetic) to establish baseline performance over while ablating k>3 and multiple hyperedge types. This remains completely un-investigated (even the genome dataset appears to be k=3?). Although, their edge/hyperedge ablation on these datasets partially addresses the point and the result (Fig 5) seems promising.\", \"response\": \"We apologize for the confusion. We have changed the equations following the reviewer's suggestion. We added a section called \\\"Experiment Setting\\\" in the Appendix to better describe the evaluation tasks and metrics. We now also listed the training hyperparameters including the optimizer, learning rate, batch size, etc in the \\\"Parameter setting\\\" section in the Appendix.\"}", "{\"title\": \"Response to Reviewer 3\", \"comment\": \"\", \"response\": \"We thank the reviewer for pointing out these prior work. We added references of these two papers and discussed them in the related work section. These methods generalize the convolution and attention operation of graphs to hypergraphs and are typically used for semi-supervised node classification task on hypergraphs with known node attributes. In contrast, our new method Hyper-SAGNN is designed to learn the function for hyperedge prediction and can be used for hypergraphs without node attributes.\", \"the_related_work_section_should_be_expanded_by_discussing_relationships_to\": \"Bai, Song, Feihu Zhang, and Philip HS Torr. \\\"Hypergraph Convolution and Hypergraph Attention.\\\" arXiv preprint arXiv:1901.08150 (2019).\\n\\nFeng, Yifan, et al. \\\"Hypergraph neural networks.\\\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a new graph neural network model capable of performing tasks over hyperedges of variable type and size (i.e. different number and types of graph nodes connected by different hyperedges). They experimentally verify its effectiveness over the previous state of the art on several datasets and tasks.\\n\\nI lean toward accepting this paper. It is a relevant topic area, the design appears novel relative to recent work, and the presentation is mostly clear. The design decisions are well-motivated and discussed (e.g. the choices to use the static embedding, the exclusion of the diagonal self-attention weights). Performance is tested on previously used datasets and tasks, for a thorough comparison against recent best methods (DHNE). The model is clearly described.\\n\\nA possible weakness of this paper is that the evaluation data sets are all k-uniform hypergraphs with k=3. This is perhaps the minimal case which their method can address. For all the discussion of generality to heterogeneous hyperedges, it would have been better to include some dataset (even if synthetic) to establish baseline performance over while ablating k>3 and multiple hyperedge types. This remains completely un-investigated (even the genome dataset appears to be k=3?). Although, their edge/hyperedge ablation on these datasets partially addresses the point and the result (Fig 5) seems promising.\", \"some_other_notes\": \"Eq 5. Given that it is the default, explicitly write that the summation is over j!=i\\nEq 8. Probably leave the ellipses out between v_{i-1} and v_{i+1}?\\nThe tasks and measures like AUC, AUPR, and network reconstruction could be described somewhere, even if in an appendix.\\nSome/many training hyperparameters not listed?\", \"edit\": \"Post rebuttal, authors have provided evidence of good performance on data sets with k=4 and k=5, addressing my main concern, and have addressed the smaller notes. Raising score but still am not expert in this area.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\n\\nThis work is tackling the problem of predicting variable-sized heterogeneous hyperedges in hypergraphs. The authors do so by introducing a multi-head attention mechanism that takes in a set of node features and predicts the probability of this set of nodes forming a hyperedge.\\n\\n- Quality and clarity: \\n\\nThe quality of the paper is good, the writing is clear and the reasoning is relatively easy to follow. The problem the authors are trying to tackle is also clearly defined. The motivation of the dynamic and static embedding choice could be explained better. The intuition that the pseudo-distance tells us how well the static embedding\\nof a node can be approximated by the features of the neighbour within the tuple somewhat makes sense, but I also have the impression that this is just one possible interpretation and what happens in practice could actually be quite different. A more formal treatment and analysis of this intuition would be very helpful.\\n\\n- Originality: \\nThe main contribution here is to propose a more flexible formulation to improve over the fixed MLP used by DHNE. This in itself is a good idea, but overall is an incremental improvement.\\n\\n- Significance: \\nThe results are convincing, DHNE is clearly outperformed on the benchmarks. Depending on the ease of training the model compared to DHNE, this work might be a useful contribution for the ICLR community.\\n\\nOther points\", \"the_related_work_section_should_be_expanded_by_discussing_relationships_to\": \"Bai, Song, Feihu Zhang, and Philip HS Torr. \\\"Hypergraph Convolution and Hypergraph Attention.\\\" arXiv preprint arXiv:1901.08150 (2019).\\n\\nFeng, Yifan, et al. \\\"Hypergraph neural networks.\\\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019.\\n\\n------------------\\nPost-rebuttal\\n------------------\\n\\nI would like to thank the authors for their in-depth response. I found the additional analysis on the dynamic node embeddings insightful and it is reassuring that it confirms the claims. Overall I think the rebuttal addresses my major concerns and I will raise my score.\"}" ] }

[ "Soochan Lee", "Junsoo Ha", "Dongsu Zhang", "Gunhee Kim" ]

[ "continual learning", "task-free", "task-agnostic" ]

{ "note_id": [ "W40Au5q6T8", "nPz0qpRdgq", "r1lgVyEYsr", "BklLaCQFsH", "r1e__A7KiB", "r1l_mbemoB", "rke8X-igiB", "rygcWQYf9r", "r1x6ECnhYr", "Syx2wD4nKB" ], "note_type": [ "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1577958547350, 1576798732793, 1573629736124, 1573629630362, 1573629551574, 1573220640161, 1573069086363, 1572143873939, 1571765813108, 1571731299596 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Paper1802/Authors" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1802/Authors" ], [ "ICLR.cc/2020/Conference/Paper1802/Authors" ], [ "ICLR.cc/2020/Conference/Paper1802/Authors" ], [ "ICLR.cc/2020/Conference/Paper1802/Authors" ], [ "~Erin_Grant1" ], [ "ICLR.cc/2020/Conference/Paper1802/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1802/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1802/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Camera-ready version is uploaded.\", \"comment\": \"We thank the anonymous reviewers and the program chair for their constructive and thoughtful comments. We just have uploaded the camera-ready version of our paper and publicly released the code through GitHub (https://github.com/soochan-lee/CN-DPM). We have largely updated section 2 for better comparison with other related papers such as (Nagabandi et al., 2019; Jerfel et al., 2019; Rao et al., 2019).\\n\\nWe also have added appendix M to compare our method with CURL (Rao et al., 2019). To summarize the differences, CURL has some components (the decoder network $p(x|z)$ and the gating network $q(y|x)$) shared across all tasks due to its generative process. As a consequence, expansion alone is not enough to fully prevent catastrophic forgetting, and the generative replay is added to mitigate forgetting in the shared components. Therefore, CURL can be regarded as a hybrid of replay and expansion methods, whereas our CN-DPM is solely expansion-based.\"}", "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper proposes an expansion-based approach for task-free continual learning, using a Bayesian nonparametric framework (a Dirichlet process mixture model).\\n\\nIt was well-reviewed, with reviewers agreeing that the paper is well-written, the experiments are thorough, and the results are impressive. Another positive is that the code has been released, meaning it\\u2019s likely to be reproducible.\\n\\nThe main concern shared among reviewers is the limited novelty of the approach, which I also share. Reviewers all mentioned that the approach itself isn\\u2019t novel, but they like the contribution of applying it to task-free continual learning. This wasn\\u2019t mentioned, but I\\u2019m concerned about the overlap between this approach and CURL (Rao et al 2019) published in NeurIPS 2019, which also deals with task-free continual learning using a generative, nonparametric approach. Could the authors comment on this in their final version?\\n\\nIn sum, it seems that this paper is well-done, with reproducible experiments and impressive results, but limited novelty. Given that reviewers are all satisfied with this, I\\u2019m willing to recommend acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer 3\", \"comment\": \"We appreciate your positive and thoughtful review. Please check appendix E and L of the updated draft for added experiment and graphs. Below are our answers to your questions.\\n\\n\\n1. Evaluation on each task\\n\\n[1] and [2] evaluate their models only on the tasks that have been presented so far. On the other hand, we always evaluate our model on the entire test set following [3] (no matter whether presented or not). Specifically, the learning curves of Figure 2 show how the accuracy on the entire CIFAR10 test set increases as more tasks are learned.\\n\\nTable 3 shows that forgetting is minimal in the classifiers of CN-DPM. At the end of each task, we measure the test accuracy of the responsible classi\\ufb01er (labeled as \\u201cClassi\\ufb01er (init)\\u201d). We also measure the accuracies after learning all tasks (labeled as \\u201cClassi\\ufb01er (\\ufb01nal)\\u201d). We observe little difference between the two scores, which con\\ufb01rms that forgetting barely occurs in the classi\\ufb01ers.\\n\\nIn appendix L of the updated manuscript, we add more learning curves following [1] and [2]. For Split-CIFAR10/100, we report (i) how the test accuracy of each task changes during training and (ii) how the average test accuracy of learned tasks decreases. Although each expert is hardly updated after their assigned task, the accuracies gradually decrease since the gating performance degrades as an increasing number of VAEs are involved.\\n\\n\\n2. The STM size\", \"there_is_a_tradeoff_in_the_stm_size\": \"as the STM size grows, the performance generally increases, but we need a stronger assumption on the data stream. In our paper, we simply set the STM size to 500 for MNIST and 1000 for the others, following the size of replay memory in [3]. We believe this is a reasonable setting since it is about 1-2% of the training set size.\\n\\nAppendix I presents the results of Split-CIFAR10/100 with a smaller STM of size 500. Compared to Reservoir whose performance drops significantly, CN-DPM keeps competitive scores even with a halved STM.\\n\\nOur basic assumption on tasks is that the data points for each task appear together. Otherwise, we may need to allow each task being revisited multiple times, for which we performed experiments in appendix H.\\n\\n\\n3. CN-DPM with known task boundaries\\n\\nOriginally, Appendix E compared the performance of CN-DPM with other task-based regularization/replay methods (EWC, Online EWC, SI, MAS, LwF, GEM, DGR, and RtF) on the Split-MNIST scenario. The compared methods are applied to a single-headed neural network such that task information is not needed at test time. Despite this handicapped setting, CN-DPM outperformed the task-based methods by significant margins thanks to the expansion paradigm.\\n\\nIn the updated draft, we add an additional experiment for CN-DPM with known task boundaries to appendix E. Given the task boundaries, we do not need to infer the responsible expert. Instead, we simply expand the model at every task boundary and train a new expert for the new task. With the task boundary, the average accuracy further increases from 93.70 to 93.81.\\n\\n\\n4. The term \\u201cBayesian\\u201d \\n\\nWe admit that our method is not \\u201cfully Bayesian\\u201d as in VCL. Given that our method uses the DPM as an overall framework and performs MAP estimation for experts, we believe that a \\u201cBayesian framework with MAP estimation\\u201d may be the best term describing our method. We will clarify the distinction more strictly.\\n\\n\\n5. The effect of lateral connections\\n\\nIt is a legitimate point that parameter sharing makes the posterior of \\u03c6 no longer independent anymore. However, the argument of [4] that the number of clusters is \\u201ctypically\\u201d upper bounded by $O(\\\\alpha \\\\log N)$ is a trend, rather than a theoretical result. More precisely, the expected number of clusters in the Chinese restaurant process, which is the prior of DPM, is $\\\\alpha \\\\log N$. In other words, $\\\\alpha \\\\log N$ is the expected number of clusters before we see any data. Therefore, depending on the distribution of the data, the number of clusters can be larger than $O(\\\\alpha * \\\\log N)$.\\n\\n\\n[1] Arslan Chaudhry, Marc\\u2019Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient Lifelong Learning with A-GEM. ICLR, 2019.\\n[2] David Lopez-Paz and Marc\\u2019Aurelio Ranzato. Gradient Episodic Memory for Continual Learning. NeurIPS 2017.\\n[3] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. arXiv, (1903.08671v4), 2019.\\n[4] Yee Whye Teh. Dirichlet process. Springer, Encyclopedia of Machine Learning:280\\u2013287, 2010.\"}", "{\"title\": \"Response to Reviewer 2\", \"comment\": \"We appreciate the thoughtful review. Below, we present our answers to the questions.\\n\\n\\n1. Feasibility of the posterior inference\\n\\nThis is closely related to the first question of R1. CN-DPM may not be successful when the discrepancy between tasks is too marginal. Conversely, the posterior inference gets easier as the discrepancy between tasks becomes larger.\\n\\nWe believe that most existing continual learning benchmarks are defined to have enough discrepancies between the tasks. In the class-incremental scenarios (e.g. Split-CIFAR100), the difference between the labels largely separates the tasks. In the domain-shifting scenario (e.g. MNIST-SVHN), the obvious differences between grayscale and color images distinguish the tasks.\\n\\n\\n2. The necessity of generative models\\n\\nIn task-free continual learning, it becomes a critical problem at test time to choose the right expert for a given data point since no task information is available. For example, if M experts are created during training, choosing an expert for a test data is a type of M-class classification problem. If we use a single gate network to solve this problem, catastrophic-forgetting may occur in the gating network. On the other hand, we propose the combination of generative models as a forgetting-free gating mechanism since each generative model is expert-specific.\\n\\n\\n3. Concerns on learning generative models \\n\\nWe agree that the generative models would not be trained properly in a fully online scenario. However, to be precise, each expert in CN-DPM is not trained in a fully online manner; during the sleep phase, each expert is trained on data in the STM (with a size of 500 or 1000) for multiple iterations. Despite the small size of STM, we empirically confirmed that the sleep phase provides enough bootstrapping for the generative models to be functional.\\n\\n\\n4. About the known concentration parameter\\n\\nAs the review mentioned, it is possible to infer the value of alpha, although it could be challenging in an online setting like ours. Instead, we want to emphasize that knowing the value of alpha ahead of time is different from knowing the number of tasks beforehand. It is rather knowing the degree of discrepancy between tasks. As long as the discrepancies between tasks remain constant, the model can still learn indefinitely many tasks. Also note that the last paragraph of Appendix B explains the underlying assumption behind choosing the concentration parameter with a concrete example.\\n\\n\\n5. More explanation on learning p(z=k|x)\\n\\nCN-DPM does not learn $p(z=k|x)$ directly. Instead, it has a set of generative models that independently learns $p(x|z=k)$. For a given task, only one generative model of the responsible expert is trained. Then, $p(z=k|x)$ is simply computed as $\\\\frac{p(x|z=k) p(z=k)}{\\\\sum_{k\\u2019} p(x|z=k\\u2019) p(z=k\\u2019)}$.\\n\\nWe also reported the gating accuracy of learned $p(z=k|x)$ on CIFAR10/100 in Table 3 as \\u201cVAE\\u201d (which is updated to \\u201cGating (VAEs)\\u201d). The gating accuracy is computed against the ground-truth task label after the whole scenario has finished.\"}", "{\"title\": \"Response to Reviewer 1\", \"comment\": \"Thank you for the constructive and encouraging comments. We have updated appendix E with an additional experiment and provide our responses below.\\n\\n\\n1. Any cases in which the proposed DPM would fail\\n\\nCN-DPM may not be successful when the discrepancy between tasks is too marginal. Since the model cannot access previous tasks in continual learning scenarios, new data need to be sufficiently different such that it is classified as new and sent to the STM. Otherwise, this slightly different data is sent to an existing expert causing weak forgetting.\\n\\n\\n2. Known task boundaries\\n\\nOriginally, Appendix E compared the performance of CN-DPM with other task-based regularization/replay methods (EWC, Online EWC, SI, MAS, LwF, GEM, DGR, and RtF) on the Split-MNIST scenario. The compared methods are applied to a single-headed neural network such that task information is not needed at test time. Despite this handicapped setting, CN-DPM outperformed the task-based methods by significant margins thanks to the expansion paradigm.\\n\\nIn the updated draft, we add an additional experiment for CN-DPM with known task boundaries to appendix E. Given task boundaries, we do not need to infer the responsible expert. Instead, we simply expand the model at every task boundary and train a new expert for the new task. With the task boundary, the average accuracy further increases from 93.70 to 93.81.\\n\\n\\n3. Applications to the RL setting\\n\\nWe believe that our method has great potential to be applied to RL settings since it is a general framework to learn distributions incrementally. As an example, CN-DPM may be applied to policy gradient methods; each expert consists of a policy network, which predicts an action for a given state, and a generative model, which recognizes whether the expert is responsible for the given state.\"}", "{\"title\": \"Clarification of our novelty over [1].\", \"comment\": \"We appreciate the authors of [1] for the detailed comments. We acknowledge that the differentiation with [1] was rather cursory in the initial submission due to the page limitation. This will be clarified in the final draft.\\n\\nFirst of all, we were surprised to find that our work has technical similarity to [1] (i.e. the use of DPM) as a parallel submission to NeurIPS 2019 (unfortunately, this work was not accepted even with very positive reviews mainly due to experimental evaluation, which was substantially improved in this ICLR submission).\\n\\nHowever, we believe there are three significant differences between our work and [1].\\n\\n1. The form of continual learning and the way DPM is applied are different.\\nIn our formulation, a data point to be fed into DPM is a single example (i.e. an image), whereas a datapoint in [1] corresponds to a task (i.e. a set of support and query examples). Thus, at both training and test time, the model of [1] receives data on a task-by-task basis and computes the responsibility for each task, not for each data sample as in our CN-DPM. That is, task-agnostic continual learning takes place at the meta-level in [1]: similar tasks are grouped into a \\u201csuper-task\\u201d for which a meta-expert is responsible, and the super-tasks are automatically discovered by the DPM framework. Therefore, [1] is a method for continual \\\"few-shot\\\" learning rather than general continual learning. Moreover, it cannot be generalized into a standard continual learning setting for the reason described in the following paragraph.\\n\\n2. Our generative gating mechanism, which lacks in [1], is critical for the test-time inference.\\nIt is not true that [1] can be applied to general continual learning such as our experimental settings. It could be possible to train the model of [1] in our continual learning scenario, but it is not possible to apply it at test time since there is no gating mechanism to choose the responsible expert at test time in which no ground truth label is available. In other words, [1] can be task-agnostic only in the settings where class labels are available at test time, such as the continual \\u201cfew-shot\\u201d learning which is the only type of continual learning experiment in [1].\\n\\nSuch limitation of [1] arises from the absence of generative component p(x | \\u03c6) when computing responsibility: p(\\u03c6 | x, y) \\u221d p(y | x, \\u03c6) p(\\u03c6). This formulation of [1] is okay if we have (x, y) pairs at test time such that we can predict the posterior: p(\\u03c6 | x, y, X, Y) \\u221d p(y | x, \\u03c6) p(\\u03c6 | X, Y). However, in the case of discriminative continual learning scenarios (Split-MNIST, Split-CIFAR10/100, \\u2026), we need to infer p(\\u03c6 | x, X, Y) at test time, which is not possible with [1]\\u2019s formulation. This is a typical problem of expansion-based continual learning, and we solve it by adding a generative component to the formulation: p(\\u03c6 | x, y) \\u221d p(y | x, \\u03c6) p(x | \\u03c6) p(\\u03c6), p(\\u03c6 | x, X, Y) \\u221d p(x | \\u03c6) p(\\u03c6 | X, Y).\\n\\n3. The training procedure is different.\\nAlthough DPM can provide the theoretical basis for task-free continual learning, many additional components are required to stably integrate with neural networks in practice. Therefore, we introduce the wake-sleep cycle with short-term memory (STM), which is crucial to handle a data stream entering one-by-one; otherwise, fitting neural networks to a single data point leads to severe overfitting. On the other hand, [1] does not need a buffer like STM since the experimental setup already assumes that the data comes in as a task (i.e. a set of supports and query examples). We also propose several practical training techniques such as parameter sharing and temperature parameters for the expert classifier.\\n\\n\\nIn summary, [1] is a method for continual \\u201cfew-shot\\u201d learning and cannot be applied to general task-free continual learning since it requires task information at test time as many other expansion methods do. Our CN-DPM, on the other hand, is the first expansion method that can be applied to general task-free continual learning.\"}", "{\"title\": \"The claim that \\\"task-free continual learning\\\" is a novel contribution neglects a prior work [1] which, while cited, is misrepresented in this manuscript.\", \"comment\": \"We congratulate the authors on their submission.\\n\\nHowever, we would like to note that the claim that \\\"expansion-based, task-free continual learning\\\" is a novel contribution ignores a prior work [1] that has been accepted to NeurIPS 2019. The authors cite [1] briefly in the related work section, stating that [1] does not extend beyond meta-learning without going into significant detail. However, in [1], we also adapt the sequential variational approximation approach of Lin et al. (2013) for the Dirichlet process mixture (DPM) and present derivations mirroring those in the submission under review.\\n\\nIn the related work section, the submission claims that [1] \\\"...lacks a generative component which is a crucial element to complete the DPM formulation.\\\" While we demonstrate parameter estimation in a mixture of regularized convnet classifiers (as opposed to using a generative component), the procedure remains a \\\"sound\\\" definition of a DPM model. Our approach is furthermore not inconsistent with a mixture of Bayes classifiers (i.e., including a generative component p(x | \\u03c6) as in the submission), as it relies only on gradient-based optimization of the components (though we do not test such an architecture empirically). More importantly, [1] also addresses the catastrophic forgetting problem with a DPM-based expansion approach.\\n\\nThe main difference lies in the fact that we apply the algorithm in [1] to the few-shot learning setting, in which the learner receives a batch of data that is known to come from the same domain, and some amount of domain-specific fine-tuning is allowed (i.e., for the bilevel optimization approach of meta-learning). This allows the computation of the assignments to experts (\\u03c1 in the submission; \\u03b3 in [1]) and the update for expert parameters (line 22 in Algo. 1 in the submission; M-STEP in [1]) to be computed using separate data points. The present submission, in contrast, computes these quantities with the same data point; it further does not allow domain-specific fine-tuning. As such, [1] can be immediately applied to the setting of this paper (with the special case of setting the number of fine-tuning steps to 0, and computing the responsibilities \\u03b3 and the M-STEP using the same datapoint).\\n\\nIn summary, we believe the methodological differences are not as great as claimed in the current submission, and that most of the differences lie in the experimental setup.\\n\\nRegards,\\nAuthors of [1]\\n\\n[1] Jerfel, G., Grant, E., Griffiths, T. L., and Heller, K. A. (2019). Reconciling meta-learning and continual learning with online mixtures of tasks. In NeurIPS 2019. https://arxiv.org/abs/1812.06080\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposes an elegant method for task-free continual learning problems. It has nicely pointed out that the conventional continual learning algorithms had the limitation of knowing the task boundaries. Followings are my summary.\", \"summary\": \"By applying DPM, the authors proposed a method of automatically determining whether to add a new expert for a new task or train the existing experts. While the Dirichlet Process Mixture (DPM) is not new, applying such nonparametric method to continual learning is new. The experimental results are impressive given the single-epoch setting.\", \"pros\": \"1. Good experimental results for the task-free setting, in which no information about task boundaries is given. Particularly, even with smaller memory-usage than the experience replay (ER) methods, the proposed method achieves better results. \\n2. Many past work should suffer from increased number of tasks due to the model capacity limit, but the proposed method efficiently expands the model capacity. \\n3. Writing flow is good and is easy to follow. \\n\\nCons & Questions: \\n1. I am not sure whether the proposed methods should work well for \\\"all\\\" cases. Is there any cases in which the proposed DPM would fail?\\n2. What happens when you actually know the task boundaries? Would following the framework with known number of experts also excel other methods?\\n3. Can you apply this to the RL setting?\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary: The paper proposes to use a Bayesian nonparametric mixture model for task-free (without explicit task labels) continual learning. The main idea is to use an expansion-based model where the number of mixture components (experts) adapts to the training data/tasks. Specifically, a Dirichlet Process Mixture Model (DPMM) consisting of a set of neural network experts is used. Empirical results demonstrate improved performance on three different datasets over some of the baselines.\\n\\nI find the methodological contribution in the paper to be somewhat limited since the main idea of the model was initially proposed in the prior work (cited in the main paper): Dahual Lin - \\u201cOnline Learning of Nonparametric Mixture Models via SVA\\u201d. In fact, the paper claims its contribution is expansion-based task-free continual learning. However, this \\u201ctask-free characteristic\\u201d is the contribution of SVA based inference. Nevertheless, I do like how the existing SVA based inference has been adapted from an online learning setting to a more general continual learning setting by using various approximations/tricks (like short-term memory with wake-sleep training, point estimates).\", \"pros\": [\"The idea of using a nonparametric model for CL is interesting and can lead to follow-up work.\", \"Results show that the approach works well.\", \"The code has been released.\"], \"overall_i_am_inclining_towards_voting_for_acceptance_if_the_authors_could_address_my_following_questions\": [\"Could you comment on the creation of test data? It is not clear to me how the model is evaluated if the task-boundaries are not known a priori. Shouldn\\u2019t the evaluation be based on tasks? How are you evaluating catastrophic forgetting? I am interested to know what was the test accuracy for the task for which the training data was seen early on during the training.\", \"It seems to me that the method works on the assumption that the number of data points for each task is at least M (size of STM) and moreover, that these data points appear together sequentially. The method should be sensitive to the size of the STM. How are you choosing M? Would the framework work if data points for each task do not appear together?\", \"Assuming you have clear task boundaries, how would you adapt this framework? Was the model compared to other methods that assume known task-boundaries like (VCL, EWC, Memory Replay)?\"], \"other_comments\": [\"The method is inspired by a Bayesian framework but calling it Bayesian wouldn\\u2019t be fair since only a point estimate is being learned for parameters. This is important to distinguish since there are other methods that are fully Bayesian like Nguyen et. al. \\u201cVariational Continual Learning\\u201d (although such methods may have other pros and cons)\", \"The samples from the base distribution of the posterior (v) are not iid anymore due to lateral connections b/w the representations. Do you think the theoretical result in Appendix B that the number of clusters is upper bounded by O(alpha*logN) is still valid?\"]}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes Continual Neural Dirichlet Process Mixture Model (CN-DPM) to solve task-free continual learning. The core idea is to employ Dirichlet process mixture model to create novel experts in online fashion when task distributions change. The proposed method is validated on various tasks and demonstrated to perform well compared to the other baselines.\\n\\nOverall I find this paper to be well-written and the experiments are conducted thoroughly. The method is compared to proper baselines in various settings, and the paper describes detailed experimental settings and architectural choices to help readers willing to reproduce. I\\u2019ve gone through the appendix and they provide enough additional experiments to support the author\\u2019s claim.\\n\\nThe main algorithm itself cannot be considered to be novel. DPM or other Bayesian nonparametric models have been extensively used for the problems requiring to adapt the model size according to the change of data. Nevertheless, the application of DPM in task-free continual learning context seems to be considered as a contribution.\\n\\nI have much experience in implementing Bayesian nonparametric models with parametric distributions and compared various methods to conduct the posterior inference of them. In my experience, even for the low-dimensional parametric models, the posterior inference algorithms for DPM usually suffer from local optima, and the sequential methods such as SVA depends heavily on the data processing order. According to the experimental setting presented in the paper, the algorithm goes through a single pass over the data stream, yet still able to reasonably train (deep) neural networks and identify mixture components jointly. Do you have any intuition about how this becomes feasible?\\n\\nI don\\u2019t fully understand why generative modeling is required. In page 4 the authors stated that the generative model prevents catastrophic forgetting. But in my understanding, using expert-specific parameters is the part that prevents catastrophic forgetting, not the generative model itself. Learning generative model in online fashion may work well in simple structured data such as MNIST, but I highly doubt that the generative model could be trained properly for CIFAR10 or CIFAR100, especially in online setting. My concern is that learning generative model part may even impede the discriminative learning. Could you elaborate more on this?\\n\\nAnother minor concern is the way the concentration parameter alpha is selected. The authors stated that they chose proper value of alpha according to the number of tasks known in advance. I think this does not make sense. Alpha should also be inferred along with other parameters, or fixed to non-informative value if the performance of the algorithm is not very sensitive to the choice of alpha.\\n\\nI think it would be more helpful to show how the task-assignment p(z=k|x) is learned. For instance, the clustering accuracy according to p(z=k|x) against the ground-truth task label can be measured, or at least qualitatively show what examples were assigned to each task.\"}" ] }

[ "Antonio Loquercio", "Alexey Dosovitskiy", "Davide Scaramuzza" ]

[ "Depth Perception", "Learning from Sparse Supervision", "Learning from Interaction." ]

{ "note_id": [ "gUuumxFCUS", "Sye0k_k8sH", "rkeMz0AHsB", "HyemLN24jr", "H1gVMVLVoB", "S1ldszjk9B", "Skgsjr36Kr", "H1l1jzoOtB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732762, 1573414885618, 1573412361743, 1573336139047, 1573311499601, 1571955359767, 1571829154755, 1571496599137 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1801/Authors" ], [ "ICLR.cc/2020/Conference/Paper1801/Authors" ], [ "ICLR.cc/2020/Conference/Paper1801/Authors" ], [ "ICLR.cc/2020/Conference/Paper1801/Authors" ], [ "ICLR.cc/2020/Conference/Paper1801/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1801/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1801/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes a deep network architecture for learning to predict depth from images with sparsely depth-labeled pixels.\\n\\nThis paper was subject to some discussion, since the authors felt that the approach was interesting and the problem-well motivated. Some of the concerns about experimental evaluation (especially from R1) were resolved due the author's rebuttal, but ultimately the reviewers felt the paper was not yet ready for publication.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Summary of changes to the manuscript.\", \"comment\": \"We thank all reviewers for their feedback, which helped us to improve our work. In the following, a summary of the manuscript\\u2019s revisions.\\n\\nSection 4.2: Added details about the training setup for ours and the baselines.\", \"table_2\": \"Added the baseline requested by Reviewer 2 (FCRN).\", \"section_5\": \"Added a discussion on failure cases.\\nSection 6.2.1 (Appendix): Added the details of the baseline tuning process we did to ensure a fair comparison with our method.\\nSection 6.5 (Appendix): Added an evaluation of fine-tuning correspondences with sparse depth loss, requested by Reviewer 2 and Reviewer 3.\\nSection 6.6 (Appendix): Added a visualization of the filters learned by the local network, requested by Reviewer 3.\"}", "{\"title\": \"Answer to AnonReviewer2\", \"comment\": \"We thank the reviewer for the detailed comments. We address the raised concerns below. We hope these clarifications help better understanding our contribution and experiment design and we are looking forward to hearing back from the reviewer.\\n\\nWe respectfully disagree that our architecture is very similar to the one proposed in DeMoN [25]. Here are the main differences:\\nWe predict the convolutional filters in an input dependent way to transform local matches in depth. There is no similar network in the literature to the best of our knowledge.\\nWe do not explicitly predict camera pose, but a hidden representation of it.\\nWe do not iteratively transform flow to depth using geometry equations.\\n\\nSince the reviewer might have slightly misunderstood our problem setup, we briefly summarize it here. We are interested in the problem of learning depth from very sparse supervision signal (available only at training time!), similar to\\u201chaptic feedback\\u201d, without explicitly using geometry. Our evaluation was not designed to thoroughly show the advantage over state-of-the-art methods, but to fairly study the new problem setting. In order to do so, we compared to a set of well known architectures for depth estimation, which we tuned to reach the best performance on our task (see Section 6.2.1 in the Appendix). Since our method and the baselines are tuned for the task and trained on exactly the same data, we believe that our evaluation is fair: any advantage our method has is due to the architecture, not the training regime. We have modified section 4.2 to clarify this point.\", \"the_experiment_with_varying_intrinsics_was_designed_with_the_explicit_goal_of_showing_that_state_of_the_art_unsupervised_depth_estimation_approaches_strongly_depend_on_the_camera_calibration_parameters\": \"when intrinsics are precisely know they can be very good, but are brittle otherwise. Even though concurrent unsupervised approaches can learn intrinsics together with depth [2], they still require intrinsics to be constant over time. Our method, in contrast, can adapt to instantaneous changes in camera parameters. Why is this evaluation not fair?\\n\\nAs requested by the reviewer, we added two additional baselines:\\nPWC-Net (Table 8, Appendix): We added a convolutional head on top of the off-the-shelf PWC-Net architecture and finetuned the entire system with only the very sparse depth loss. We found its performance to be generally poor, mainly due to overfitting. Details could be found in the Appendix in Table 8.\\nThe FCRN architecture from [1] (Table 2, main manuscript): In order to make a fair comparison, we tuned this architecture for our task of learning from sparse supervision. Specifically, we changed its input layer to provide it with the same input as ours: a pair of frames and the optical flow. We also had to change the activation function of the network from Relu to Leaky Relu to have better performance (details in the appendix in section 6.2.1). We then trained this baseline with exactly the same input as ours, and we found its performance to be lower than our method, but similar to our DispNet baseline [48, manuscript].\\n\\n[1] Laina, Iro, et al. \\\"Deeper depth prediction with fully convolutional residual networks.\\\" 2016 Fourth international conference on 3D vision (3DV). IEEE, 2016.\\n[2] Gordon, Ariel, et al. \\\"Depth from Videos in the Wild: Unsupervised Monocular Depth Learning from Unknown Cameras.\\\" arXiv preprint arXiv:1904.04998 (2019).\"}", "{\"title\": \"Answer to AnonReviewer3\", \"comment\": \"We thank the reviewer for the positive feedback about our work.\\n\\nAs requested by the reviewer, we have improved the description of our architecture in the methodology section. More details are provided in the Appendix in Section 6.2. To maximize reproducibility, we will publish our code upon acceptance.\\n\\nAs the reviewer is interested in the relationship between outliers in optical flow and estimated depth, we invite the reviewer to check out Fig. 5 in the Appendix. This figure experimentally supports the claim that that our approach is more robust to outliers than methods based on projective geometry. \\n\\nUpon reviewer recommendations, we have added the results of fine-tuning the optical flow module in Table 8 in the Appendix. This methodology gave good results, but did not improve on our previous approach of keeping them fixed throughout training. The reason is that the flow module is too large to be effectively trained with the very sparse depth loss, and ends up overfitting to the observed points. We have validated this claim with additional experiments in which we fine-tune PWCNet end-to-end with only the very sparse depth loss.\\n\\nIn addition, we added a visualization of the filters generated by the local network in Fig.9 of the Appendix. Note that these filters are different for every image pair. However, we discovered that when the relative transformation between images is similar, filters also tend to be similar (first and second row of Fig. 9). In contrast, when the relative transformation between views is completely different, filters tend to acquire a dissimilar pattern (first and third row of Fig. 9).\\n\\nFinally, we have added a brief summary of failure cases in the conclusion section. In general, our approach suffers from outliers in the optical flow field and by small baseline between views. In future work, we plan to alleviate this problem by fine tuning the optical flow module with the photo-consistency loss and exploiting monocular cues.\"}", "{\"title\": \"Answer to AnonReviewer1\", \"comment\": \"We thank the reviewer for the detailed review and try to address the raised concerns below. We hope these clarifications can help better understanding our experiment design and we are looking forward to hearing back from the reviewer.\\n\\nWe believe that the reviewer might have misunderstood the motivation of our paper, so we briefly restate it here. Existing learning-based depth estimation methods are trained either with a large amount of labeled data or using geometry. Biological agents, instead, acquire 3D awareness by directly interacting with the environment, mainly with haptic feedback. The main research question driving our work is to figure out if artificial agents could also learn from such limited supervision signal,without explicitly using geometry. Investigation of this scientific question may also bring practical benefits: such an approach can be easy to use and robust in applications where appropriate ground truth is available. With this work, we are making the first steps in investigating these questions.\\n\\nThus, the main goal of our evaluation is not to thoroughly show the advantage over state-of-the-art methods, but to fairly study the new problem setting. To this end, we compare our proposed architecture to several baselines based on the well-known Eigen and DispNet architectures. Note that our architecture is not provided with any privileged information with respect to these baselines: all of them get access to optical flow, so the comparison is fair. Additionally, we compare to the current state-of-the-art unsupervised depth estimation method [9], which outperforms other methods mentioned by the reviewer (i.e. Huangying et al. CVPR-18, Anurag at al. CVPR-19, Chaoyang et al. CVPR-19) by a large margin on the KITTI benchmark. We do not compare to the rest of the vast literature on unsupervised depth estimation because those methods have only been applied in a driving setting, which is completely different to our problem setup, and thus each method would require careful and time-consuming tuning to ensure a fair comparison. \\n\\n\\nWe respectfully disagree with the statement that \\u201chaptic measurements probably have large errors\\u201d. Robots (e.g. manipulators) can achieve millimeter accuracies in haptic sensing with the latest technical developments. Haptic sensing has indeed been used both in robotics [1,2] and augmented / virtual reality applications [3]. In addition, neurophysiology studies indicate that such sensing is indeed very fine for biological agents as well. For example, humans can perceive up to 2.5 degrees in rotation and 3.6mm in translation of the objects they are manipulating [4, Section \\u201cQuantifying Haptic Sensitivity\\u201d].\\n\\nThanks for spotting the typo on \\u201c179% error reduction\\u201d. We will correct it in the updated version of the manuscript.\\n\\n[1] Marchal-Crespo, Laura, et al. \\\"Haptic error modulation outperforms visual error amplification when learning a modified gait pattern.\\\" Frontiers in neuroscience 13 (2019): 61.\\n[2] Lew, Thomas, et al. \\\"Contact Inertial Odometry: Collisions are your Friend.\\\" arXiv preprint arXiv:1909.00079 (2019).\\n[3] Robles-De-La-Torre, Gabriel. \\\"The importance of the sense of touch in virtual and real environments.\\\" Ieee Multimedia 13.3 (2006): 24-30.\\n[4] Henriques, Denise YP, and John F. Soechting. \\\"Approaches to the study of haptic sensing.\\\" Journal of Neurophysiology 93.6 (2005): 3036-3043.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper provides a depth learning architecture (global-local structure) from two images. It claims that SoTA depth can be estimated from the supervision of a very sparse ground truth by leveraging the optical flow information between two images. In the experiments, it shows superior performance than other baseline methods such as Eigen's network and DispNet.\", \"pros\": \"\", \"1\": \"The claim of robustness to camera intrinsics is not solved in principle but due to training using ground truth from multiple dataset . It still suffer from depth motion confusion if there is no ground truth depth guidance when testing. It is also unfair for comparison of this metric with unsupervised approach where a universal intrinsic is assumed.\", \"2\": \"I think the comparison might not be fair since the baselines are all single image estimation networks, while the approach has two images, where disparities are serving as a strong cue for depth. Other possible architectures such as flow net, pwc net, DeMoN[25] and stereo networks such as (gc-net or psm-net) should be considered since these networks are more focus on feature matching.\", \"cons\": \"1`: It is a fairly standard network design similar to DeMoN[25], where motion network for global pose and local dense network for local matching.\", \"3\": \"Even for single image network, eigen's method is not SoTA, the author may consider (1) as one of the baseline, etc.\\n\\n(1) \\\"Deeper Depth Prediction with Fully Convolutional Residual Networks\\\"\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposed a novel global-local network, which can be trained with extremely sparse ground truth, to predict dense depth. Though widely applied on the task of segmentation, the use of only uncalibrated input and extremely sparse label for depth estimation is novel. By incorporating optical flow and decoupling global and local modules, this pipeline aligns well with disparity-based geometric methods. The paper is generally well-written and easy to follow.\", \"here_are_some_concerns\": \"Some important details are missing in order to reproduce the results. For example, the forward pipelines of the global module and the local module are covered with only few sentences. The authors should expand the section of methodology and give better formulations of their pipeline.\\n\\nHow much does the quality of input optical flow affect the results? Also, the author claimed \\u2018the network has the opportunity to correct for small inaccuracies or outliners in the optical flow input\\u2019, could you show any result related to this claim? (e.g. flawed optical flow but good output)\\n\\nCan you try to incorporate the optical flow module and fine-tune for better results? \\n\\nAre there some failure cases, and some visualizations on the filter banks of the local network?\\n\\n=========================================================\", \"after_rebuttal\": \"I thank the author for the response. \\n\\nSince there is a possibly more realistic setting that we train the model with dense depth annotations we currently have and use very sparse annotations to adaptive to new environments where we are hard to gain dense annotations. Also, the optical flow module relies on extra annotations to train. I will keep my origin scores.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper presents a method for learning to predict a depth map given a pair of images. Training is done is an unsupervised way except for a small set of image locations for which the absolute 3D locations are known. The optical flow between the 2 images is computed automatically. A network is trained to predict the camera motion between the 2 images from the images and the optical flow. From this global motion, convolutional filters are predicted, in order to transform the optical flow into a depth map.\\n\\nThe use of image locations with known 3D locations is motivated in the paper to \\\"simulate\\\" the knowledge of depth coming from a haptic system.\\n\\nI have several concerns about this paper.\\n\\n* My main concern is that the paper compares against only a few papers (1 from 2014, 1 from 2016, and one more recent from 2019) while the literature is extremely vast, and visually, the results seem far from the state-of-the art. See for example:\\n\\nHuangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. In CVPR, 2018.\\n\\nAnurag Ranjan, Varun Jampani, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In CVPR, 2019.\\n\\nChaoyang Wang, Jose Miguel Buenaposada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using\\ndirect methods. In CVPR, 2018.\\n\\nIn fact, these papers have simpler requirements as they do not need 2 images at run-time, nor 3d data at training time. I thought for a moment that the advantage of the paper was to be able to predict an absolute motion and an absolute depth (this is not impossible if the method is able to estimate the scale from known objects, as it is suggested from the introduction). This was however incorrect as the text says in the middle of Section 4.1 \\\" we resolve the inherent scale ambiguity by normalizing the depth values such that the norm of the translation vector between the two views is equal to 1\\\".\\n\\n* The network used to predict optical flow was trained with a large amount of supervised data. As image matching is the most difficult task, it is difficult to claim that the method is unsupervised.\\n\\n* The motivation for having sparse measurements is to \\\"simulate\\\" haptic. This is fine for me in principle, however haptic measurements would probably have large errors, while it seems that the experiments use ground truth values for these points.\", \"one_minor_remark\": \"End of Section 4.4: What is a \\\"179% error reduction\\\"? How is computed the percentage?\"}" ] }

[ "Ian Fischer", "Alex A. Alemi" ]

[ "Information Theory", "Adversarial Robustness" ]

{ "note_id": [ "7OztopXgxO", "rJxEZfH3iH", "BJluZv0jiH", "ryx9RUAiiH", "HJx49U0sjB", "B1e4B8Assr", "SJlrME_6Yr", "B1xkJ3EhtB", "r1gbwXWiFr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732733, 1573831164452, 1573803776107, 1573803730361, 1573803660173, 1573803580000, 1571812364775, 1571732438854, 1571652440543 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1800/Authors" ], [ "ICLR.cc/2020/Conference/Paper1800/Authors" ], [ "ICLR.cc/2020/Conference/Paper1800/Authors" ], [ "ICLR.cc/2020/Conference/Paper1800/Authors" ], [ "ICLR.cc/2020/Conference/Paper1800/Authors" ], [ "ICLR.cc/2020/Conference/Paper1800/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1800/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1800/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes CEB, Conditional Entropy Bottleneck, as a way to improves the robustness of a model against adversarial attacks and noisy data. The model is tested empirically using several experiments and various datasets.\\n\\nWe appreciate the authors for submitting the paper to ICLR and providing detailed responses to the reviewers' comments and concerns. After the initial reviews and rebuttal, we had extensive discussions to judge whether the contributions are clear and sufficient for publication. In particular, we discussed the overlap with a previous (arXiv) paper and decided that the overlap should not be considered because it is not published at a conference or journal. Plus the paper makes additional contributions.\\n\\nHowever, reviewers in the end did not think the paper showed sufficient explanation and proof of why and how this model works, and whether this approach improves upon other state-of-the-art adversarial defense approaches.\\n\\nAgain, thank you for submitting to ICLR, and I hope to see an improved version in a future publication.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Adding public comment for the missing reference in the review and reply\", \"comment\": \"The reference [A2] in the review and reply above is:\\n\\n[A2] Zhang et al., Defending against Whitebox Adversarial Attacks via Randomized Discretization, In AISTATS, 2019.\"}", "{\"title\": \"Response to Reviewer2\", \"comment\": \"Thank you for your review. We have addressed your minor comments in the updated text. We agree that the early figures are difficult to read, and have improved them in the updated version. We appreciate that you found the presentation clear, the approach original, and the benefit of CEB to model robustness of interest in itself. We agree that a careful comparison of the effects of regularization techniques on robustness would be interesting for future work. We note that we implicitly have some comparisons in the paper, in that the ImageNet models are all trained with both L$_2$ weight normalization and BatchNorm, neither of which confers substantial robustness to the average case, and essentially no robustness to the worse case, when used without CEB.\"}", "{\"title\": \"Response to Reviewer3\", \"comment\": \"Thank you for your review. We address your major concerns below.\\n\\nAdversarial Training.\\nNone of our models is adversarially trained. Only the Madry model (which we compare to as a baseline for robustness on CIFAR10) is adversarially trained. We agree that this is a surprising result, and hope that other researchers will find it as exciting as we do, since we additionally demonstrated that CEB improves clean test accuracy as well, while the adversarial training techniques proposed so far are all harmful to clean test accuracy.\\n\\nAdversarially-Trained Robustness to L$_2$ Attacks.\\nYes, the result is correct. The baseline deterministic model is **substantially more** robust to the PGD L$_2$ attack than the adversarially-trained model from Madry (2017). We conferred with Madry extensively to ensure that we were measuring everything in exactly the same way for Figure 3. While we did not write this paper to show that adversarial training fails to generalize to other attacks, we consider this to be a very important result for the research community on adversarial robustness. We discuss that fact in the caption for Figure 2 (which also shows the same result) and in the Conclusion. Based on your suggestion, we additionally are mentioning this result in the Introduction.\\n\\nAlternative Attacks.\\nThe transfer attacks shown for the large CEB model is an example of a blackbox attack. They show that the large CEB model is even more robust to PGD attacks generated by both the less robust large CEB model and the adversarially trained model from Madry (2017). We may have time to add other attacks to the appendix if the paper is accepted, but we note that PGD is the standard attack used in the literature, due to its flexibility, relative speed, and effectiveness. We note here that preliminarily we have tested the large CIFAR10 CEB models on all of the attacks in CleverHans. CEB helped in all instances, but we chose to focus on the PGD style attacks for the paper to try to tell a compelling story. \\n\\nTraining with AutoAugment.\\nWe mention in the Introduction that good data augmentation policies are the one currently-known way to improve average-case robustness (i.e., robustness to perturbations like those found in the Common Corruptions benchmark). This is the core justification for the use of AutoAugment. There are additional pragmatic reasons for its use. For example, for CIFAR10 models, some augmentation policy is required to get good accuracy for all models we are aware of. The standard set of augmentations was shown to be inferior to the learned policies from AutoAugment, so there seems to be little point in using the standard augmentations. Furthermore, we are comparing with adversarial training, which (as we mention in the paper) can be viewed as a particular choice of data augmentation policy. Thus, comparing to models trained with AutoAugment seems reasonable. Finally, we wanted to demonstrate (with the large CEB models) that it was possible to exceed the previous SOTA accuracy on CIFAR10 for a WideResNet model that was presented in the AutoAugment paper. A core point of the experiments is to show in multiple settings that CEB not only improves the robustness of the model, it also improves classical generalization performance. Outperforming the previous SOTA results makes that claim much stronger. We discuss those comparisons in detail in Section 3.2.\\n\\nClarity of Figure 3.\\nWe agree that it is difficult to interpret Figure 3 in the current version. We have improved it in the new version and will continue to work on its legibility for the camera ready.\\n\\nOverfitting of $\\\\rho$.\\nWhile the main point of the paper was to demonstrate the effect of $\\\\rho$ on robustness, in order to clarify and clean up our argument, based on your question we have modified the results to be the ones achieved by cross validation. I.e. we evaluate the robustness on a validation set, choose the rho that is best on the validation set and then report the test set performance, where the test set is the set of perturbations: Common Corruptions, Imagenet-A, and targeted PGD attacks. We suggest using cross validation in practice and have emphasized this in the text of the paper as well.\\n\\nQuestions about the Impact of BatchNorm.\\nWe agree that the empirical fact that BatchNorm seems to harm robustness of CIFAR10 models is intriguing. Note that we do not disable BatchNorm in the ImageNet models and still show real robustness improvements on the common corruptions benchmark as well as on targeted PGD attacks in the updated version of the paper.\"}", "{\"title\": \"Response to Reviewer1\", \"comment\": \"Thank you for the detailed review. We respond to each of your concerns below.\\n\\nOverlap with Fischer (2018).\\nWe believe our paper offers novel contributions on top of Fischer (2018) and tried to fairly attribute it as necessary (despite it being unpublished work). In particular, this paper offers a unique rederivation and justification of the CEB objective as well as a novel set of experiments. In this work, we used a simpler invocation of the CEB objective and applied to larger datasets (Imagenet vs FashionMNIST). Additionally, all of the Common Corruptions work is novel in comparison. The major contributions of this work are that the same technique scales easily to large classification problems like ImageNet and CIFAR10, and that the robustness conferred by training using the CEB objective applies to untargeted attacks and common corruptions as well. In our updated experiments, we are additionally showing robustness improvements on ImageNet models against ImageNet-A [1] and L$_2$ and L$_\\\\infty$ targeted PGD attacks. \\n\\nComparisons with other adversarial defenses.\\nThank you for the suggestion. In the context of this paper, we aim to demonstrate not that CEB is the best adversarial defense, but that it offers some substantial defense, which is interesting given that it is a general purpose objective function. In principle this could be composed with other defense techniques and we agree that would make interesting follow up work. In particular, with regards to [A2], we'd like to emphasize that our CEB models do not show any decrease in their predictive performance and in fact show increased test accuracy in the regime in which they show adversarial robustness, a property that to our knowledge is unique.\\n\\nImageNet Baselines.\\nDue to feedback from the authors of [3] that the community is standardizing around reporting numbers on 224x224 Common Corruptions, in the updated draft we are switching from 299x299 Imagenet training and testing to 224x224, so the numbers in the new draft are slightly different. However, in both cases, our top-1 accuracies were well in-line with what we have seen reported in the literature for ResNet50, with and without AutoAugment. We imagine that the baselines could be improved by tuning learning rate schedules and weight decay factors carefully, but we don\\u2019t see any reason that similar tuning of the CEB models wouldn\\u2019t yield similar improvements. We have updated the appendix to mention that our implementation is closely based on the open-source Cloud TPU ResNet50 model from [2], and also to include a minimal set of modifications to that model to get the same basic CEB training setting that we used for the ResNet50 models without AutoAugment in this paper. We stuck with the ResNet models since many of our experimental results are meant to compare to other published Common Corruptions results, which are given largely for ResNet50 style models. Are there particular architectures you were hoping to see compared instead?\\n\\nBatchNorm Baseline for CIFAR10.\\nThis is a good suggestion. We trained a deterministic 28x10 Wide Resnet model with BatchNorm and AutoAugment. It got 97.3% accuracy on the CIFAR10 test set, which is in-line with the 97.4% accuracy for that model reported in the AutoAugment paper. It got 0% accuracy on the baseline L$_\\\\infty$ PGD attack at $\\\\epsilon=8$ throughout training (we measured test robustness on L$_2$ and L$_\\\\infty$ PGD attacks for all of our models every two training epochs). Interestingly, final L$_2$ PGD attack accuracy was noticeably better than the deterministic baseline without BatchNorm (73% vs. 66%), though still is much worse than the best CEB models on that attack (over 80% accuracy). We have added these results to the paper.\\n\\nComparisons between VIB and CEB.\\nWe agree that a more complete comparison with VIB would be nice, but we note that we are already at the page limit. We would also like to make certain that the reviewer saw Section 3.1, and in particular the right two panels of Figure 1, where we directly compare CEB and VIB Fashion MNIST models at the same range of $\\\\rho$ values against targeted PGD L$_2$ and L$_\\\\infty$ attacks. In that setting, the CEB models dominate the VIB models at both general classification accuracy and robustness, even though the measured lower bounds on $I(X;Z)$ were almost identical for the VIB and CEB models at each value of $\\\\rho$. We think this section provides the type of comparison with VIB that you are asking for in your review. Please tell us if we have misunderstood.\\n\\n[1] Hendrycks et al. \\u201cNatural Adversarial Examples\\u201d. 2019. https://arxiv.org/abs/1907.07174.\\n[2] https://github.com/tensorflow/tpu/tree/master/models/official/resnet.\\n[3] Yin et al. \\u201cA Fourier Perspective on Model Robustness in Computer Vision\\u201d. 2019. https://arxiv.org/abs/1906.08988.\"}", "{\"title\": \"General Response\", \"comment\": [\"We have given detailed replies to reviewer questions and concerns below. Based on reviewer feedback, we have additionally updated the paper as follows:\", \"We have added ImageNet results for ResNet 152 models, to show that the same trend of larger CEB models improving robustness over their deterministic baselines, with particular dramatic effect on the new PGD attacks.\", \"We have added targeted PGD L$_2$ and L$_\\\\infty$ attacks to the ImageNet experiments, as well as results on Natural Adversarial Examples [1].\", \"We now point out in the Introduction the incidental but important contribution that the adversarially-trained CIFAR20 model of Madry (2017) failed to generalize to L$_2$ PGD attacks, having much worse performance than a deterministic baseline model of the same architecture on that attack.\", \"We have improved the legibility of Figures 1 through 3.\", \"We have added a deterministic baseline CIFAR10 model that uses BatchNorm to the comparison.\", \"We have added a code listing to the Appendix showing the core changes that are needed to turn the CloudTPU ResNet50 model [2] into a CEB model.\", \"We hope that our discussion in the replies has helped the reviewers understand the paper in much greater detail, and that the changes listed above have adequately addressed reviewer\\u2019s comments. We look forward to answering any additional questions that reviewers may have in the remaining review period.\", \"[1] Hendrycks et al. \\u201cNatural Adversarial Examples\\u201d. 2019. https://arxiv.org/abs/1907.07174.\", \"[2] https://github.com/tensorflow/tpu/tree/master/models/official/resnet.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper studied the effectiveness of Conditional Entropy Bottleneck (CEB) on improving model robustness. Three tasks are considered to demonstrate its effectiveness; generalization performance over clean test images, adversarially perturbed images, and images corrupted by various synthetic noises. The experiment results demonstrated that CEB improves the model robustness on all considered tasks over the deterministic baseline and adversarially-trained classifiers.\\n\\nThe proposed idea is simple, easy to implement, and generally applicable to various classification networks. I especially enjoyed the nice experiment results; indeed, it has been widely observed that many existing approaches on adversarial training sacrifice the classification performance to improve the robustness over adversarial examples [1]. The proposed approach sidesteps such a problem since it does not require adversarial retraining. It is surprising to see that the proposed method is still able to achieve stronger robustness than the models based on adversarial training.\\n\\nOne of my major concerns is that it has a large overlap with Fischer (2018) in terms of methodology, experiment settings, and empirical observations, which limits the general contribution of the paper. Fischer (2018) first proposed CEB objective and showed its effectiveness in various tasks, including the adversarial defense. Although this paper extends the results on adversarial defense to CIFAR 10 dataset and includes additional ablative experiments and experiments on other types of input noises, it ends up confirming very similar observations/conclusions in Fischer (2018). Although Fischer (2018) is an unpublished work, I think that it is fair to consider that as a prior work since it is properly cited in the paper. \\n\\nMy other concern is that the experiment misses comparisons against other adversarial defense approaches, which makes it difficult to understand the degree of robustness this model can achieve. The current comparisons are mostly focused on deterministic and VIB baselines, which are useful to understand the core argument of the paper, but insufficient to understand how useful CEB could be in the purpose of adversarial defense. Especially, I believe that some recent approaches that do not require adversarial training, such as [A2], are worth comparisons. \\n\\nBelow are some minor comments/questions on the paper.\\n1. Section 3.2: For this experiment, BN is removed from the classification network; it would be still beneficial to see the baseline performance with BN (deterministic model) to better compare the classification performance on clean test data.\\n2. Section 3.3: The performance on both baseline and the proposed model on clean data is far lower than the SOTA models. Some clarification would be helpful.\\n3. It would be great to see further clarifications of improvement in CEB over VIB; currently, it is very simply justified that it is because CEB optimizes tighter variational bound on Eq.(3) than VIB. But it would also be great to see justifications from various angles (e.g. in the context of adversarial defense).\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": [\"The paper modifies existing classifier architectures and training objective, in order to minimize \\\"conditional entropy bottleneck\\\" (CEB) objective, in attempts to force the representation to maximize the information bottleneck objective. Consequently, the paper claims that this CEB model improves general test accuracy and robustness against adversarial attacks and common corruptions, compared to the softmax + cross entropy counterpart. This claim is supported by experimental results on CIFAR-10 and ImageNet-C datasets.\", \"In overall, the manuscript is easy-to-follow with a clear motivation. I found the experimental results are also promising, at least for the improved test accuracy and corruption robustness. Regarding the results about adversarial robustness, however, it was really confusing for me to understand and validate the reported values. I would like to increase my score if the following questions could be addressed:\", \"It is not clear whether adversarial training is used or not in CEB models for the adversarial robustness results. If the results were achieved \\\"without\\\" adversarial training, these would be somewhat surprising for me. At the same time, however, I would want to see more thorough evaluation than the current, e.g. PGD with more n, random restart, gradient-free attacks, or black-box attacks.\", \"I wonder if the paper could provide a motivation on why the AutoAug policy is adopted when training robust model. Personally, this is one of the reasons that makes me hard to understand the values presented in the paper.\", \"Figure 3, right: Does this plot indicates that \\\"28x10 Det\\\" is much more robust than \\\"28x10 Madry\\\"? If so, it feels so awkward for me, and I hope this results could be further justified in advance.\", \"Figure 3: It is extremely difficult to understand the plots as the whole lines are interleaved on a single grid. I suggests to split the plots based on the main claims the paper want to demonstrate.\", \"How was the issue of potential over-fitting on rho handled, e.g. using a validation set?\", \"In general, I slightly feel a lack of justification on why the CEB model improves robustness. In Page 5: \\\"... Every small model we have trained with Batch Normalization enabled has had substantially worse robustness, ...\\\" - I think this line could be a critical point, and need further investigations in a manner of justifying the overall claim.\"]}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"TITLE\\nCEB Improves Model Robustness\\n\\nREVIEW SUMMARY\\nAs a thorough empirical evaluation of a new method for training deep neural networks, this paper is a useful contribution.\\n\\nPAPER SUMMARY\\nThe paper presents an empirical study of the robustness of neural network classifiers trained using a conditional entropy bottleneck regularization.\\n\\nCLARITY\\nThe paper is well written and easy to follow. Results are presented in a way that gives the reader a good overview, but figure quality / legibility could be improved.\\n\\nORIGINALITY\\nThe CEB method is an original (but minor) modification of a well known method. \\n\\nSIGNIFICANCE\\nThe primary contribution of the paper is a thorough experimental evaluation of the particular regularization method presented in the paper. While this is certainly useful, it would have been even better with more comparison with other regularization methods. The paper clearly demonstrates the benefits of CEB, which is of interest in itself.\\n\\nFURTHER COMMENTS\\nConsider making a more clear distinction between generalization and robustness.\\n\\\"are are\\\" Double word.\\n\\\"all machine-learned systems...highly vunerable\\\" Too strong statement.\\n\\\"VIB\\\" Abbreviation not defined in the text.\\nFigure 1, 2 and 3 are a bit hard to read (especially the dotted lines)\"}" ] }

[ "Alexandre Defossez", "Nicolas Usunier", "Leon Bottou", "Francis Bach" ]

[ "source separation", "audio synthesis", "deep learning" ]

{ "note_id": [ "fzDEFA7Dx", "Hyld4J2niB", "HJgSX4chjB", "HJgR9oY2jr", "S1lrCvKhsB", "rkl1c1t3sr", "SJg58JKhsH", "BJllrDO3sr", "SJlSfDunjB", "Hkl_LuUQcS", "SJxIHvP15H", "H1g-lHiCKB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732705, 1573859120059, 1573852189153, 1573850005821, 1573849037499, 1573846918766, 1573846865801, 1573844792202, 1573844748669, 1572198480182, 1571940157875, 1571890409343 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1799/Authors" ], [ "ICLR.cc/2020/Conference/Paper1799/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1799/Authors" ], [ "ICLR.cc/2020/Conference/Paper1799/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1799/Authors" ], [ "ICLR.cc/2020/Conference/Paper1799/Authors" ], [ "ICLR.cc/2020/Conference/Paper1799/Authors" ], [ "ICLR.cc/2020/Conference/Paper1799/Authors" ], [ "ICLR.cc/2020/Conference/Paper1799/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1799/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1799/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper proposed a waveform-to-waveform music source separation system. Experimental justification shows the proposed model achieved the best SDR among all the existing waveform-to-waveform models, and obtained similar performance to spectrogram based ones. The paper is clearly written and the experimental evaluation and ablation study are thorough. But the main concern is the limited novelty, it is an improvement over the existing Wave-U-Net, it added some changes to the existing model architecture for better modeling the waveform data and compared masking vs. synthesis for music source separation.\", \"title\": \"Paper Decision\"}", "{\"title\": \"SIR reply\", \"comment\": \"There is significantly more interference with Demucs, you can find the SIR on page 15 of the paper, in the appendix. SIR for Demucs is 10.39, 11.47 for Tasnet.\\nWe don't pretend that Demucs is better than Tasnet, but we support that the two have different approaches and capabilities, one being better at separation and the other keeping better naturalness. While it would be ideal to have the best of both world, we believe it is beyond the scope of this paper. We provide a solid experimental comparison of spectrogram methods, Tasnet and Demucs, both with standard source separation metrics and human evaluations. We also provide the code for the training pipeline of Demucs and Tasnet on MusDB which we are going to open source.\\nOverall, we agree with Reviewer 3 that this paper is not a revolution in the field but a solid foundation for future work by all interested parties.\"}", "{\"title\": \"SIR\", \"comment\": \"I see, but I am also hearing more interference with Demucs, which might explain lower SDRs. Do you have numbers for SIR values?\"}", "{\"title\": \"Comparison of Tasnet and Demucs\", \"comment\": \"Computational requirements are of a similar order. Tasnet requires more memory and computation for the same duration but because it only uses local information it can be trained on much smaller audio extracts, which compensates in the end.\\n\\nPlease have a listen to the audio samples included in the link. You will see that there is something like static noise on the output of Tasnet, as well as hollow attacks, especially noticeable on the drums and bass sources. We confirmed this observation with mean opinion scores evaluating both the quality (i.e. absence of artifacts) and purity (absence of contamination from other sources) of the separated audio. Tasnet has slightly better purity (3.42 against 3.30 for Demucs) but Demucs has significantly higher quality (3.22 against 2.85).\\nThose results are detailed in Table 2 and 3 in the new version of the paper.\"}", "{\"title\": \"Thanks for the honesty and the effort\", \"comment\": \"Thanks for including the Tasnet result, and showing the results transparently. I am seeing that the models with your model are in the same ballpark as Tasnet, do we have any advantage with the proposed model compared to Tasnet? How do the computational requirements compare?\"}", "{\"title\": \"Reply to Reviewer 3\", \"comment\": \"We thank the reviewer for the constructive and positive review.\\n\\nFor the discussion of \\\"upsamplied convolution\\\" vs \\\"transposed convolution\\\", We clarify that what we mean by \\\"upsample-convolution\\\" is a (e.g., bilinear) upsampling method followed by a non-strrided convolution vs a strides .\\n\\nWe also made significant updates of the paper (please see the general comments for a summary of the main changes).\"}", "{\"title\": \"Reply to reviewer 2\", \"comment\": \"We thank the reviewer for the constructive comments. The reviewer's main points were:\\n\\n1- \\\"it is not clear to me that the conclusions drawn based on the study in this work will be generally applicable to other related tasks\\\"\", \"we_argue_that_there_are_two_contributions_in_that_paper\": \"a) showing that now the state-of-the-art is on multi-instrument source separation is obtained by wave-form-to-waveform models (see general comments for updated results)\\nb) showing that approaches based on convolutional encoders/decoders are effective for source separation tasks. Of course, experiments on other source sepraration problems with similar architectures would be a big plus, but it is out of the scope of the current paper and we leave such studies as future work.\\n\\n2-\\\"The section describing the evaluation metric SDR wasn\\u2019t very clear to me\\\"\\n \\nWe rephrased it and followed the reviewer suggestion to refer to Vincent et al. and only explain what the metric is supposed to capture\\n\\n3- multiplying by +1 or -1 corresponds to a global shift of phase. we use that as a form of data augmentation, since the model should be equivalent to such a shift\\n\\nWe also updated the paper to correct the reviewer's minor comments.\"}", "{\"title\": \"Reply to Reviewer 1\", \"comment\": \"We thank the reviewer for pointing out the Tasnet paper. We made experiments with Tasnet and updated the paper accordingly. Please, see the general comments for more details.\\n\\nWe also improved the model performance (mostly the inference procedure that takes an average over random initial shifts of the input). The results of Demucs (and Tasnet) are now very significantly better than the best models operating on spectrograms (+0.4 SDR points for Tasnet and +0.3 SDR for Demucs).\"}", "{\"title\": \"general comment: comparison to Tasnet and update of the paper\", \"comment\": \"We first would like to thank the reviewers for their constructive comments. We made significant updates to the paper to respond to the reviews:\\n\\n1- We performed several improvements to the model, and updated the corresponding empirical study (increased the number of channels, changed the batch size, and updated the inference procedure using an average over random initial shifts of the song (see Section 4.5). The results are now of 5.6 SDR, while the best previous results on MusDB are 5.3 (see the updated Table 1). You might notice that the validation loss has increased between the two versions, this is because we initially used pytorch DistributedSampler which repeats some of the tracks to make the dataset a multiple of the number of GPUs. This biased the validation loss in the first version of the paper and we fixed that since.\\n\\n2- We added experiments with Tasnet, as suggested by reviewer 1. Because Tasnet was developed for monophonic source separation with a sampling rate of 8kHz, we had to adapt it to make it work on stereo sampled at 44kHz (which requires increasing the receptive field of the overall network to obtain reasonable performances). With these modifications (described in a new Section 3), Tasnet obtains state-of-the-art SDR of 5.7, slightly better than Demucs. The difference diminishes when adding more training data, as they both obtain 6.3 SDR with 150 additional songs, whereas the previously best reported result on MusDB test set was 6 with 800 additional songs.\\n\\n3- We performed additional human evaluations between Tasnet and Demucs (trained without additional data) in terms of quality and contamination with other sources. The lower SDR for Demucs is explained by a bit more contamination. On the other hand, Demucs obtains significantly better results in terms of quality. \\n\\n4- We updated the paper overall to reorganize the discussion on the differences between Demucs and Tasnet instead of the differences between Demucs and Wave-U-Net.\\n\\n5 - We added the source code to the supplementary material. Pre-trained models are also available for evaluation and can be downloaded on demand from the demucs.separate script. All instructions the README file in the code folder in the ICLR code link. The link will download both the code and audio samples for Demucs, Conv-Tasnet and all the baselines. The entire supplementary material is 100MB and might take a few minutes to download.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"In this work, the authors consider the task of supervised music source separation, i.e., separating out the components (bass, drums, voice, other) out of a mixed music track. In particular, the work considers the task of supervised source separation where the individual target tracks are available during training time. The main contribution of this work is the improvement of an end-to-end waveform-to-waveform separation models through a number of architectural changes that allow such waveform-to-waveform models to perform comparably with other current state-of-the-art methods that instead operate in the spectrogram domain.\\n\\nOverall, the paper is generally well written and the method is easy to follow. However, I have a couple of concerns about this work, based on which I would rate the work as \\u201cweak reject\\u201d:\\n\\n1. The specific architecture proposed in this task does appear to improve performance for this particular task, but it is not clear to me that the conclusions drawn based on the study in this work will be generally applicable to other related tasks. Also, the final model is still only comparable to the spectrogram-based methods overall, and appears to do a slightly worse job in separating the \\u2018vocal\\u2019 and \\u2018other\\u2019 tracks than these baseline methods. As such, I\\u2019m unsure about the impact of this work.\\n2. The section describing the evaluation metric SDR wasn\\u2019t very clear to me. Perhaps it would be better to just refer back to (Vincent et al., 06) and just describe what SDR captures instead? Or alternatively, more details could be added to explain the computation more clearly.\\n3. Section 4.2: The authors mention that they multiply each source by +/- 1. I didn\\u2019t follow why this is being done. Could the authors please clarify.\", \"minor_comments\": \"1. The notation in Equation 1 is slightly confusing because \\u2018s\\u2019 is used to index both the mixture and the sources. I think x \\\\in \\\\mathcal{D} would be clearer. Similarly I wonder if x_s \\\\in \\\\mathbb{R}^C \\\\times \\\\mathbb{R}^T would be clearer than x_s \\\\in \\\\mathbb{R}^{C,T}\\n2. The reference (Oord et al., 2017) should be rendered as (van den Oord et al., 2017) \\n3. The Figure uses K, and S to denote Kernel Width and stride, but this isn\\u2019t explicitly mentioned in the text. Perhaps it would be useful, in Section 3.1 to write: \\u201c... convolution with kernel width, K=8, and stride, S=4, ...\\u201d or something similar, and updating the caption to explain the notation.\\n4. Section 3.2: \\u201cThe decoder is almost the symmetric of the encoder\\u201d --> \\u201cThe decoder is almost the inverse of the encoder\\u201d would perhaps be better?\\n5. Section 3.2: \\u201cThe final layer \\u2026 S.C_0 ...\\u201d --> \\u201cThe final layer \\u2026 S * C_0 ...\\u201d\\n6. The notation in Equation 1 and Equation 2 is slightly inconsistent (lower-case vs. upper-case L for the loss function)\\n7. Section 4.3: \\u201cadditionnaly\\u201d --> \\u201cadditionally\\u201d; \\u201c32 GB or RAM\\u201d --> \\u201c32 GB of RAM\\u201d\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper suggests using a Unet type architecture to perform end-to-end source separation. They are reporting performance improvement over another architecture which uses Unets architecture in conjunction with Wavenet decoders. They also report a very marginal performance improvement over an STFT based model (open unmix) The improment is 0.02 dB (table 1), and I am not sure if it is statistically significant.\\n\\n\\nAlso, there is no mention of algorithms which adaptively learn the basis and then do masking similar to what we do in the STFT domain. A very popular example for this is tasnet, which performs well on speech source separation tasks. I would like to see comparisons with this model. If you think there is a specific reason why not to use adaptive basis approaches such as TASNET, please do let me know.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"The authors present modifications to the state of the art (waveform based) source separation model (Wave-U-Net) and improve the state of the art to be comparable to spectral masking based methods. They clearly outline all the architectural changes they make (GLU nonlinearities, strided upsampling, bidirectional RNN), perform thorough evaluations against strong baselines, and a complete ablation study to demonstrate the value of each component. The paper reads well and quickly informs newcomers to the field with proper motivation and context. For all these reasons I think it should be accepted. One weakness of the paper is that the relative incremental nature of the study, but given the thoroughness and clarity of the experiments, and the non-triviality of the architecture search, I think this is a valuable contribution for the ICLR community.\\n\\nOne small suggestion, the language difference between \\\"upsampled\\\" convolution and \\\"transposed\\\" convolution is a bit confusing. I might suggest focusing on the striding of the convolution vs. bilinear upsampling, as a non-strided convolution can still be \\\"transposed\\\" from a standard API point of view in pytorch or tensorflow.\"}" ] }

[ "Yuanhao Wang*", "Guodong Zhang*", "Jimmy Ba" ]

[ "minimax optimization", "smooth differentiable games", "local convergence", "generative adversarial networks", "optimization" ]

{ "note_id": [ "FBYDmXnMSN", "BylKxgP3oH", "BJxZKhUnoB", "S1lLhU1hsB", "SJlGh5Aoor", "r1g7wOsooH", "rke54KqoiH", "BJx9afqiiB", "SJxTI6tior", "SJeHw7FsiH", "rylHGhdjjH", "SJlg77SioH", "B1eWY0Voir", "rkls3yVosH", "HyxYimzssH", "SJePYIDciH", "SklQELDcjB", "BJlOStb5ir", "H1gvHQuxir", "ryxdoBDeoS", "ByxHav0yoB", "S1gN8fC1jS", "r1x_rSVkqr", "r1xED6s2YB", "r1g-eWsqKH", "r1lnJfUHtr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment" ], "note_created": [ 1576798732677, 1573838833014, 1573837944657, 1573807789789, 1573804713649, 1573791835455, 1573787953705, 1573786305677, 1573784917501, 1573782364948, 1573780492524, 1573765912031, 1573764729437, 1573760946606, 1573753761123, 1573709439242, 1573709355108, 1573685568055, 1573057342827, 1573053855641, 1573017533199, 1573016140145, 1571927359601, 1571761499627, 1571627241056, 1571279331911 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1797/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ], [ "ICLR.cc/2020/Conference/Paper1797/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1797/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1797/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1797/Authors" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The submission proposes a novel solution for minimax optimization which has strong theoretical and empirical results as well as broad relevance for the community. The approach, Follow-the-Ridge, has theoretical guarantees and is compatible with preconditioning and momentum optimization strategies.\\n\\nThe paper is well-written and the authors engaged in a lengthy discussion with the reviewers, leading to a clearer understanding of the paper for all. The reviews all recommend acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Further Response\", \"comment\": \"Thank you for all the comments and suggestions.\\n\\nIn terms of our debate around the right notion of game underlying GANs, we agree that this issue is controversial, so we will modify our paper as it's not the main focus of our paper. We have already toned down our claim in the related work section in the current revision. We were spending most of our time discussing with reviewer #1 and didn't get enough time to go through the paper to modify all those statements, and are really really sorry about that. We promise that we will update the paper according to your suggestions in the camera ready version should our paper be accepted.\"}", "{\"title\": \"Final BEST attempt. We updated the statement around Theorem 3.\", \"comment\": \"On the one hand, we thank you for all detailed feedback and comments. We really appreciate your consistent dialogue with us. Your comments definitely helped make our paper stronger. On the other hand, we are a bit frustrated and upset. We believe we have addressed all your concerns so far, especially your concern about the validity of Theorem 3. It's a bit unfair to us since we cannot further participate in the reviewer discussion period. However, we won't give up and would like to give it our final best try. Hope we can reach consensus before rebuttal deadline.\\n\\nIn your last response, you raised concern with that fact that we used $y = r(x)$ in our proof of Theorem 3. We would clarify that in the proof of Theorem 3, we are indeed just calculating the Jacobian at a fixed point (we can therefore safely use $y = r(x)$, see the statement before the calculation of Jacobian in page 16). Theorem 3 itself only says about stability of fixed points, which is a proxy for convergence. For local convergence, we have to combine it with Proposition 4 (see the newly added Remark 2). We apologize that we didn't refer to Proposition 4 in previous version, we've updated the paper to make this connection explicit.\\n\\nWe stress again that the key is that we are allowed to do Taylor expansion around the fixed point (see Proposition 4). Therefore, we only need to calculate the Jacobian matrix for fixed points in Theorem 3. The combination of Theorem 3 and Proposition 4 implies local convergence (see Remark 2).\\n\\nWe believe our paper deserves a score of 6 or even higher given our theoretical and empirical contributions. We've resolved all your concerns so far, and we hope you can reconsider your rating.\"}", "{\"title\": \"The main concern has been addressed.\", \"comment\": \"Thank you for your response and valuable input over the discussion period. We believe the discussion has benefited both of us. However, we respectfully disagree with your assessment of our general-sum results. In particular, your main concern of our result is already answered above.\\n\\n- Regarding why Theorem 3 uses $y=r(x)$, but we can reason about dynamics (8)\\n\\nWe believe we have answered this question already in the paragraph \\\"Connecting the result of Theorem 3 to our local convergence claim\\\". Essentially, we can do this is because we can use the first-order Taylor series at the fixed point to characterize the behavior of the algorithm in the neighborhood of the fixed point. Thus, we only need to compute the Jacobian at fixed points in Theorem 3. We would like to know exactly which part of this argument you find not convincing.\\n\\n- Regarding the statement of Theorem 3:\\n\\nWe believe our statement of Theorem 3 is clear. The stability of fixed points is defined in Definition 2, and our requirement of the learning rate is given explicitly in the proof. Should you find the statement of Theorem 3 vague, we would like to ask you to point out exactly what requires further clarification.\\n\\nWe would like to ask you to point out what exactly is wrong *in our reasoning*. We also think that stating your doubts clearly and exactly would greatly benefit the following reviewer discussion period.\"}", "{\"title\": \"Some Concerns Remain; Will leave it up to the AC\", \"comment\": \"Thanks for the response.\\n\\nI want to reiterate that I think this paper studies an interesting problem, looks at an interesting set of dynamics, and has some impressive empirical results.\\n\\nHowever, in my opinion, either there is some problems with the results that are given for general-sum games or the assumptions and the result are not stated rigorously enough. It seems to me you are drawing conclusions about a system using y_t iterates which are not the same as r(x) and you have not shown that r(x) is attracting under the y_t dynamics surely. I can also try to say this another way. In the proof of theorem 3, you effectively plug in r(x) in the first line and after applying the similarity transform, in the upper left block, it appears you take the total derivative, as if y=r(x), to be able to get back the condition in prop 8, but you have not shown that y=r(x) under the FR dynamics. I think that under the right assumptions you could get the result you have, but it has been left up to some question since theorem 3 is stated in a vague manner.\\n\\nAgain, thanks for being willing to have a cordial discussion. I hope that the prolonged back and forth has helped make your paper stronger in your view. For now I have left my score as is, and am willing to change during the reviewer discussion period if there is further evidence that the conclusion in the paper is fine.\"}", "{\"title\": \"More clarifications. FR is sort of a replacement of timescale separation.\", \"comment\": \"As we notice in your last response, it seems that you are still reasoning about the issues of GDA + best-response gradient dynamics. Actually, that's exactly our motivation to propose FR. FR and timescale separation are two solutions to overcome the undesirable convergence of GDA + best-response gradient dynamics (in main paragraph, it was stated in the setting of zero-sum games). That also explains why we don't need two timescale separation. We stress again that FR is sort of a better replacement of timescale separation.\\n\\nLet us know if you have any other questions/comments. Hope we can reach consensus soon. And if you agree that our explanation makes sense and therefore theorem 3 holds. Please update your rating as you acknowledged that our method is novel and our contributions are significant.\"}", "{\"title\": \"Your reasoning about best-response gradient dynamics does NOT apply to FR dynamics. Please pinpoint which step of our proof is wrong or give a specific example where FR fails.\", \"comment\": \"Thank you for your presenting your arguments. Nevertheless, we are still not convinced of the need to use timescale separation. We also believe that there is no discrepancy between our statement of Theorem 3 and our claim in the main text.\\n\\n- An minor unstated assumption:\\n\\nWe have noticed that we indeed made a mistake in our results in that, to apply Proposition 4, one needs the update rule to be differentiable. This in turn requires the objective function(s) to be thrice differentiable (in Assumption 1 we stated twice differentiable). We will modify Assumption 1 in our next revision.\\n\\n- About the example you mentioned in the second paragraph:\\n\\nThe example you bring up is very interesting. However, we believe that as long as the update rule of FR is differentiable so that its Jacobian changes smoothly, we will be able to apply Proposition 4, i.e. use Taylor expansion at a equilibrium $(x^*,y^*)$. In that case, how Hessian changes becomes immaterial, as that does not affect the Jacobian of the update rule. In other words, our results will hold in your example if the cost functions are thrice differentiable.\\n\\n- Connecting the result of Theorem 3 to our local convergence claim:\", \"our_claim_for_local_convergence_is_basically\": \"there exists a neighborhood $U$ with $(x^*,y^*)\\\\in U$ such that when initialized in $U$, FR converges to $(x^*,y^*)$. We believe such convergence follows from Theorem 3 just as the non-zero-sum convergence claim follows from Theorem 1. In particular, from Theorem 3, we know that at Stackelberg equilibria, the Jacobian eigenvalues fall in $[-1,1]$. Suppose that we consider a \\\"strict Stackelberg equilibrium\\\" $(x^*,y^*)$, i.e. $\\\\tilde{H}_{xx}$ and $G_{yy}$ are both positive definite, then the Jacobian eigenvalues would fall in $(-1,1)$. Then, from Proposition 4 (well known result), there exists a neighborhood $U$ of $(x^*,y^*)$ such that when initialized in $U$, the following iterates will be contractions in $\\\\Vert\\\\cdot\\\\Vert_2$.\\n\\nWe kindly welcome the reviewer to point out the discrepancy between Theorem 3 and our claim of local convergence in the line of reasoning above.\", \"note\": \"\\\"Strict Stackelberg equilibrium\\\" is essentially the definition for differentiable Stackelberg equilibrium in [1].\\n\\n- About the arguments you presented about IFT:\\n\\nThe reasoning you give concerns best-response gradient dynamics. However, FR is fundamentally different from best-response gradient dynamics, so currently we do not see how the analogy works for FR.\\n\\nFirst, what FR does in general-sum games is applying a correction to best-response gradient dynamics. Recall that GDA (with best-response gradient) is problematic, but FR's correction of GDA leads to satisfactory convergence. Therefore, problems with best-response gradient dynamics (the example you mentioned) does not apply to FR at all. We hope the reviewer to read our previous response about the relationship between FR and best-response gradient dynamics. They're fundamentally different.\\n\\nSecond, the problem you raised is exactly one of the motivations underlying FR (see Section 4, although it is stated for zero-sum games). The correction term of FR allows the follower to move toward $r(x)$ at every iteration (finally it hits $r(x^*)$, see Figure 1), even in non-zero-sum games. As we argued in our previous responses, the correction term of FR can be thought of as a replacement of timescale separation. Though the exact dynamics of FR is different from $x_{t+1} = x_t - \\\\gamma_1 h_1(x_t, r(x_t))$ (as we don't require y to be on the ridge $r(x)$), but it doesn't affect the fact that our FR converges to the right solution. And that's exactly our main contribution.\\n\\nTo be clear, our FR dynamics is given by:\\n$x_{t+1} = x_t - \\\\gamma_1 h_1(x_t, y_t)$\\n$y_{t+1} = y_t - \\\\gamma_2 \\\\nabla_y g(x_t, y_t) + \\\\gamma_1 (\\\\nabla_{yy} g)^{-1}\\\\nabla_{yx}g h_1(x_t, y_t)$\\n\\nTo sum up, we are still not convinced that there is any need to use two timescale update in FR for our results to hold. We kindly ask the reviewer for providing other evidences to support his/her arguments. Otherwise, our theorem 3 should be taken as correct.\"}", "{\"title\": \"Yes\", \"comment\": \"We're working on our response now, will get back to you very soon.\"}", "{\"title\": \"Clarifications\", \"comment\": \"Is it more clear now?\"}", "{\"title\": \"Elaborate a bit about the notations.\", \"comment\": [\"Thanks you for the explanation. Nevertheless, we feel rather confused about your notations in IFT part.\", \"What do $x$ and $y$ mean? The leader and the follower?\", \"What's the definition of $f$ and $g$? Is $f$ the objective of the leader while $g$ the objective of the follower?\", \"Why you sometimes use $h$, sometimes $h_1$ and $h_2$? What's the difference between them?\", \"What's the definition of $r$? Is it the same as ours? Why it sometimes take two arguments and sometimes just one?\", \"What does $D_2$ mean?\", \"Why would you expect the update of $y_t$ (the second last equation) coincides with $x_t$ (the last equation) if they are different players?\"]}", "{\"title\": \"Response Regarding General-Sum Result\", \"comment\": \"Thanks for the response on the general-sum results and going through your thoughts with me. I will try to be more clear and concrete in this explanation. Again, please let me know if something is unclear following this response or if you believe there is an issue with my reasoning.\\n\\nTo begin, I agree with you that there is nothing mathematically wrong in the steps of your proof. However, I believe that there is a problem with the conclusions you draw from the steps in your proof regarding the convergence guarantees of the FR dynamics. Let $x$ be the player 1 variable and $y$ be the player 2 variable and $f$ be the objective of player 1 and $g$ be the objective of player 2. To summarize my argument, consider $(x_1^{\\\\ast}, y_1^{\\\\ast})$ to be a critical point of the FR dynamics and $(x_2^{\\\\ast}, r(x_2^{\\\\ast}))$ to be a stackelberg equilibrium where $x_2^{\\\\ast} = x_1^{\\\\ast}$, it can be that $y_1^{\\\\ast} \\\\neq r(x_2^{\\\\ast})$ which would contradict the statement of Theorem 3. To motivate the more detailed analysis consider the following example. Suppose that $\\\\nabla_{yy}^2g(x, r(x))$ has positive eigenvalues everywhere, but anywhere off of it you have negative eigenvalues. Let us further consider that the change from positive eigenvalues to negative eigenvalues is continuous so that for any small neighborhood a perturbation of $r(x)$ drives you away from the equilibrium. Then the only initial conditions for which the FR dynamics would reach the equilibrium would be given an initialization of $r(x)$, which is a set of measure zero. Now I go into the details of why I believe that your result may only hold if you explicitly assume the FR dynamics use $r(x)$ or if you introduced a timescale separation carefully. \\n\\nConsider what the implicit function theorem (IFT), which says the following (where for only this statement $f$ is an arbitrary function and not the cost function of player 1):\\n\\nLet $U \\\\subset E$, $V \\\\subset F$ be open and $f : U \\\\times V \\\\rightarrow G$ be $C^q$, $q \\\\geq 1$. For some $x_0\\\\in U$, $y_0 \\\\in V$ assume the partial derivative in the second argument $\\\\nabla_{y}f(x_0,y_0) : F \\\\rightarrow G$ is an isomorphism. Then there are neighborhoods $U_0$ of $x_0$ and $W_0$ of $f (x_0 , y_0 )$ and a $r:U_0\\\\times W_0 \\\\rightarrow V$ such that for all $(x,w)\\\\in U_0\\\\times W_0$, $f(x,r(x,w))=w$.\\n\\nNow consider a follower update\\n$$y_{t+1}=y_t-\\\\gamma_2 \\\\nabla_y g(x_t,y_t)$$\\nand suppose at a critical point $(x_0,y_0)$ \\n$$\\\\nabla_y g(x_0,y_0)=0$$\\nand $\\\\nabla_{yy}^2g(x_0,y_0)$ is non-degenerate (hence, an isomorphism). Then IFT says that there exists \\n$$\\\\exists \\\\ U_0, \\\\ W_0, \\\\ r \\\\in C^q(U_0\\\\times W_0, V)$$\\nsuch that $x_0\\\\in U_0$, $\\\\nabla_yg(x_0,y_0)\\\\in W_0$ and such that\\n$$\\\\nabla_yg(x,r(x,0))=0, \\\\ \\\\forall x\\\\in U_0$$\\nHowever, even on $U_0$, this does not mean that the $y$ iterates generated by the system\\n$$x_{t+1}=x_t-\\\\gamma_1h_1(x_t, y_t))$$\\n$$y_{t+1}=y_t-\\\\gamma_2 \\\\nabla_yg(x_t,y_t)$$\\ncoincide with the $r(x)$ iterates when player 1 is doing\\n$$x_{t+1}=x_t-\\\\gamma_1h_1(x_t,r(x_t)).$$\\nwhere for simplicity we can just consider $h_1$ to be the total derivative of $f$.\\n\\nThis same line of reasoning should analogously apply to the FR dynamics. So this is effectively why I believe that your results would only hold if the dynamics were explicitly using $r(x)$. To fix this issue, I think you would need to show that $r(x)$ is attracting on the ball you get from the IFT and this would also require the timescale separation. Essentially, the issue in the conclusion given in Theorem 3 is why singular perturbation theory is needed. A good reference that may make things more clear is [4]. \\n\\nSo to summarize, I don't think what has been stated in Theorem 3 is what has been proved or it has not been stated and the underlying assumptions given rigorously enough. \\n\\nThanks for being willing to hear out my concerns regarding this result and having a cordial dialogue about it. Please let me know if you do not agree with my statements or if you have any questions.\\n\\n[4] Skinner, \\\"Singular Perturbation Theory\\\"\"}", "{\"title\": \"Thank you for quick reply!\", \"comment\": \"First, we really appreciate your further comments and clarifications. Sorry if our tone was a bit aggressive.\\n\\nRegarding the figure with the eigenvalues, we will update the paper once you post the response for Theorem 3. In the next version, we will include all eigenvalues. I meant to say positive semi-definite and sorry for the confusion.\\n\\nWe're looking forward to discussing more with you on Theorem 3.\"}", "{\"title\": \"Response Regarding Non-General-Sum Things\", \"comment\": [\"Before responding, I want to point out that [1] is the closest paper in the literature to this paper and so naturally it would be what the majority of the comments regarding related work would concern. I also want to make clear that my primary concerns in the review process have concerned the general-sum results, which is an independent issue. Moreover, I was not asking for experimental comparisons to the dynamics in [1] during the rebuttal period since this would be time consuming for you in a short time window and unreasonable to request. Sorry if that was not clear. I was simply making the point that the dynamics are similar for zero-sum games and so it may have been the most natural experimental comparison in the initial submission that could have made the paper stronger. With that being said, I do believe the example you provided along with the reasoning for the empirical advantage of FR makes sense and provides justification for why the dynamics may be preferable for training GANs.\", \"You are right about the gradient norm that it should reflect the same thing, thank you for pointing this out to me. As I stated in my initial review, the fact that the FR dynamics quickly drive the gradient norm to zero is impressive in my opinion. I am, and have been, of the opinion that the empirical results are strong and the example you provided along with the reasoning you provided reinforces this fact.\", \"Thanks for including the figure with the eigenvalues. However, I am not sure why you showed the top eigenvalues instead of the top and bottom eigenvalues since that is what would allow for evaluation of definiteness. It is not super clear to me that the Schur complement is actually positive from what you have shown. If I am missing something in this line of reasoning, please let me know.\", \"Thanks for reiterating your convergence results in zero-sum games. I did not mean to sound dismissive of them in my previous comment. My intention was to try and understand why the FR dynamics would be preferred to the Stackelberg gradient dynamics in zero-sum games. The example you provided and the justification gives some evidence for why it may be the case. You are right that that the deterministic results are not direct corollaries of any results in [1], you provide interesting results using momentum and preconditioning, and you devised a set of dynamics that inherit the properties from the Stackelberg gradient dynamics but may come with empirical advantages. I acknowledge that this set of contributions is meaningful.\", \"I will now work on explaining further the potential issue with the general-sum results I see and post shortly.\"]}", "{\"title\": \"Thanks for the reply\", \"comment\": \"Thanks for the reply. You write\\n\\n\\\"A main observation of [1] is that when GANs are trained to generate good samples, the generator seems to be closer to a saddle point than a local minimum of the loss (see Figure 5 and 6). Thus the GANs are not at local Nash equilibria, but achieve good empirical performance. Since (local) Nash equilibrium is the typical solution concept for simultaneous games, we consider this an evidence against viewing GANs as simultaneous games.\\\"\\n\\nThis is indeed one possible interpretation, but not arguably not the only one. I would therefore suggest giving the reader the above argument and letting them decide for themselves. The way it is written presently, it might be misunderstood as [1] investigating the question of simultaneous vs sequential game, which it does not.\\n\\nOn page 1, you also write\\n\\\"Unlike simultaneous games, many practical machine learning algorithms, including generative adversarial networks\\n(GANs) (Goodfellow et al., 2014; Arjovsky et al., 2017),\\nadversarial training (Madry et al., 2018) and primal-dual\\nreinforcement learning (Du et al., 2017; Dai et al., 2018),\\nexplicitly specify the order of moves between players and\\nthe order of which player acts first is crucial for the problem.\\\"\\n\\nThis, again, is a bit misleading since in the case of GANs, for instance, simultaneous gradient descent is commonly used and was recommended, for instance, in the tutorial of goodfellow (https://arxiv.org/pdf/1701.00160.pdf page 20).\\n\\nI'm sorry if I'm appearing pedantic, but the question of the right notion of game underlying GANs is an interesting research question that deserves its own treatment.\\nTo my knowledge there is no firm evidence one way or the other and therefore preliminary claims should be avoided. \\nI would appreciate if the authors could modify passages in question (or tell me why they disagree)\"}", "{\"title\": \"Further Clarifications and Comparisons with [1]. The reason why we don't need timescale separation to work.\", \"comment\": \"As the reviewer is concerned with our FR algorithm in general-sum games, we'd like to further clarify the relationship between our FR, timescale separation and best-response gradient dynamics (i.e. the dynamics of [1]). Hopefully, we can reach consensus before rebuttal deadline. Again, we appreciate all you detailed comments and suggestions.\\n\\nFirst, we present a view that FR in general-sum games is applying an additional preconditioner to the best-response gradient dynamics so as to fix the second-order condition of local Stackelberg equilibrium.\\n\\nFrom the sufficient/necessary conditions of local Stackelberg equilibrium, it can be seen that using best-response gradient to update is a way to match the first-order conditions. Therefore, it can be argued that best-response gradient dynamics is in fact the counterpart of GDA in general-sum games. Comparing FR in zero-sum games and non-zero-sum games ((5) and (8)), one can see that FR in non-zero-sum games essentially applies the same preconditioning matrix to best-response gradient dynamics. When applied to GDA in zero-sum games, the preconditioner fixes the second-order condition for local minimax and eliminates the need for timescale separation (Theorem 1). What we show in Theorem 3 is that, when applied to best-response gradient dynamics, it fixes the second-order condition for local Stackelberg equilibrium. From this point of view, it is reasonable why our algorithm does not require a timescale separation.\\n\\nIn short, our FR algorithm adds a correction term to the follower's update (or equivalently applys an asymmetric preconditioner for the whole dynamics), which can be thought as a replacement of timescale separation rather than the best-response gradient term (which is to fix the first-order condition). We believe that our FR is better than timescale separation in the sense that 1. timescale separation can lead to slow convergence due to the use of small learning rate. By contrast, we are allowed to use large learning rates while still converge to the right solution. 2. FR in general-sum games converges exactly to local Stackelberg equilibria, whereas best-response graident dynamics with timescale separation [1] is not yet shown to have this desirable property (see Remark 3 in [1]).\\n\\nNext, we would like to discuss a potential reason why the best-response dynamics in [1] works for zero-sum-games (in the sense that stable fixed points are exactly local Stackelberg equilibria) without timescale separation, but not necessarily in general-sum-games. Our understanding is that, in zero-sum games, the best-response term $\\\\nabla_{xy}^2 g (\\\\nabla_{yy} g)^{-1} \\\\nabla_y f$ becomes $\\\\nabla_{xy}^2 f(\\\\nabla_{yy} f)^{-1} \\\\nabla_y f$. This provides the best-response gradient dynamics with a nice alternative explanation, namely predicting the gradient of $\\\\nabla_{x} f(x,r(x))$ by estimating $y-r(x)$ to be $(\\\\nabla_{yy}f)^{-1}\\\\nabla_yf$ (a Newton step). It can be seen that this alternative explanation is no longer valid for general-sum games, since $y-r(x)\\\\approx (\\\\nabla_{yy}g)^{-1}\\\\nabla_y g$, whereas the best-response term uses $ (\\\\nabla_{yy} g)^{-1} \\\\nabla_y f$. This coincident is also reflected in the Jacobian calculation (see (4) in [1]): the upper-right block of the Jacobian exactly cancels to zero.\\n\\nTo sum up, we believe the key novelty of our algorithm is orthogonal to best-response gradient dynamics. Instead, we view it as a way to match the second-order conditions for local minimax/Stackelberg equilibria by combining it with GDA or best-response gradient dynamics. In the zero-sum setting, using best-response gradient dynamics indeed works (as shown in [1]). However, if our conjectured explanation is true, this could be somewhat coincidental.\"}", "{\"title\": \"Further Response [1/2]\", \"comment\": [\"Thank you for the detailed feedback! We will try to answer your questions more clearly this time.\", \"About Proposition 7 and 8: We believe these results are straightforward results of the definition of local Stackelberg equilibrium, which we already acknowledged to be proposed in [1] (see page 5). We've added another reference to [1] before Definition 4 for further clarity.\", \"About the correctness of Theorem 3: The dynamics we consider is always (8), where all derivatives are evaluated at $(x_t,y_t)$, so we do not require the follower to play the best response in the algorithm. We would like to emphasize that $r(x)$ is the global solution to a non-convex problem that we never explicitly use. We would like to breakdown the proof of Theorem 3 so that the reviewer may point out which step does not hold if under the dynamics of (8).\"], \"step_1\": \"We care about the fixed points of (8), and since the preconditioner is invertible, a fixed point of (8) necessarily satisfies $\\\\nabla_{y}g(x,y)=0$ and $D_x f(x,y)=0$.\", \"step_2\": \"At a point such that $\\\\nabla_{y}g(x,y)=0$ and $D_x f(x,y)=0$, the Jacobian of (8) can be calculated to be the last equation on page 16.\", \"step_3\": \"The eigenvalues of the Jacobian of (8) at such a point can be shown to be those of $I-\\\\eta_{x}\\\\tilde{H}_{xx}$ and $I-\\\\eta_{y}G_{yy}$.\\n\\nWe cannot see which step won't hold under our dynamics, namely Eqn (8). We kindly welcome the reviewer to point out which step is incorrect under the current dynamics. Otherwise, we should conclude that Theorem 3 is mathematically correct. If you are still concerned about Theorem 3, we are happy to discuss more before the rebuttal deadline.\\n\\n\\n[1] Fiez et al., \\\"Convergence of Learning Dynamics in Stackelberg Games\\\", https://arxiv.org/pdf/1906.01217v2.pdf\"}", "{\"title\": \"Further Response [2/2]\", \"comment\": \"- About the contribution of this paper in zero-sum-games: On a second read we now agree that the some stochastic results in [1], namely remark 4, hold under less restrictive assumptions. However, we do not believe this undermines our contribution significantly, even in the zero-sum setting.\\n\\nAmong other things, on the theory side, we: 1. proposed a novel algorithm and shown local convergence result for it (Theorem 1); 2. incorporated preconditioning and momentum into the algorithm with the same convergence guarantees (Proposition 6 and Theorem 2); 3. gave explicit local convergence rates (Theorem 2). None of these are corollaries of results in [1], and they hold under arbitrary constant learning rate ratios, which is much more favorable in practice.\\n\\nOn the practical side, we demonstrated effectiveness of our algorithms in both simple examples and GANs. In the GAN setting in particular, we show that the gradient norm quickly diminishes to zero. In contrast, the algorithm in [1] does not achieve this in GAN training (see details below).\\n\\nFor these reasons, we believe the contribution of our work is significant. This is not to mention our result for non-zero-sum games (Theorem 3).\\n\\n- About empirical comparison of FR and algorithm in [1]: We provide empirical results in GANs comparing FR and the dynamics in [1]. Here is the link for the results (https://docs.google.com/document/d/1HQbbd3nphY9bmhAu8Mir1M5gqpo5ZnlmewrVNVNdbsM/edit?usp=sharing). Currently, we find that the algorithm in [1] doesn't perform as well as FR (cf. Fig. 4 fourth column). In particular, the discriminator's output is not constant, and gradient norm does not decrease to zero. To be noted, we cannot exclude the factor of hyperparameter tuning for now, but we have tried our best to tune their method before the rebuttal deadline.\\n\\n- About the timescale separation, the majority of modern GAN papers use roughly same scale learning rates for both the generator and discriminator (e.g., BigGAN [2]). In [1], the learning rate of the generator has to be infinitely smaller than that of the discriminator. In practice, it potentially leads to mode collapse issue when a much smaller learning rate is used for the generator since the discriminator becomes too good and the gradient w.r.t the generator vanishes (see WGAN [3] paper for more discussions).\\n\\n- About plotting the gradient norm: First, we would like to point out that points where individual gradients are 0 have to be fixed points of FR. Note that when gradients are 0, all updates of FR are zero since the preconditioners are positive definite. In fact, the set of fixed points of FR is exactly the points where individual gradients are 0 (which is argued in our proof of Theorem 1). We are confused about the reviewer's remark that \\\"The norm of the individual derivative does not indicate if it is at a critical point or not\\\".\\n\\nSecond, we mean to use the gradient norm of different algorithms as a measure for convergence (see e.g. Fig. 5). These algorithms use different update rules, but are all based on individual gradients. Therefore, we believe that it makes more sense to plot the norm of individual gradients for a comparison.\\n\\n- Regarding the description of the results in [1]: We would clarify this in our next revision.\\n\\n- Regarding the eigenvalues, we plot the top-20 eigenvalues (sort according to the magnitude) for the Hessian of the discriminator and the Schur compliment in Appendix E.3.\\n\\n\\nLastly, we thank reviewer #1 for all the comments again. We hope our responses address your concerns, especially your primary concern on Theorem 3. \\n\\n\\n\\n[1] Fiez et al., \\\"Convergence of Learning Dynamics in Stackelberg Games\\\", https://arxiv.org/pdf/1906.01217v2.pdf\\n[2] Brock et al., \\\"Large Scale GAN Training for High Fidelity Natural Image Synthesis\\\", https://arxiv.org/pdf/1809.11096.pdf\\n[3] Arjovsky et al., \\\"Wasserstein GAN\\\", https://arxiv.org/pdf/1701.07875.pdf\"}", "{\"title\": \"Response\", \"comment\": \"Sorry for the delayed response.\", \"comments_on_the_general_sum_results\": [\"I appreciate the explanation on proposition 7 and 8. I do however, still have some significant concerns. To begin, I do not think you should be stating the propositions as a result in this paper given that the conditions are already given in [1]. The primary concern involves the analysis of the Jacobian for the FR dynamics in general-sum games. It seems that the results you have only hold if the follower is playing an exact best response along each step of the learning trajectory, that is y = r(x) for every x iterate. This would correspond to a set of decoupled dynamics where the only solution being learned is for the leader. Even though the dynamics push the follower to be close to r(x) it is not always exactly at r(x). To fix this issue and show the dynamics using y coincide with using r(x) you would need to use singular perturbation theory and introduce a timescale separation between the players. For this reason, I do not believe that Theorem 3 holds. If you do not agree with this line of reasoning, please feel free to explain why you believe it holds and hopefully we can discuss and resolve things before the discussion period ends. I am also happy to elaborate on my statements overhead if it is not clear what the issue is that I am raising.\", \"I have to respectfully disagree that the dynamics in this paper are more intuitive than the dynamics in [1]. The dynamics from that paper directly reflect the underlying game structure. However, I do see some utility in the interpretation in this paper. Since the dynamics have the same critical points in [1], similar deterministic guarantees for that set of dynamics could be given. For this reason, I am not sure that giving the deterministic convergence results is any more significant that stochastic results. The results in [1] also gives guarantees under relaxed assumptions beyond what you mentioned. I do not agree with the claim that timescale separation makes results not applicable to GANs since to my knowledge they are often trained with a timescale separation. This essentially what makes me skeptical about the level of contribution in zero-sum games.\", \"Thanks for fixing the issue with the gradient penalty regularization by removing it in the paper.\", \"Why do you plot the gradient norm for each player instead of the norm of the vector field? The norm of the individual derivative does not indicate if it is at a critical point or not.\", \"Thanks for including the details regarding the damping scheme in the paper and explaining how the eigenvalues were computed. I noticed that you have not included a plot of the eigenvalues in the paper yet. I believe that would be good to do. That is interesting to know that the generator hessian eigenvalues are all zero since that would correspond to the realizable assumption made in some GAN papers.\", \"That is a good point that the example in [3] does not directly contradict your claim since all stable critical points of simultaneous gradient descent in zero sum games will be stackelberg. That being said, I appreciate you removing your statement since it is not known that all non-nash stable critical points of the simultaneous gradient dynamics are stackelberg equilibrium in zero-sum games in higher dimensions.\", \"While you may say that [1] can converge to non-stackelberg points, the statement in remark 3 only regards general-sum games. The way it is stated in the paper now, it seems that you are saying the dynamics in [1] can converge to non-stackelberg points in zero-sum games. I believe you could edit this section to make it clear in zero-sum games, the dynamics in [1] will only reach critical points which are stackelberg, just as in this paper.\", \"Finally, in point 4 from my review, I was trying to ask why you did not compare empirically to the dynamics in [1] since they are the closest to the dynamics in this paper and have similar properties. Sorry that this was not clear.\"]}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"Thank you for your detailed comments and kind words about our work.\\n\\nRegarding the claim that GANs training is a sequential game, we note that the majority of GAN papers consider GAN training as a divergence minimization problem. For example, f-GAN is minimizing f-divergence, the original GAN is minimizing JS-divergence and W-GAN is minimizing Wasserstein distance. By taking this perspective, we are implicitly modeling GAN training as a sequential game since the divergence interpretation involves solving the maximization in the inner loop. Hence minimax should be an appropriate solution concept. We will clarify this point in our next revision. \\n\\nA main observation of [1] is that when GANs are trained to generate good samples, the generator seems to be closer to a saddle point than a local minimum of the loss (see Figure 5 and 6). Thus the GANs are not at local Nash equilibria, but achieve good empirical performance. Since (local) Nash equilibrium is the typical solution concept for simultaneous games, we consider this an evidence against viewing GANs as simultaneous games.\\n\\nThe reason why we used batch size 2,000 for MNIST is that our analysis was done for noiseless setting (full-batch). To exclude the factor of subsampling noise, we used large batch training. We leave the stochastic version of our algorithm to future work as it is highly non-trivial.\\n\\nRegarding the comparison with consensus optimization, we measured the training steps instead of wall-clock time. In terms of wall-clock time, consensus optimization does take fewer computation at each step. But we note that is not the main focus of our work.\\n\\nRegarding how we invert the Hessian, we discussed the details in section 6.2.1 in our submission and we've added more details in the Appendix D.4. Specifically, we solve the linear system $\\\\mathbf{H}_{yy}^2 z = \\\\mathbf{H}_{yy} \\\\mathbf{H}_{yx} \\\\nabla_{x} f$.\\n\\n\\n\\n[1] Berard et al., \\\"A closer look at the optimization landscapes of generative adversarial networks\\\", 2019.\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"Thank you very much for your kind words about our work. It's really encouraging that you think the problem we study is important.\\n\\nRegarding the use of Hessian (and its inverse) in our method, we agree that it seems hard to generalize to large-scale machine learning tasks for exact Hessian computation. However, we note that conjugate gradient method (or Hessian-free method) only involves Hessian-vector product which has roughly the same computation cost as one backpropagation. Particularly, conjugate gradient has been successfully applied to a wide range of tasks such as reinforcement learning[2], image classification[3] and meta learning[4]. In the paper, we used conjugate gradient for GAN training and we've added more details in the Appendix. Lastly, we would like to note that Hessian is not necessary for standard supervised learning tasks since first-order methods like gradient descent converges to the right solution (local minima)[1]. However, it might not be the case for sequential games, we believe that the use of Hessian is necessary for problem we study, otherwise we might find a wrong solution. In terms of the approximation error of conjugate gradient and how it affects our convergence guarantees, we leave it for future work.\\n\\n[1] Lee et al., \\\"Gradient descent only converges to minimizers\\\", 2017\\n[2] Schulman et al., \\\"Trust Region Policy Optimization\\\", 2015\\n[3] Martens, \\\"Deep learning via Hessian-free optimization\\\", 2010\\n[4] Rajeswaran et al., \\\"Meta-Learning with Implicit Gradients\\\", 2019\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"Thank you for your detailed comments and feedback. We hope that our responses below adequately resolve your concerns. Although we believe the current revision does a much better job of presenting these arguments, we warmly encourage you to provide any criticisms that may help us further express these points more clearly.\\n\\n- Regarding the comparison with [1]:\\n\\nWe are aware of and has cited the work in [1], and agree that more discussion is due. Compared to the dynamics in [1] (that is, Eqn (1) without noise), we believe that FR has two advantages. First, FR has a more intuitive interpretation, as it is trying to follow the follower's best response function, i.e., the ridge. In comparison, when positioned on the ridge, the update of the dynamics in [1] coincides with gradient descent-ascent and will drift away from the ridge. Second, the guarantees of FR carries over to non-zero-sum games. We believe this is important since it is a much more general setting and has applications such as hyperparameter tuning. In comparison, the authors of [1] acknowledged that in non-zero-sum games, their algorithm can converge to a point that is not differential/local Stackelberg equilibrium (see Remark 3 [1]).\\n\\nWe also note that the main focus of [1] is different, which is proving convergence of gradient dynamics in the presence of noise. However, to do so, they assume that there is only one local maximum for the follower (Assumption 2), and that there is a timescale separation between the leader and the follower. Because of these two assumptions, their main results are not directly applicable to many problems including GAN training. Roughly speaking, they showed stronger results under stronger assumptions.\\n\\n\\n- Regarding the definition of local Stackelberg equilibrium:\\n\\nOur definition for local Stackelberg equilibrium agrees with the definition in [1] up to edge cases. In particular, the Hessian of $\\\\phi(x) := f(x, r(x))$ is exactly $D^2f_1$ in Definition 4, [1]. Thus, requiring $D^2f_1$ to be positive semi-definite is almost the same as requiring $x$ to be a local minimum of $\\\\phi(x)$.\\n\\nWe apologize for stating Proposition 7 and 8 without further explanation. We have updated our paper with a detailed explanation for their derivation.\\n\\nIn particular, we can see that $\\\\nabla\\\\phi(x)=\\\\nabla_{x}f+\\\\nabla r(x)^T\\\\nabla_{y}f$, which is the same as the transpose of $Df(x,r(x))$ using the notation in [1]. It then follows that $\\\\nabla^2\\\\phi(x)=\\\\nabla_{x}(\\\\nabla_{x}f+\\\\nabla r(x)^T\\\\nabla_{y}f)+\\\\nabla_{y}(\\\\nabla_{x}f+\\\\nabla_{x}r(x)^T\\\\nabla_{y}f)\\\\nabla r(x)=DD f$. Substituting $\\\\nabla r(x)$ with $-G_{yy}^{-1}G_{yx}$ would prove Proposition 7 and 8.\\n\\n- Regarding the gradient penalty regularization, we removed this section as it has little connection with our main contributions.\\n\\n- Regarding the gradient norm, we meant the individual derivative for each player.\\n\\n- Regarding the detailed damping scheme, we added the details in Appendix D.4 of current revision.\\n\\n- Regarding the eigenvalues of the second order equilibria condition, we are able to compute the Hessian and its inverse exactly for networks we used for mixture of Gaussian experiments. The networks we used were 2-hidden-layer MLP with 64 hidden units for each layer. To be specific, we compute each row of the Hessian by multiplying the Hessian with a vector (with all entries $1$). To be noted, the Hessian-vector product can be done efficiently by doing reverse-mode autodiff (backpropgation) twice. For exact values, we notice that the eigenvalues of the generator's Hessian are all zero (which is not surprising since the discriminator outputs a flat line) while the eigenvalues of the discriminator's Hessian are all positive (as we added $L_2$ regularization). Therefore, it is easy to see that the Schur compliment $H_{xx} - H_{xy}H_{yy}^{-1}H_{yx} = - H_{xy}H_{yy}^{-1}H_{yx}$ is positive definite.\\n\\n- Regarding the local minimax convergence claims in other works: It is indeed our mistake to cite [2] for our claim; we have removed this citation in our revision. However, we do not believe Proposition 11 in [3] contradicts our claim. In their example, the min player moves faster than the max player, so the dynamics converge to a local maximin. By \\\"[1] can converge to non-local Stackelberg points\\\", we meant that stable limit points of [1] can be points that are not local Stackelberg, which is acknowledged by the authors of [1] in Remark 3.\\n\\n\\n\\n[1] Fiez et al., \\\"Convergence of Learning Dynamics in Stackelberg Games\\\", 2019.\\n[2] Heusel et al., \\\"GANs trained by a two time-scale update rule converge to a local Nash equilibrium\\\", 2017.\\n[3] Mazumdar et al., \\\"On Finding Local Nash Equilibria (and only Local Nash Equilibria) in Zero-Sum Games\\\", 2019.\"}", "{\"title\": \"First Revision\", \"comment\": \"We thank each of the reviewers for their time and comments.\\n\\nWe have uploaded our first revised version of our paper which addresses main concerns of reviewer #1. Particularly, we added details of our damping scheme in CG (Appendix D.4) and detailed derivation of proposition 7 and 8. Besides, we also removed section 4.1 due to the mistake reviewer #1 spotted.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"In this paper, the authors introduce a new optimization algorithm for minimax problems, or finding equilibria in sequential\\ntwo-player zero-sum games. Such problems are common in machine learning, including generative adversarial networks or primal-dual reinforcement learning. The commonly used gradient descent-ascent algorithm, corresponding to taking a gradient step for both players (or for both variables being minimized and maximized over), does not converge, in general, to local minimax points. Moreover, it can converges to fixed points which are not local minimax. To address these issues, the authors introduce the \\\"follow the ridge\\\" algorithm for minimax optimization problems. Given a minimax problem min_x max_y f(x, y), this algorithm consists in adding a correction term to the gradient corresponding to the y variable (corresponding to the max). This term is derived from the observation that minimax optimization should follow ridges (i.e. local maximum w.r.t. to y) of the function. Ridges can be defined as the implicit functions such that the gradient w.r.t. y is equal to zero, allowing to design an update that would stay \\\"close\\\" to the ridge. The correction term corresponding to the update thus involve the inverse of the Hessian w.r.t. y. The authors prove that all the fixed points of this algorithm are minimax, and that all local minimax are fixed points of the algorithm. The proof use first and second order conditions for local minimax points, which were recently derived in a paper by Jin et al. The proposed algorithm can also be used with momentum and preconditioning, and be generalized to Stackelberg games. Finally, the authors evaluate the follow the ridge algorithm on toy low dimensional GAN problems, as well as experiments on the MNIST dataset, showing better convergence that other methods used for minimax optimization problem.\\n\\nThe problem studied in this paper is an important one, as it arises in multiple area of machine learning such as adversarial \\ntraining or reinforcement learning. It has also received significant attention from the community in the recent years. This paper propose a simple solution, which is well motivated, to the problem as well as a proof of convergence. A limitation of the proposed method is that it uses the Hessian of the problem, probably making it hard to apply on large scale problems that are common in deep learning. I believe that it would make the paper stronger to discuss potential ways to mitigate this issue (e.g. inspired by L-BFGS), and their impact on theoretical guarantees. (Note that the authors briefly mention using the conjugate gradient algorithm in the experimental section).\\n\\nOverall, the paper is well written, and easy to follow (even for non-expert like me). I believe that it does a good job at introducing the problem and existing work on which it builds, and to motivate the proposed solution. I have not checked the proofs carefully, but they seem sensible. A small weakness of the paper is the experimental section: for example, I am not sure the MNIST experiments bring much to the paper, and would have preferred more convincing experiments. However, this is mostly a theoretical paper, and I do not think this is a big concern.\\n\\nTo summarize, I think the paper study an important problem, proposes a sound solution and is clearly written. For these reasons, I believe that the paper should be accepted to the ICLR conference. However, as I am not an expert on this area, my recommendation is a low confidence one.\", \"minor_comment\": \"I believe that at the beginning of second paragraph of section 4, \\\"Suppose that y_t is a local minimum of f(x_t, .)\\\" should be \\\"maximum\\\".\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper focus on sequential games, which is the common formulation of GANs and a number of games in machine learning applications. From this perspective, it is natural to look at Stackelberg equilibria. In my opinion, the objective of the paper is important and relevant. The theoretical and empirical results are reasonably convincing. However, I do have some rather serious concerns about the general-sum game results and several questions regarding the relation to related work and the experiment details that need to be addressed.\\n\\n1. The FR dynamics in algorithm 1 are closely related to the dynamics in [1]. In particular, the Jacobian of the FR dynamics is a similarity transform of the Jacobian of the dynamics in [1]. As a result, each algorithm has the same set of stable attractors. This should probably be mentioned in the paper. Given this relation, it is not clear what the advantage of the FR dynamics are over the dynamics in [1]. Could you please discuss this?\\n\\n2. The gradient penalty regularization connection does not make sense in section 4.1. The optimization problem presented has an issue because the dimensions do not align in the constraint. The quantity \\\\nabla_x f(x, y)^T H_yy^{-1}\\\\nabla_x f(x, y) would not be defined if the dimensions of the players are not equal. \\n\\n3. In the related work it is claimed that two time-scale GDA converges only to local minimax and [2] is cited. I would avoid using this claim with respect to that paper since the statement following the main result in the paper is not right (see proposition 11 of [3] for proof). It is not clear what is meant when it is claimed that [1] can converge to non-local Stackelberg points. The dynamics in [1] only converge to local minimax points in the special case of zero-sum games.\\n\\n4. Since the dynamics in the paper are the closest to those in [1], it seems that the paper would be stronger by comparing with that set of dynamics. \\n\\n5. I found it to be quite impressive that the vector field is driven to zero in the GAN examples. Just to clear, for each algorithm when the \\u2018gradient norm\\u2019 is shown, does this mean the norm of the update for each norm or does it mean the individual derivative for each player. For example in FR, would it be the norm of the derivative with respect to the follower variable of the function or the norm of the update including the second order information?\\n\\n6. The path angle plot was interesting to see for the GAN example. The authors claim that the eigenvalues of the second order equilibria condition are non-negative. It would be nice if the authors could show the eigenvalues in the appendix and discuss how they were computed since it may be non-trivial to compute depending on the network size.\\n\\n7. The damping method to stabilize training is not quite clear. Could you provide more details about how this was done?\\n\\n\\nMy primary concerns have to do with the portion of the paper considering general-sum games. I do not understand where proposition 7 and 8 come from. I am not convinced the definitions provided are necessary and sufficient conditions for Stackelberg equilibria. In [1], a differential Stackelberg equilibrium is defined. The definition in this paper does not appear to agree with the definition in [1]. The final positive definite condition in proposition 7 and 8 does not appear to be taking the total derivative 2 times when I evaluate the derivatives, so I am not sure what the quantity is. If this is not a proper set of conditions for the equilibria, then it would also mean that the dynamics do not only converge to equilibria in general-sum games. It is important that the authors clear up this concern since I do not believe Theorem 3 holds as a consequence of problems with propositions 7 and 8.\\n\\n[1] Fiez et al., \\\"Convergence of Learning Dynamics in Stackelberg Games\\\", 2019.\\n[2] Heusel et al., \\\"GANs trained by a two time-scale update rule converge to a local Nash equilibrium\\\", 2017.\\n[3] Mazumdar et al., \\\"On Finding Local Nash Equilibria (and only Local Nash Equilibria) in Zero-Sum Games\\\", 2019.\", \"post_author_response\": \"Thanks to the authors for the effort in discussing the paper with me. The authors made several changes to the paper in response to my comments including removing section 4.1, fixing comments about related work, including details on the damping procedure, showing experimental comparisons to [1] along with an explanation of why the dynamics in this paper may be preferred for training GANs, providing details on propositions 7 and 8 and including reference to [1], adding further assumptions on the functions, and attempting to make theorem 3 more clear. Overall, I think this paper proposes an interesting set of dynamics, several meaningful theoretical guarantees, and impressive empirical results. I would be curious to see how it performs on even more large-scale GAN problems in the future. As a result, I have changed my original score from a weak reject to a weak accept. My primary concerns with the paper regarded the general-sum convergence results and I appreciated the explanations from the authors. I am still of the opinion that theorem 3 could be stated more rigorously in the sense that the neighborhood on which the local convergence holds should be more explicit. It seems to me that the convergence result may only hold in a ball around an equilibrium in which the implicit function is well-defined and the FR dynamics will be attracted to r(x) and that this space could be arbitrarily small for some problems. Nonetheless, this result is only in the appendix, and the paper includes enough contributions beyond this to warrant acceptance.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary\\nThe present work proposes a new algorithm, \\\"Follow the Ridge\\\" (FR) that uses second order gradient information to iteratively find local minimax points, or Stackelberg equilibria in two player continuous games. The authors show rigorously that the only stable fixed points of their algorithm are local minimax points and that their algorithm therefore converges locally exactly to those points. They show that the resulting optimizer is compatible with heuristics like RMSProp and Momentum. They further evaluate their algorithm on polynomial toy problems and simple GANs.\\n\\nDecision\\nI think that this is a solid paper that addresses the well-defined goal of finding an optimizer that only converges to local minimax points. This is established based on both theoretical results and numerical experiments. Since there has been a recent interest in minimax points as a possible solution concept for GANs, I believe the paper should be accepted.\\n\\nThe paper occasionally makes claims that the solutions of GANs should consist of local minimax points (\\\"We emphasize that GAN training is better viewed as a sequential game rather than the simultaneous game, since the primary goal is to learn a good generator.\\\"), which are not backed up by empirical results or reference to existing literature. If anything, the empirical results in this paper do not show improvement of the resulting generator (with the exception of the 1-dimensional example that has a particular rigidity since low discriminator output can easily restrict the movement of generator mass based on first order information). The right solution concept for GANs is not what the paper is about, but before publication the authors should remove these claims, identify them as speculative, or substantiate them .\\n\\nSuggestions for revision\\n(1) In the last displayed formula on page 4 it should be the gradient w.r.t x.\\n(2) Remove, substantiate, or mark as speculative the claims regarding the right notion of solution concept for GANs.\\n\\nQuestions to the authors\\n(1) You write \\\" There is also empirical evidence against viewing GANs as simultaneous games (Berard et al., 2019). \\\". Could you please elaborate, why Berard et al. provides empirical evidence against viewing GANs as simultaneous games?\\n(2) The Batch size for MNIST of 2000 is much larger than the values I have seen in other works. What is the effect of using more realistic batch sizes in training?\\n(3) When measuring the speed with which consensus optimization and FR converge, shouldn't you allow consensus optimization five times as many iterations, since you are using five iterations of CG to invert the Hessians in each step?\\n(4) You mention that you use CG to invert the Hessian, but the Hessian is not positive definite? Do you apply CG to the adjoint equations?\"}", "{\"comment\": \"We noticed a small visualization problem in Fig. 4. The KDE plots (first row) were generated by the seaborn package. The function kdeplot (https://seaborn.pydata.org/generated/seaborn.kdeplot.html) chooses a Gaussian kernel with improper bandwidth by default, so the modes in our figures look wider than they actually are. We emphasize that this does not affect our claim that FR learns the true distribution and our comparison of FR with other algorithms.\", \"title\": \"Small visualization problem in Fig. 4\"}" ] }

[ "Shiori Sagawa*", "Pang Wei Koh*", "Tatsunori B. Hashimoto", "Percy Liang" ]

[ "distributionally robust optimization", "deep learning", "robustness", "generalization", "regularization" ]

{ "note_id": [ "Yw6Tpflcm", "HJgQOqWqiB", "HkeL89WqsB", "ryxpHc-qoS", "Syge19-5jr", "BJexEXB09r", "HJg4BjDy5B", "rJgFdjSYFr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732649, 1573685866881, 1573685837772, 1573685828794, 1573685720370, 1572913959978, 1571941180242, 1571539824769 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1796/Authors" ], [ "ICLR.cc/2020/Conference/Paper1796/Authors" ], [ "ICLR.cc/2020/Conference/Paper1796/Authors" ], [ "ICLR.cc/2020/Conference/Paper1796/Authors" ], [ "ICLR.cc/2020/Conference/Paper1796/AnonReviewer5" ], [ "ICLR.cc/2020/Conference/Paper1796/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1796/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper proposes distributionally robust optimization (DRO) to learn robust models that minimize worst-case training loss over a set of pre-defined groups. They find that increased regularization is necessary for worst-group performance in the overparametrized regime (something that is not needed for non-robust average performance).\\n\\nThis is an interesting paper and I recommend acceptance. The discussion phase suggested a change in the title which slightly overstated the paper's contributions (a comment which I agree with). The authors agreed to change the title in the final version.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer #5\", \"comment\": \"Thank you very much for your comments. We appreciate the emergency review!\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"Thank you very much for your detailed and helpful comments.\\n\\n\\n1. \\u201cThe title is a little bit vague and may overstate the paper's contribution. In fact, the problem addressed in the paper is not as general as the title suggests and has little to do with the generic DRO framework.\\u201d\\n\\nWe will change our title to \\u201cDistributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization.\\u201d We hope the new title highlights our focus on the group DRO setting as well as our specific contributions. In addition, we will edit the main text to further emphasize that our work is on the group DRO setting.\\n\\n\\n2. \\u201cThe groups need to be chosen using prior knowledge of \\u2018spurious associations\\u2019, which may arouse doubts about the algorithm's actual effectiveness and undermines the value of the algorithm on the application side.\\u201d\\n\\nPlease refer to our response to all reviewers. This is an important point that we will further discuss in the updated manuscript.\\n\\n\\n3. \\u201cSome terms used in the paper are ambiguous. For example the concept of \\\"group\\\" is not defined properly. Sometimes it refers to a subset of the dataset while sometimes it denotes a sub-distribution in the mixture distribution. The authors should use more precise terminology.\\u201d\\n\\nBy a \\u201cgroup\\u201d, we refer to a sub-distribution in a mixture distribution. If the mixture distribution is a population distribution, then the group refers to a sub-distribution of the population distribution; and if the mixture distribution is an empirical distribution (e.g., the set of training examples sampled from the population), then the group refers to sub-distribution of the empirical distribution, which corresponds to a subset of the sampled dataset. We will define groups more carefully in Section 2 and distinguish between the empirical and population distributions in other sections. If there are any specific parts in which such clarifications would be helpful, please let us know.\\n\\n\\n4. \\u201cThere is a typo in Proposition 1, where the uncertainty set Q should be a set of probability distributions, not a subset of R^m.\\u201d\\n\\nThat is indeed a typo, and we will fix it as suggested. Thank you.\"}", "{\"title\": \"Response to all reviewers\", \"comment\": \"We thank all of our reviewers for their detailed and very helpful comments. We have responded to each reviewer individually and will update our paper with the corresponding changes.\\n\\nHere, we discuss Reviewers 1 and 2's comment about how our method uses pre-specified groups. Prior work has shown that it is difficult to learn distributionally robust models without some prior knowledge or structural assumptions on groups. Without any assumptions, the worst-case group of training data could be exactly the group of training points with the highest loss, but minimizing the loss on these worst-case points often yields models that are too conservative to be effective (Hu et al., 2018; Oren et al., 2019). Asking domain experts to pre-specify groups is a direct way of incorporating such prior knowledge, and as we discuss below, we believe it is feasible in many applications. As Reviewer 1 suggests, exploring other approaches that build upon the group DRO framework (e.g., learning groups by making assumptions on group structure) would be promising future work.\\n\\nExisting work in many application areas has identified realistic distributional shifts that can be practically expressed in terms of pre-specified groups. For instance, in biology, batch effects like the day on which an experiment was performed are a significant source of spurious variability (e.g., Leek et al., 2010); and in medicine, patient demographics like ethnicity are spuriously associated with patient outcomes (e.g., Badgeley et al., 2018). In NLP, we studied a specific annotation artifact (negations), but recent work has categorized many others (Gururangan et al., 2018; Naik et al., 2018; Liu et al., 2019; McCoy et al., 2019). In all of these examples, the group identity of each training example is already present in the dataset, and we can use it to build models that are robust to the corresponding distributional shifts.\\n\\nMoreover, we need not be too precise when specifying groups. Concretely, it can be sufficient to specify a set of candidate attributes for spurious correlations, as long as one is related to the ground-truth spurious attribute. In a new CelebA experiment, group DRO maintains high robust accuracy when: 1) instead of the ground-truth spurious attribute Male, we specify a related attribute Wearing_Lipstick instead, and 2) we also specify four distractor/non-spurious attributes (Eyeglasses, Smiling, Double_Chin, and Oval_Face). Optimizing for worst-case performance over all $2^6 = 64$ groups, the DRO model attains 78.9% robust accuracy across the 4 original groups (dark-haired males and females, and blond males and females). These robust accuracies are not far off from the original DRO model with just the ground-truth spurious attribute (86.7%) and significantly outperform the ERM model (37.8%).\\n\\nOf course, there are still many applications that do not lend themselves to pre-specified groups. We agree with Reviewer 1 that learning groups is an interesting direction for future work; the key challenge there is to understand what structural assumptions on groups (e.g., we might assume data points within a group cluster together) lead to meaningful groups in practice.\", \"references\": \"1. W. Hu, G. Niu, I. Sato, and M. Sugiyama. Does distributionally robust supervised learning give robust classifiers? In International Conference on Machine Learning (ICML), 2018.\\n2. Y. Oren, S. Sagawa, T. Hashimoto, and P. Liang. Distributionally robust language modeling. In Empirical Methods in Natural Language Processing (EMNLP), 2019\\n3. R. T. McCoy, E. Pavlick, and T. Linzen. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Association for Computational Linguistics (ACL), 2019.\\n4. S. Gururangan, S. Swayamdipta, O. Levy, R. Schwartz, S. Bowman, and N. A. Smith. Annotation artifacts in natural language inference data. In Association for Computational Linguistics (ACL), 5. 2018.\\n5. A. Naik, A. Ravichander, N. Sadeh, C. Rose, and G. Neubig. Stress test evaluation for natural language inference. In International Conference on Computational Linguistics (COLING), pp. 2340\\u20132353, 2018.\\n6. N. F. Liu, R. Schwartz, and N. A. Smith. Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets. In North American Chapter of the Association for Computational Linguistics (NAACL), 2019. \\n7. M.A. Badgeley, J.R. Zech, L. Oakden-Rayner, B.S .Glicksberg, M. Liu, W. Gale, M.V. McConnell, B. Percha, T. M. Snyder, and J. T. Dudley. Deep learning predicts hip fracture using confounding patient and healthcare variables. npj Digital Medicine, 2, 2019.\\n8. J. T. Leek, R. B. Scharpf, H. C. Bravo, D. Simcha, B. Langmead, W. E. Johnson, D. Geman, K. Baggerly, and R. A. Irizarry. Tackling the widespread and critical impact of batch effects in high-throughput data. Nature Reviews Genetics, 11(10), 2010.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"Thank you very much for your detailed comments and for going through our paper so carefully.\\n\\n\\n1. \\u201cThe algorithm has a downside in that the groups must be known a priori, is it possible for these groups to be learnt?\\u201d\\n\\nThanks for raising this point; please refer to our response to all reviewers.\\n\\n\\n2. \\u201cCan using a hinge loss also improve robustness to the worse case examples?\\u201d\\n\\nWe ran new experiments using the hinge loss in the CelebA and Waterbirds datasets (since these are binary classification tasks), but performance on the worst-case group did not improve. Compared to the logistic loss, performance with the hinge loss was similar or slightly worse across each group. For CelebA (on the test set and with early stopping), robust accuracy was 84.4% with the hinge loss and 88.3% with the logistic loss; and similarly for Waterbirds, robust accuracy was 84.7% with the hinge loss and 84.9% with the logistic loss.\\n\\n\\n3. \\u201cWhat is the effect on the training time of this algorithm? Is it just the time for an additional forward prop?\\u201d\\n\\nOptimizing for the worst-case group loss instead of the average loss has little effect on the run time of the algorithm. In practice, optimizing for the worst-case group loss vs. average loss takes a similar amount of time (<5% difference) for a given number of epochs. For example, it took 12h 50min to train the CelebA ERM model for 50 epochs on an NVIDIA TITAN Xp, and 13h 20 min to train the corresponding DRO model.\\n\\nThe small difference above is expected because the computation of the loss $\\\\ell(\\\\theta^{t-1}; (x,y))$ and its gradient $\\\\nabla\\\\ell(\\\\theta^{t-1}; (x,y))$ dominates the run time of each iteration in both optimization algorithms. The robust optimizer in Algorithm 1 requires only a few additional computations over the standard optimizer: a. Multiplying the weights $q$ by exponentiated losses and normalizing (second-to-last line in Algorithm 1), and b. Multiplying the loss gradient by $q$ (last line in Algorithm 1). These are relatively cheap operations.\\n\\nThis comparison of training times is an important point to clarify in the paper, and we will make it explicit in our revision. Thanks for highlighting it.\\n\\n\\n4. \\u201cWhat is the effect on the worse-case examples of weight decay on the BERT model? Even though it hurts the average performance does it improve the worse case at all? \\n\\nOn the BERT model, we found that weight decay (specifically, $\\\\lambda \\\\in \\\\{0.01, 0.03, 0.1, 0.3, 1.0\\\\}$) did not seem to affect the model\\u2019s accuracy on the worst-case group in comparison to the model without weight decay; training the above models to convergence resulted in similar (bad) performance on the worst-case group. \\n\\nMuch larger values of weight decay ($\\\\lambda \\\\in \\\\{3.0, 10.0\\\\}$) were too conservative and significantly lowered model performance across all groups. In CelebA and Waterbirds datasets, we similarly observed that weight decays that are too high result in poor overall performance.\\n\\nWhile it is possible that we have not found the appropriate value of weight decay for the BERT/MultiNLI model, these results may suggest that weight decay is not an effective form of regularization (compared to early stopping) for this particular model. \\n\\n\\n5. \\u201cIn the last line of algorithm 1 why is $q_G$ used instead of $q_g$?\\u201d\\n\\nThat is a typo, and it should indeed be $q_g$. We will fix it. Thank you.\\n\\n\\n6. \\u201cAt the bottom of P5: The ordering of 93.4% and 97.1% seem to be reversed.\\u201d\\n\\nThat is also a typo, and we will fix it. Thank you.\\n\\n\\n7. \\u201cAbove eq. 5 , $\\\\delta_g$ seems to be overloaded. In the paragraph, it first refers to the generalization gap and then later to a heuristic.\\u201d\\n\\nThanks for catching the overloading. We will refer to the heuristic as $\\\\hat{\\\\delta}_g$ instead.\\n\\n\\n8. \\u201cTable 2: The drop in average accuracy for waterbirds does not seem 'small'.\\u201d\\n\\nYou\\u2019re right that it is not a small change in relative error. Thanks for the catch; we will clarify the text.\\n\\n\\n9. Bottom of P8: 'background is more unique', it seems this is supposed to mean the background appears less often?\\n\\nWe will explain the example more clearly in the updated version. The idea is the following: Assume that the actual spurious association is the same as in the original Waterbirds dataset (whether the background is water or land). However, instead of a single water background and single land background, we now have fine-grained labelings of water and land backgrounds, such that waterbirds appear in 9 different water backgrounds (e.g., \\u201clake\\u201d, \\u201cpond\\u201d, \\u201csea\\u201d, etc.) and 1 land background, while landbirds appear in 9 different land backgrounds and 1 water background. In this setting, each group of (waterbird/landbird, background) is the same size, so resampling yields the same model as ERM. However, the DRO model would correctly upweight the waterbirds on a land background (and vice versa).\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #5\", \"review\": \"To the best of my knowledge, this is the first paper to carefully address and propose an algorithm (with guarantees) for distributionally robust learning in the overparametrized regime, which is typical of modern large deep neural networks. Following other work, the paper formalizes distributionally robust learning as the minimization of a worst-case loss over a set of possible distributions. The main message of the paper is that for distributionally robust learning, regularization plays an important role by avoiding perfect fitting to the training data, at the cost of poor generalization (thus lack of robustness) in some of the possible distributions. Furthermore, the paper proposes a new stochastic optimization algorithm to minimize the loss that corresponds to distributionally robust learning and gives convergence guarantees for the proposed algorithm.\\n\\nI think this is an interesting and solid paper, with a clear presentation style, and well-supported contributions. To the best of my knowledge, this is novel work and, in my opinion, it is relevant work, both in terms of applicability as well as in terms of contribution to the understanding of the generalization behavior of overparameterized deep neural networks.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper describes a method of training neural networks to be robust to a worse case mixture of a set of predefined example attributes. This is done with a loss in accuracy in the average case but improvements in the worse case. The proposed algorithm is relatively simple and convergence rates are also given for this new algorithm.\\n\\nBuilding neural networks that perform well in the face of group-level worse case test-set distributions is a very important problem particularly in areas such as health and safety-critical applications as past work points out. This paper shows good results in the worse case and additionally shows that the common technique of importance reweighting cannot arrive at the same solution. The convergence analyses also yield additional insight into this new algorithm. The paper is well written and relatively easy to understand with good details on the experimental setup. The algorithm has a downside in that the groups must be known a priori, is it possible for these groups to be learnt? Also, can using a hinge loss also improve robustness to the worse case examples?\\n\\nHowever, there are some unanswered questions in the paper. What is the effect on the training time of this algorithm? Is it just the time for an additional forward prop? What is the effect on the worse-case examples of weight decay on the Bert model? Even though it hurts the average performance does it improve the worse case at all? In the last line of algorithm 1 why is q_G used instead of q_g?\", \"other_comments\": \"\", \"at_the_bottom_of_p5\": \"The ordering of 93.4% and 97.1% seem to be reversed.\\nAbove eq. 5 , \\\\delta_g seems to be overloaded. In the paragraph, it first refers to the generalization gap and then later to a heuristic.\", \"table_2\": \"The drop in average accuracy for waterbirds does not seem 'small'.\", \"bottom_of_p8\": \"'background is more unique', it seems this is supposed to mean the background appears less often?\\n\\n======================================================================================================\", \"update_after_rebuttal\": \"Thanks for the detailed answers to my comments and the additional experiments done with a hinge loss. I will keep my rating.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a novel training method based on group DRO that makes it possible for overparameterized neural networks to achieve uniform accuracy over different groups of the data, where the data is required to obey a mixture distribution composed of these groups. In addition, the authors identify that increased regularization plays a critical role for worst-group performance in the overparameterized regime by a series of empirical studies. The authors also compare their method with the traditional importance weighting method. Finally, they introduce a stochastic optimizer for their group DRO method.\", \"pros\": [\"The experiments are well designed and the results show the effectiveness of the approach.\", \"Most parts of the paper are well written and easy to follow.\"], \"cons\": [\"The title is a little bit vague and may overstate the paper's contribution. In fact, the problem addressed in the paper is not as general as the title suggests and has little to do with the generic DRO framework.\", \"In the group DRO setting the data is distributed as a mixture of different groups. However, the groups need to be chosen using prior knowledge of \\\"spurious associations\\\", which may arouse doubts about the algorithm's actual effectiveness and undermines the value of the algorithm on the application side.\", \"Some terms used in the paper are ambiguous. For example the concept of \\\"group\\\" is not defined properly. Sometimes it refers to a subset of the dataset while sometimes it denotes a sub-distribution in the mixture distribution. The authors should use more precise terminology.\", \"There is a typo in Proposition 1, where the uncertainty set Q should be a set of probability distributions, not a subset of R^m.\"]}" ] }

[ "Nikita Moriakov", "Jonas Adler", "Jonas Teuwen" ]

[ "Generative models", "CycleGAN" ]

{ "note_id": [ "6-NwgO-Bfir", "63qKMbhxrd", "sUqMNEuHlP", "HygDhZNnjH", "SkelelV3sB", "H1exJyVnoB", "B1l_X-GJcS", "rylknOkRFr", "H1e26UE6tr" ], "note_type": [ "official_comment", "comment", "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1587498706205, 1584444108547, 1576798732620, 1573826990636, 1573826536065, 1573826263738, 1571918112161, 1571842215177, 1571796676366 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Paper1794/Authors" ], [ "~Tomer_Galanti1" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1794/Authors" ], [ "ICLR.cc/2020/Conference/Paper1794/Authors" ], [ "ICLR.cc/2020/Conference/Paper1794/Authors" ], [ "ICLR.cc/2020/Conference/Paper1794/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1794/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1794/AnonReviewer2" ] ], "structured_content_str": [ "{\"title\": \"Thank you for your feedback\", \"comment\": \"Dear Tomer,\\n\\nFirstly, thank you very much for your feedback for bringing this work to our attention, it is very interesting. Naturally, we will be adding references to these contributions in the updated version of the paper. Concerning the existence of the automorphisms. The cited sections indeed explain informally that the cycle consistency constraints are not enough to guarantee uniqueness of solutions in general, but the assumptions on the underlying probability space are never fully formalized as far as we can see. Existence of automorphisms in general is a very well-defined property that holds for many (but not all) probability spaces. This assumption, why it matters and how it can be satisfied in practice for domains like natural images and smooth mappings between them is what we wanted to highlight in our submission.\"}", "{\"title\": \"The Role of Minimal Complexity in Unsupervised Learning of Semantic Mappings\", \"comment\": \"Thank you for contributing to the theory of unsupervised image to image translation. The existence of automorphisms discussed in Sec. 2.2 was already investigated in \\\"The Role of Minimal Complexity in Unsupervised Learning of Semantic Mappings\\\" by Galanti et al. 2018 https://openreview.net/pdf?id=H1VjBebR- (see Secs. 2 and 4) and in \\\"Generalization Bounds for Unsupervised Cross-Domain Mapping with WGANs\\\" by Galanti et al. 2018 https://arxiv.org/pdf/1807.08501.pdf (see Secs. 4.1 and 4.2). The solution to this problem was further discussed in Sec. 5 of the latter. We would appreciate it if you cite our work.\"}", "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper theoretically studied one of the fundamental issue in CycleGAN (recently gained much attention for image-to-image translation). The authors analyze the space of exact and approximated solutions under automorphisms.\\n\\nReviewers mostly agree with theoretical value of the paper. Some concerns on practical values are also raised, e.g., limited or no-surprising experimental results. In overall, I think this is a boarderline paper. But, I am a bit toward acceptance as the theoretical contribution is solid, and potentially beneficial to many future works on unpaired image-to-image translation.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for your feedback!\", \"comment\": \"Thank you very much for your feedback and additional references, this is very interesting!\\n\\nWe opted to use the measure theory language because the language of probability distributions is not flexible enough to accomodate commonly occurring distributions, e.g. those (strictly) supported on lower dimensional manifolds. Unless we require the PDF explicitly to have some symmetries, it is not clear why the corresponding probability space would have any smooth automorphisms at all. Thus we show that the existence of automorphisms is a very general property, and in the setup of e.g. latent space with spherical Gaussian PDF we show that there are smooth automorphisms as well.\\n\\nWe would like to point out that the goal of the paper is to provide a well-grounded and mostly self-contained analysis for the basic CycleGAN approach and to analyze theoretically the effect of the commonly used identity loss, along with some experiments to justify the claims. While the problem of multiple solutions for the CycleGAN is commonly realised, a good theoretical explanation for this is lacking in ML literature. \\n\\nNaturally, there are other approaches to unsupervised image-to-image translation with different losses and architectures. While analyzing all of them in a single paper is not realistic, we think that the philosophy we suggest in this paper can help researchers better understand the potential and the limitations of these newer image-to-image translation models. The underlying automorphisms can always pose a problem, and the question then becomes if a new loss/new architecture explicitly restricts this set.\\n\\nWe have added some additional experiments on BRATS 2015 dataset. In this set of experiments we will show how the loss and PSNR (since we have a ground truth) change when we vary the weight for identity loss. We introduce a (approximate) probability automorphism in the form of left/right flips and show that this highly unwanted transformation still obtains low loss values unless an identity loss is used.\\n\\nWe have also added a discussion about the MISO paper you suggested, where we hypothesize that the MISO approach does not 'solve' the issue of unwanted automorphisms, but rather restricts the set of these automorphisms to those that leave the style of the image fixed. Therefore the amount of uncertainty in this solution is connected to the capacity of the style encoder, and should ideally be quantified. When some important style content is present - e.g., anatomical landmarks - it seems reasonable that one should make sure that the style encoder learns this information. We think it is a an interesting question for future work.\"}", "{\"title\": \"Thank you for your feedback!\", \"comment\": \"Thank you very much for your feedback!\\n\\nAs for the MNIST2MNIST task, we observed that adding identity loss here forces the network to preserve the original image quite easily. This is in line with our expectations since the domains are identical and the identity loss should trivially remove the ambiguity. The high values in the confusion matrix correspond to the digit class which is very definitive, and the smaller ones correspond to the cases when the digit class is somewhat ambiguous. It can be for instance digit \\u20182\\u2019 which is written a bit like \\u20186\\u2019 with a closed loop in the bottom, and it happens with digits '3', '5', '8' as well. \\n\\nFollowing your feedback, we have added some additional experiments on the BRATS 2015 dataset with medical images. In this set of experiments we show how the loss and PSNR (since we have a ground truth available) change when we vary the weight for identity loss, and we compare these values with flipped version of the image. We see that in the absence of identity loss the final CycleGAN loss is very similar for both original and flipped network output, while the PSNR drops significantly for the flipped version. Increasing the identity loss weight does not always result in improved performance in terms of PSNR.\\n\\nWe have also added additional discussion of some newer image-to-image translation models from the 'automorphism point of view', and we hope that some of the questions we pose can be answered in future work.\"}", "{\"title\": \"Thank you for your feedback!\", \"comment\": \"We kindly thank for your feedback!\\n\\nWe have added some additional experiments on BRATS 2015 dataset to expand the experimental section, and provided an additional discussion of some newer multimodal image-to-image translation models from the 'automorphism point of view'.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This is an interesting, timely study. CycleGAN has attracted a lot of attention in unpaired image-to-image translation. Although the basic idea of CycleGAN seem sensible, its precise behavior is not totally clear--can one really avoid mismatch with CycleGAN? Do we need additional constraints? This paper provides a nice answer the the first question.\\n\\nOverall I enjoyed reading this paper. The addressed issue is important, the investigation is reasonable, and the results are intuitive and plausible, with clear practical implications. I think it is a good paper.\\n\\n I acknowledge I read the authors' response and other reviews and would like to keep my original rating.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": [\"This paper focuses on CycleGAN method to show theoretically when the exact solution space is invariant with respect to authomorphisms of the underlying probability spaces for unpaired image-to-image translation.\", \"The paper provides interesting theoretical results on identifying conditions under which CycleGAN admits nontrivial symmetries and has a natural structure of a principal homogeneous space. Proposition 2.1 provides interesting insights into the invariance of the kernel space.\", \"Propositions 2.5 and 2.6 are interesting in that they show that the existence of authomorphisms can worsen the performance of CycleGAN, however, it is unclear that in practice, how could one verify the conditions efficiently before applying CycleGAN.\", \"The experimental results are interesting, however, they are very limited. Having a toy experiment is a good sanity check, but it would be more interesting to see the performance on a real-world applications, such as medical images or other use-cases brought in the introduction. Also, more discussion on the results provided in Fig 3, confusion matrices, would be very helpful. Are there any intuitions behind the large and low values in the table? It could be interesting to see what are the effects of other parameters such as alpha in producing the results.\", \"Overall this paper presents interesting results regarding the theory of CycleGANs, however, the numerical results are very limited, and do not justify the motivations discussed in the introduction and the abstract. Moreover, although the paper introduces novel attempts and theoretically analyzing the CycleGAN, the scope of the work seems to be limited, and thus, it does not have a sufficient significance to be published in ICLR. I strongly suggest the authors to expand and provide more experimental evaluations.\", \"** update:\", \"Thanks for your comments! I found the additional experiment useful and better aligned with the purpose of the model. The discussion added clarified the confusion about the automorphism. That is why I decided to change my score.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"I have read the rebuttal of the authors . Thank you for you answer and for addressing some concerns. While the question addressed is important, the theory presented here does not seem to hint to a solution, hence I am keeping my score.\\n\\n###\", \"summary_of_the_paper\": \"This paper shows that the cycle GAN loss suffers from the presence of lot of symmetries that make the existence of a unique solution not possible , and moreover adding a regularizer that uses the identity loss is not enough to make the problem less prone to those invariances.\", \"review_of_the_paper\": \"The notations and the formalism in the paper are heavy and cumbersome and don't come with any surprising result, since the transforms between unpaired spaces will be found always up to symmetries since we have the composition of one map with another. The use of the identity loss is also shown to not to help either in fixing this invariance issue. \\n\\nExperiments are not interesting since without any structure on the map of F and G , the source domain and the target domain, one is expected to get permutations.\\n\\nThe paper points in the conclusion that the use of skipconnection in F and G has the major influence.\\n\\n The study of cycle GAN might need some assessment of what is the mutual information between the domains , as on what information needs be preserved , and information needs to match , skip connection maintain the content in image generation as the information is kept from lower layer and its modified to target the style of the target images. \\n\\nAn information theoretic analysis of cycle gan is needed using for example the objective of \\\"MISO: Mutual Information Loss with Stochastic Style Representations for Multimodal Image-to-Image Translation\\\". \\nor by using a radically new approach for cycle gan such as the Gromov Wasserstein distance as done in \\\" Learning Generative Models Across Incomparable Spaces\\\"\"}" ] }

[ "Tao Lin", "Sebastian U. Stich", "Kumar Kshitij Patel", "Martin Jaggi" ]

[ "local sgd", "large", "sgd", "stochastic gradient methods", "state", "art", "distributed training", "deep neural networks", "drastic increases", "sizes" ]

{ "note_id": [ "LF2Cb7DMZq", "HklP6PCtiH", "HJx9y-jOsS", "SJffksdsr", "rkeltAcOir", "r1gp2D8yqS", "r1xplIO6YS", "H1lcegDaKr", "S1g1nOZDKr", "rylU6gfSYB", "SkeiU0ZHFH", "H1eGZobrYS", "H1e8UK-SKS", "ryxvs7ZBFS", "rkxB3y-BKH", "HkxrQ7yHKH", "rJgWOs2EYS", "B1xFASpXFS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_comment", "comment", "official_comment", "official_comment", "comment" ], "note_created": [ 1576798732590, 1573672895423, 1573593314269, 1573592841516, 1573592695534, 1571936181421, 1571812853449, 1571807217817, 1571391654993, 1571262654068, 1571262034928, 1571261177686, 1571260750199, 1571259294562, 1571258284828, 1571250972819, 1571240809319, 1571177937218 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1789/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1789/Authors" ], [ "ICLR.cc/2020/Conference/Paper1789/Authors" ], [ "ICLR.cc/2020/Conference/Paper1789/Authors" ], [ "ICLR.cc/2020/Conference/Paper1789/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1789/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1789/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1789/Authors" ], [ "ICLR.cc/2020/Conference/Paper1789/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1789/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1789/Authors" ], [ "~Stone_Jamess1" ], [ "ICLR.cc/2020/Conference/Paper1789/AnonReviewer1" ], [ "~Stone_Jamess1" ], [ "ICLR.cc/2020/Conference/Paper1789/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1789/Authors" ], [ "~Stone_Jamess1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The authors propose a simple modification of local SGD for parallel training, starting with standard SGD and then switching to local SGD. The resulting method provides good results and makes a practical contribution. Please carefully account for reviewer comments in future revisions.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Author's feedback sounds reasonable\", \"comment\": \"Since the other reviewers all give positive reviews, and I do not have strong reasons to doubt the empirical results after reading the author's feedback, I increase the score.\"}", "{\"title\": \"Response to Reviewer1\", \"comment\": \"Thank you for your review. Please check our responses below.\\n\\n[We correctly finetune the learning rate for mini-batch SGD]\\nWe would like to clarify that our choices of learning rates are in favor of the baselines, not our method. We have carefully fine-tuned the learning rate for all mini-batch SGD baselines, while we rely on the \\u2018linear scaling rule\\u2019 for (post)-local SGD. As mentioned in the paragraph \\u2018The procedure of fine-tuning\\u2019 of Appendix A.4.1, instead of using \\u2018linear scaling\\u2019 [2] or \\u2018square root scaling\\u2019 [1] to automatically determine the scaling factor, we did perform grid search for the optimal scaling factor for all mini-batch SGD baselines (this grid search comprises linear scaling and square root scaling as special cases).\\n\\nWe believe that these clarifications give sufficient evidence to show that your concern raised on the \\u201cquestionable empirical results\\u201d, and specifically the learning rates in mini-batch SGD, is without cause. As this was one of your main concerns in your review, we would therefore like to ask you if you could reconsider your score?\\n\\n[Novelty of Algorithms]\\nReviewers 2 and 3 seem to perceive the simplicity and ease-of-use as an advantage of our method. We show that these known algorithmic building blocks, if combined correctly, can lead to remarkable improvements in generalization error, and at the same time alleviate problems in large batch training, and provide a comprehensive empirical study of the involved trade-offs.\\n\\n[Theoretical analysis of generalization performance]\\nWe agree that a theoretical proof of improved generalization ability of Local SGD variants over large batch SGD would be an impressive result. However, we are not aware of any such results even in the single machine case. Generalization results are typically much harder to prove than optimization results in the non-convex setting. The analysis would have to depend on batch size and learning rate, in order to be useful for the large batch setting. And most results which do, don\\u2019t tell anything useful about the generalization of deep neural networks. Thus we believe that at this point, we can only talk about intuitive explanations (such as injecting noise) as opposed to \\u201cwhy Local SGD works?\\u201d (or large-batch doesn\\u2019t). \\n\\nMoreover, our work can provide new insights for the theoretical investigation of the local SGD. Please check our first item in our response to Reviewer2.\\n\\n\\n[2] Goyal, et al. \\\"Accurate, large minibatch SGD: Training imagenet in 1 hour.\\\" arXiv preprint arXiv:1706.02677 (2017).\"}", "{\"title\": \"Response to Reviewer3\", \"comment\": \"Thank you for your review. We\\u2019ve updated the draft and answer your concerns below:\\n\\n[Scalability, and generalize post-local SGD with other compression techniques]\\nAbout scalability improvements over mini-batch SGD, Figure 2 does in fact show the transition between fast communication to the setting when communication does become a limiting factor, by giving complete time-to-accuracy numbers including communication.\\n\\nGradient compression techniques are indeed orthogonal to our approach. In contrast to gradient compression, local SGD does change the trade-off between the number of communications and back-propagations, which can be crucial in high latency scenarios. Nevertheless, in the updated submission, we include a table (Table 3) of new results in Section 4.2 to demonstrate the compatibility of post-local SGD w.r.t. this aspect. Post-local SGD can be combined with the SOTA sign-based compression techniques [3, 4], not only addressing the quality loss issues introduced by the compressed communication, but also further improve the scalability. We will try to add more results to this aspect.\\n\\n[Experiments on language models]\\nSome preliminary results for training 3-layer LSTMs for language modeling on WikiText2 are included in Table 12 (Appendix C.5.2), where we show that post-local SGD can still improve large-batch SGD without any hyper-parameter tuning. \\n\\nGeneralizing post-local SGD to Adam for other NLP tasks is interesting but out of the scope of the current paper. We focus on addressing the generalization issue of deep learning, where mini-batch SGD with momentum is the SOTA optimizer and adaptive methods have not been equally successful for generalization [5]. \\n\\n[The theoretical contributions]\\nPlease check our response to Reviewer1 and Reviewer2 for more details.\\n\\n\\n[3] Bernstein, et al. \\\"signSGD: Compressed optimisation for non-convex problems.\\\" ICML, 2018.\\n[4] Karimireddy, et al. \\\"Error feedback fixes signSGD and other gradient compression schemes.\\\" ICML, 2019.\\n[5] Wilson, et al. \\\"The marginal value of adaptive gradient methods in machine learning.\\\" Advances in Neural Information Processing Systems. 2017.\"}", "{\"title\": \"Response to Reviewer2\", \"comment\": \"Thank you for your review. We answer your specific questions below.\\n\\n[New insights for theoretical investigation of local SGD]\\nThe established theoretical work for local SGD considers convex loss or non-convex training loss, but none of these work discusses the generalization issues. Even the prior theoretical work cannot explain the failure of large-batch SGD for non-convex deep learning. \\n\\nIn that light, it is a far-fetched claim that Local SGD is well understood in the non-convex case. In fact, compared to the local SGD variant in [1] who borrowed the theory from the optimization analysis, post-local SGD is an opposite algorithm based on the understanding of generalization in deep learning.\\n\\nOur empirical experiments and superior performance provide new valuable insights for the community to better understand the convergence analysis and deep learning generalization theory.\\n\\n[Post-local SGD is a simple plugin w/o involving heavy tuning]\\nWe performed extensive ablation studies and provided practical guidelines for post-local SGD.\\nOur algorithm does not have any additional hyperparameters when compared to local SGD. As in local SGD, the only hyper-parameter which needs to be tuned is the number of local update steps H, which (e.g. illustrated in Figure 4) determines (1) how much better the post-local SGD can be over the mini-batch SGD, and (2) how much communication cost can be reduced. \\n\\nBelow we summarize the detailed evidence to justify the effectiveness of post-local SGD.\\n1. In all our evaluations, we fixed the local batch size B_loc to the value reported in the literature.\\n2. Figure 4 studies how different numbers of workers, and different numbers of local update steps, will impact the generalization performance. Post-local SGD in general improves over mini-batch SGD.\\n3. We provide extensive ablation studies on the switching time of post-local SGD in the Appendix (Figures 11, 12, 13). For our post-local SGD evaluations, we turn on post-local SGD after the first learning rate decay, for better generalization performance and communication efficiency.\\n\\n[The meaning of the dominant eigenvalue of the Hessian]\\nThe eigenvalues of the Hessian of the loss characterize the local curvature of the loss. Some techniques are developed to better understand the landscape of the loss surface, as well as the generalization properties [2, 3, 4]. Borrowing the explanation from [3]: \\u201cthe spectrum of the Hessian is composed of two parts: (1) the bulk centered near zero, (2) and outliers away from the bulk which appears to depend on the data\\u201d. It implies most directions in the weight space are flat, and leads to little or no change in the loss value, except for the directions of eigenvectors that correspond to the large eigenvalues of the Hessian.\\n\\nFor more details, please check [3, 4].\\n\\n\\n[1] Wang, et al. \\\"Adaptive communication strategies to achieve the best error-runtime trade-off in local-update SGD.\\\" SysML, 2019.\\n[2] Yao, et al. \\\"Hessian-based analysis of large batch training and robustness to adversaries.\\\" NeurIPS. 2018.\\n[3] Sagun, et al. \\\"Empirical analysis of the Hessian of over-parametrized neural networks.\\\" ICLR workshop, 2018.\\n[4] Ghorbani, et al. \\\"An investigation into neural net optimization via hessian eigenvalue density.\\\" ICML 2019.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a new distributed computation technique for SGD training of deep neural networks. The proposed method is a modification of the local SGD which updates models distributed to several workers in a parallel way and synchronize the model parameters at every few epochs. The local SGD shows a nice performance but it is not robust against large mini-batch size. The proposed method is called post-local SGD that starts the local SGD after some epochs of standard mini-batch SGD. This modification makes the local SGD robust against the large mini-batch size. The authors conducted thorough experiments to investigate the performance of the proposed method. The experiments reveal that the proposed method gives better performances than the mini-batch SGD and the vanilla local SGD.\", \"pros\": [\"As far as I checked, the numerical experiments are strong. They checked several points of view. Several settings of mini-batch sizes and number of workers are compared. For more detailed points, they compared the proposed method for not only the vanilla SGD but also other optimizers such as momentum SGD. Moreover, different choices of timing of starting local SGD (t') are compared.\", \"The proposed method is simple and easy to implement.\"], \"cons\": [\"The local SGD itself is not new and has been already proposed (indeed, the authors are also explaining this point in the paper), and theoretical investigation of local SGD has also been well exploited. This method is just a combination of normal mini-batch SGD and local SGD. In that sense, there are not a so much eye-opening idea in the proposed method.\", \"SGLD interpretation is instructive, but it does not explain why the \\\"post\\\"-local SGD outperforms the local SGD. We can intuitively guess why post-local SGD is better but it is not a rigorous theory.\", \"The post local SGD algorithm has more freedom than existing method, which results in more tuning parameters. There are parameters such as the batch-size before and after switching, the number of workers, the timing of switching and so on. It could be disadvantage of the proposed method.\"], \"minor_comment\": [\"Please give the precise meaning of the dominant eigenvalue.\", \"I think the setting that the data are distributed to all workers is acceptable. It is also an interesting setting from the HPC point of view (on the other hand, it would be less interesting in terms of federated learning).\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a variant of local SGD, post-local SGD, for distributed training of deep neural networks. It targets to mitigate the generalization gap caused by large batch training. The idea is straightforward and easy to understand-- start the training with standard mini-batch SGD and later switch to local SGD. The rationale behind this scheme is that switching to local SGD helps the training converge to flatter minima compared to using large-batch SGD, which correlates with sharper minima, and that helps close the generation gap. Switching to local SGD at the second phase also helps improve communication efficiency by reducing the amortized communication volume. The authors perform empirical studies using ResNet and DenseNet to conclude that post-local SGD outperforms large-batch SGD in terms of generalization performance while also with improved communication efficiency.\", \"strengths\": [\"The post-local SGD technique is simple yet seems to be useful in practice.\", \"Provide a thorough evaluation of the communication efficiency and generalization performance of local SGD.\", \"Introduce a hierarchical version of post-local SGD to better adapt to the topology of the GPU cluster, which often consists of heterogeneous interconnects.\"], \"weaknesses\": \"- The design and experiments are largely empirical without theoretical derivation.\\n- It is less clear about the benefit of post-local SGD when applied to ADAM, which is widely used for distributed training of NLP tasks. \\n- Scalability improvements over mini-batch SGD are largely done by ignoring other optimization techniques that also reduce the communication volume, such as gradient compression[1], Terngrad[2].\\n\\nOverall, the post-local SGD proposed by the paper seems to be a promising technique for large-scale distributed training. The motivation and explanation of the work are clear. My major concern is about the generalizability of this work. ResNet50 is not that interesting from a distributed training perspective. It is less clear whether the performance gains are consistent across tasks. The authors are encouraged to report experimental results on distributed training of large LM models.\\n\\n[1]\\\"Deep Gradient Compression: Reducing the Communication Bandwidth for Distributed Training\\\", Lin et. al., ICLR 2018\\n[2]\\\"Terngrad: Ternary gradients to reduce communication in distributed deep learning\\\", Wen et. al., NeurIPS 2017\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"In this paper, the authors propose a variant of local SGD: post-local SGD, which improves the generalization performance compared to large-batch SGD. This paper also empirically studies the trade-off between communication efficiency and performance. Additionally, this paper proposes hierarchical local SGD. The paper is well-written, the experiments show good performance.\\n\\nHowever, there are several weakness in this paper:\\n\\n1. The post-local SGD is a simple extension of local SGD. Roughly speaking, post-local SGD uses fully synchronous SGD to warm up local SGD. The novelty of this algorithm is limited.\\n\\n2. The main statement that post-local SGD improves the generalization performance, is only supported by empirical results. No theoretical analysis is provided. Thus, the contribution of this paper is limited.\\n\\n3. In this paper, it is reported that in some experiments, local SGD significantly outperforms mini-batch SGD. As shown in Figure 3, when the number of workers is large enough, mini-batch SGD has extremely bad performance. However, such bad result is potentially caused by a bad choice of learning rates. In this paper, the authors use \\\"linearly scaling the learning rate w.r.t. the global mini-batch size\\\". However, some recent papers also suggest using square root scaling instead of linear scaling [1]. I think the mini-batch SGD fails simply because the learning rates are too large. However, the authors claim that the learning rates for mini-batch SGD are fine-tuned (in Figure 3), which makes the empirical results questionable.\\n\\n\\n-----------\\nReferences\\n\\n[1] You, Yang, et al. \\\"Large batch optimization for deep learning: Training bert in 76 minutes.\\\" arXiv preprint arXiv:1904.00962 (2019).\\n\\n===================\\nUpdate after author's feedback:\\n\\nI do not have strong reasons to doubt the empirical results after reading the author's feedback, so I increase the score.\"}", "{\"comment\": \"We acknowledge the theoretical contribution of the work [2,3] and will add it to the related work - however the contribution of our empirical paper is completely orthogonal to them. We focus on the generalization issue for *large-batch training*, an issue which is still unsolved even for iid data but of strong importance for distributed deep learning. We believe our contribution is significant in this aspect. [2,3] primarily focus on the optimization aspect in terms of theoretical insights. The global mini-batch size in their experiments is up to 512 for CIFAR-10 and 1024 for ImageNet; while ours reaches 4096 for CIFAR and 16K for ImageNet, focusing on the generalization challenge due to large mini-batch sizes.\\n\\nOn a side-note, we would like to clarify that only [3] proves a convergence rate for both the iid and non-iid setup. The convergence proof in [2] relies on bounded gradients (Assumption 1) and iid data (Assumption 2) with might be difficult to generalize to the non-iid setting, as argued in a recent preprint [5].\\n\\n[5] A. Khaled, K. Mishchenko, P. Richt\\u00e1rik, First Analysis of Local GD on Heterogeneous Data, arxiv.org/abs/1909.04715\", \"title\": \"Orthogonal contributions\"}", "{\"comment\": \"I agree that federated learning (FL) is beyond the scope of this paper.\\n\\nHowever, I think [2,3] are local SGD, not FL.\\n\\nIn my opinion, compared to local SGD, FL has the following major differences (and difficulties in theoretical analysis):\\n1. subsampled workers: only a subset of workers participates the training in each epoch\\n2. heterogeneous number of local steps: in local SGD, the number of local steps (H in your paper) must be the same across different workers, so that we can use an auxiliary (fake-average) variable to analyze the convergence. However, for federated learning, different workers can have different H, which makes it difficult for theoretical analysis.\\n\\nDecentralized/non-IID data is just another standard setting of traditional distributed SGD.\\nNon-IID data itself does not make [2,3] FL.\", \"title\": \"Federated learning is different from local SGD\"}", "{\"comment\": \"As I mentioned, in practice, random sampling without replacement is preferred.\\nReshuffling the indices at the beginning of each epoch and then sequentially load the data samples according to the shuffled indices is the easiest way to implement sampling without replacement.\\n\\nSince it is \\\"without replacement\\\", the same data sample should not be processed/sampled twice by different workers in the same epoch.\\nIn a single epoch, although a worker have access to the data of the other workers, it won't use it. \\nOtherwise, there will be some extra computation.\\n\\nAnother point of view is to make a fair comparison between local SGD and fully synchronous SGD. If we get the random seeds fixed, the workers in these 2 algorithms should process the same sequences of data. The only difference is that local SGD skips some rounds of synchronization. \\nYes, it is possible for different workers to share some data in the same epoch in local SGD, however, that will make it unfair when compared to fully synchronous SGD.\", \"title\": \"Reshuffling is just a way to implement random sampling without replacement\"}", "{\"comment\": \"Thank you for the thoughtful comments. We agree that the decentralized/non-IID setting is important and much more difficult.\\n\\nThe local SGD algorithm in that setting is also known as federated averaging [FA], and there is a very extensive recent literature on it, including the two papers you mention. The federated learning setting is beyond the scope of our paper here as it will introduce additional generalization challenges beyond the large-batch scenario of our interest here. Nevertheless we'll include both mentioned papers in the revised version.\\n\\n[FA] McMahan, B., Moore, E., Ramage, D., Hampson, S., & y Arcas, B. A. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data. AISTATS 2017\", \"title\": \"Federated learning is beyond the scope of this paper\"}", "{\"comment\": \"Thanks for your quick response.\\n\\nThe goal of this paper is to reduce the communication cost, so it will not aggregate the gradients for every iteration. So for each node, it can only get information from other data through the gradient aggregation. \\n\\nSince each user can get access to other data, why we confine each node reshuffle the data every epoch? The setting here allows each node get access to all data, then it is no necessary for other nodes to get information from gradients only. It is my understanding from the setting of this paper. \\n\\nThanks.\", \"title\": \"Thanks for your reply\"}", "{\"comment\": \"In practice, it is observed that random permutation/random sampling without replacement usually converges faster than sampling with replacement [4] (there should be some better references but I don't remember...).\\nThat's why we usually use random permutation in experiments. If you are familiar with the dataloader of pytorch, you will see that it is also implemented by random permutation.\\n\\nBTW, I don't really understand why \\\" each user sample a batch of data from the entire dataset and performs the computation. I think it will make the computation more faster since each user doesn't need to be confined for a disjoint subset.\\\"\\nHere you mean \\\"sampling with replacement is faster than sampling without replacement in computation\\\"? why is that true?\\n\\n\\n------\\n[4] Ying, Bicheng, et al. \\\"On the performance of random reshuffling in stochastic learning.\\\" 2017 Information Theory and Applications Workshop (ITA). IEEE, 2017.\", \"title\": \"In practice, random permutation usually gives better convergence\"}", "{\"comment\": \"Thanks for all of your responses.\\nI agree that the approach cannot be used for distributed mobile users, so it weakens the contribution. \\n\\nIn addition, I have another concern that each node has access to the full dataset and you just need to shuffle the index. Consider the following case, in every iteration, each user sample a batch of data from the entire dataset and performs the computation. I think it will make the computation more faster since each user doesn't need to be confined for a disjoint subset. From this point, I don't see the benefit of the disjoint setting. \\nFurthermore, when we do the mini-batch setting for single GPU training, we also don't have the constraint each sample cannot be used twice since the mini-batch is randomly selected.\\n\\nThanks again for all of your response.\", \"title\": \"Thanks for your replies\"}", "{\"comment\": \"First, I agree with the authors that the experimental setup is standard and correct.\\nBasically, each epoch is a full pass of the entire dataset, and the same data sample won't be processed twice in the same epoch.\\n\\nThis paper takes the assumption that each worker has the access to the entire dataset.\\n\\nHowever, typically, decentralized/non-IID setting is considered more difficult. \\nFurthermore, when the dataset is extremely large, it is reasonable to disjointly partition the dataset to the workers and never reshuffle and re-partition it.\\nNote that the previous comment never claims that the experimental setup is incorrect.\\nAnd, since there are other papers of local SGD [2,3] that assume decentralized/non-IID local datasets on the workers, the contribution of this paper is potentially weakened.\\n\\n----------\\n[2] Yu, Hao, Sen Yang, and Shenghuo Zhu. \\\"Parallel restarted SGD with faster convergence and less communication: Demystifying why model averaging works for deep learning.\\\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019.\\n[3] Yu, Hao, Rong Jin, and Sen Yang. \\\"On the Linear Speedup Analysis of Communication Efficient Momentum SGD for Distributed Non-Convex Optimization.\\\" International Conference on Machine Learning. 2019.\", \"title\": \"Experimental setup is correct, though the IID setting is a little bit weak.\"}", "{\"comment\": \"Thank you for your interest. The setting in which data is kept locally to the workers is normally called 'decentralized'. Note that the \\u2018distributed\\u2019 setting does not always have to be \\u2018decentralized\\u2019.\\n\\nWe consider the 'distributed' (data center) setting where each worker has fast access to all data points. This is a standard setup and very important in deep learning. See for instance [1], the current state-of-the-art baseline for large-batch training in the datacenter.\\n\\nWe use the formulation 'disjointly partitioned' to indicate that data points are used only once (across workers) in each epoch. To achieve this, we don't reshuffle the data points themselves. We just shuffle the data indices with a shared random seed across workers at the beginning of each epoch. Each worker then takes a disjoint part of the data indices to generate local mini-batches.\\n\\nWe hope this resolves your concern.\\n\\n------\\n[1] Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour, https://arxiv.org/abs/1706.02677\", \"title\": \"We believe the experimental setup is standard and correct\"}", "{\"comment\": \"I think the author uses the distributed deep learning is misleading.\\n\\nIn the experiment part, the author say \\\" The data is disjointly partitioned and reshuffled globally every epoch. \\\". So it violates the setting in most distributed deep learning that data is kept locally at each node. \\nConsider the case, if the data is shuffled many times, then each node will get access to the full dataset, so it weakens the experimental part. In fact, in many real applications, it is challenging and time consuming to reshuffle the data again and again.\", \"title\": \"Use distributed deep learning is misleading\"}" ] }

[ "Francesco Croce", "Matthias Hein" ]

[ "adversarial robustness", "provable guarantees" ]

{ "note_id": [ "jOnPym6Ro9", "rJlbXqZsjS", "B1x6FtZisB", "BJlX4K-jjB", "BylLN_bojr", "rkeaaOD7qS", "BygCdma6FH", "ryegXTSpYr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732557, 1573751321150, 1573751172891, 1573751083030, 1573750830473, 1572202692935, 1571832694038, 1571802392073 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1788/Authors" ], [ "ICLR.cc/2020/Conference/Paper1788/Authors" ], [ "ICLR.cc/2020/Conference/Paper1788/Authors" ], [ "ICLR.cc/2020/Conference/Paper1788/Authors" ], [ "ICLR.cc/2020/Conference/Paper1788/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1788/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1788/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper extends the degree to which ReLU networks can be provably resistant to a broader class of adversarial attacks using a MMR-Universal regularization scheme. In particular, the first provably robust model in terms of lp norm perturbations is developed, where robustness holds with respect to *any* p greater than or equal to one (as opposed to prior work that may only apply to specific lp-norm perturbations).\\n\\nWhile I support accepting this paper based on the strong reviews and significant technical contribution, one potential drawback is the lack of empirical tests with a broader cohort of representative CNN architectures (as pointed out by R1). In this regard, the rebuttal promises that additional experiments with larger models will be added in the future to the final version, but obviously such results cannot be used to evaluate performance at this time.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Answer to reviewer 1\", \"comment\": \"We thank the reviewer for the encouraging comments. We address below the questions:\\n\\n1) \\\"``I am very curious about the landscape or decision boundary analysis and visualization\\\"\\n\\nIn general, MMR-Universal significantly expands the regions where the classifier is affine, similarly to what MMR does. The difference is that MMR-Universal enlarges the linear regions so that all the $l_p$-balls for $p\\\\geq 1$ are contained. Since the decision boundary inside each linear region consists in a hyperplane and the linear regions are especially wide, the decision boundary is likely to be the union of portions of a few hyperplanes. The problem however is that this effect is in $2$ dimensions relatively limited as the differences between $l_p$-balls become much more pronounced with increasing input dimension $d$. In this short amount of time we could not yet come up with a good example in $2$ dimensions resp. with an appropriate projection of the case in higher dimensions. \\nWe agree with the reviewer that such an illustration will further provide intuition and is thus helpful and we will try hard to come up with this for the final version.\\n\\n2) \\\"``Try a few more classical CNN architectures\\\"\\n\\nWe will train for the final version the usual LeNet architecture and the large network from (Wong et al, 2018), as it is a standard choice for provable robustness, which is roughly 4 times larger than our current one. Moreover, we will extend the implementation of our method to handle residual connections, which do not make any difference to the algorithm from a theoretical point of view. At the moment we are running low on GPU resources due to the upcoming CVPR deadline - we are sorry that we have to postpone the additional results for larger models to the final version.\"}", "{\"title\": \"Typo\", \"comment\": \"Sorry, it should $\\\\|x\\\\|_1$ in the answer to 6)\"}", "{\"title\": \"Answer to Reviewer 2\", \"comment\": \"We thank the reviewer for the encouraging comments. We address the raised questions individually.\\n\\n1) \\\"The proposed regularization scheme does not come with any explicit robustness guarantee. Is it possible to show that a model that minimizes the regularized loss is guaranteed to be $l_p$ robust for some radius?\\\"\\n\\nIf the MMR-Universal regularization (stated in page 6) is successfully minimized during training and reaches zero for a particular training point, then for this point we have achieved robustness of $\\\\gamma_1$ wrt $l_1$ and $\\\\gamma_\\\\infty$ of $l_\\\\infty$, since no hyperplane, either region or decision boundary, is closer than those thresholds to any of the points of the training set (please see Theorem 2.1).\\nTheorem 4.1 provides then for any $p\\\\in(1, \\\\infty)$ in Equation (7) the radius of the $l_p$-ball on which the decision does not change and is thus robust. In practice, on the test points some hyperplanes happen to be closer than the given thresholds even after training, so that the guarantees from Theorem 4.1 are smaller than the desired ones. However, since MMR-Universal expands also neighboring linear regions (where there are no training points), we can certify with the mixed-integer programming certification of (Tjeng et al, 2019) or the approach of (Wong and Kolter, 2018) $l_p$-balls with larger radii, achieving good provable robustness. As for all current certification methods we observe that the enforced robustness on the training set generalizes to the test set.\\n\\n2) \\\"In Appendix C.3, it is mentioned that the best values for $\\\\gamma_p$ were empirically found to be 1-2 times larger than the desired $\\\\epsilon_p$ robustness. Could this be formalized in theory?\\\"\\n\\nThe effectiveness of larger $\\\\gamma$ is an empirical observation, see also e.g. (Gowal et al, 2018). In practice, we use $\\\\gamma > \\\\epsilon$ in order to counter the fact that potentially the training is able to push the hyperplanes only partially, and not up to the desired threshold. It is an interesting question if one can \\nback up this procedure with theoretical arguments.\\n\\n3) \\\"From Figure 2, it seems that there is an optimal ratio of $l_1$ to $l_\\\\infty$ robustness for which the $l_2$ guarantee is maximized. It would be interesting to see what this optimal ratio is for maximizing the $l_p$ guarantee, as a function of dimension d and p.\\\"\\n\\nThanks for this interesting question. We added in the Appendix A.3 the derivation of the optimal ratio as a function of $d$ for $p=2$, as well as the corresponding maximal improvement of the bounds given by the convex hull over those obtained considering only the union of the $l_1$ and $l_\\\\infty$ balls. It turns out that the gain of the $l_2$-robustness guarantee of the convex hull over to the union of the $l_1$-and $l_\\\\infty$-ball increases with dimension $d$ as $d^\\\\frac{1}{4}$. For $p\\\\neq 2$, it needs to be approximated numerically.\\n\\n4) \\\"In the experiments, the MMR-Universal scheme is compared with MMR+AT-$l_2$ and MMR+AT-$l_\\\\infty$. It would be interesting to compare it with MMR-$l_1$+AT-$l_\\\\infty$ or with MMR-$l_\\\\infty$+AT-$l_1$.\\\"\\n\\nWe have started the training and the evaluation of such models of MNIST and report them in Appendix C.5. We will add the other datasets in the final version. The preliminary analysis on MNIST suggests that the models have good provable robustness only in the norm for which MMR is used. Moreover, the model trained with MMR-$l_1$+AT-$l_\\\\infty$ shows non-trivial provable robustness also wrt $l_2$ (as it happens sometimes in Table 4), but comes with no guarantees wrt $l_\\\\infty$. In particular, we achieve no non-trivial robustness for the union of the $l_1$, $l_2$, $l_\\\\infty$-balls. This stresses again that enforcing provable robustness wrt\\n$l_1$ and $l_\\\\infty$ as in MMR-Universal is the key for provable robustness wrt intermediate $l_p$-balls.\\n\\n5) \\\"In eq(1), is c intended to be the true label at x?\\\"\\n\\nYes, thanks for noticing this. We have fixed this.\\n\\n6) \\\"In Figure 2, the red curves are missing in the first and third plots.\\\"\\n\\nThe red curve is hidden by the green one, as they almost everywhere overlap and they are difficult to distinguish. The red curves can be seen for values of $\\\\norm{x}_1$ close to 1, zooming into the plots. We have added a note on this in the caption.\"}", "{\"title\": \"Answer to Reviewer 3\", \"comment\": \"We thank the reviewer for the detailed comments. We have addressed the suggested changes:\\n\\n1) \\\"The author should provide more empirical analysis on the MMR regularization, like how it changes during the training process.\\\" \\n\\nWe have added in the Appendix C.4 plots showing the evolution of test error and upper bounds on the robust $l_p$-test error for $p \\\\in \\\\{1,2,\\\\infty\\\\}$ evaluated using the approach of (Wong and Kolter, 2018) together with an upper bound on the robust test error over the union of these $l_p$-balls during training using our regularizer MMR-Universal for all four datasets. For all datasets we see improving test error as well as robustness with increasing number of epochs. Please note that the reported results differ from the ones reported in the main paper as we had to retrain the models to do the evaluation during training but most importantly due to the fact that we do not use the expensive mixed-integer programming of approach (Tjeng et al, 2019) for $l_\\\\infty$ which would improve the $l_\\\\infty$ upper bounds on the robustness quite significantly but is very slow so that a full evaluation every 10 epochs or less is quite time consuming (please note that we are running a bit low on computational resources due to the upcoming CVPR deadline). We will include the evaluation with MIP in the final version.\\n\\n\\n2) \\\"The captions of the table and figure are too small.\\\" \\n\\nWe have increased the size of the captions for better readability.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\nThe author proposed MMR regularization for the provable robustness of union of l-1 and l-infty balls, which is robust to any l-p norm for p>=1.\", \"strengths\": \"1. The paper is well organized. \\n2. The theoretical part is completed and experiments are done on MNIST, Fashion-MNIST, GTS, CIFAR-10. \\n3. The proposed method presents a significant improvement over SOTA.\", \"weakness\": \"1. The author should provide more empirical analysis on the MMR regularization, like how it changes during the training process. \\n2. The captions of the table and figure are too small.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary of the paper's contributions:\\n\\nThis paper proves a result on the l_p robustness (p \\\\neq 1, \\\\infty) of a piecewise affine classifier in terms of its l_1 and l_\\\\infty robustness. This result is based on the insight that a guarantee for l_1 and l_\\\\infty robustness of a piecewise affine classifier also guarantees robustness within the convex hull of the union of the l_1 and l_\\\\infty balls. The paper then proposes a regularization scheme called MMR-Universal for a ReLU neural network that simultaneously trains for l_1 and l_\\\\infty robustness. This scheme is based on maximizing the linear regions of the network as in Croce et al., AISTATS 2019. Using the main result of the paper, it is implied that the proposed regularization scheme also enforces l_p robustness for any p \\\\geq 1.\", \"main_comments\": \"(1) The paper provides an interesting result that guarantees l_p robustness for a piecewise affine classifier based on only l_1 and l_\\\\infty robustness. (2) The experiments show that the proposed regularization scheme is indeed effective in simultaneously guaranteeing robustness with respect to all l_p balls.\", \"detailed_comments\": [\"The proposed regularization scheme does not come with any explicit robustness guarantee. Is it possible to show that a model that minimizes the regularized loss is guaranteed to be l_p robust for some radius? In Appendix C.3, it is mentioned that the best values for \\\\gamma_p were empirically found to be 1-2 times the desired \\\\epsilon_p robustness. Could this be formalized in theory?\", \"From Figure 2, it seems that there is an optimal ratio of l_1 to l_\\\\infty robustness for which the l_2 guarantee is maximized. It would be interesting to see what this optimal ratio is for maximizing the l_p guarantee, as a function of dimension d and p.\", \"In the experiments, the MMR-Universal scheme is compared with MMR+AT-l_2 and MMR+AT-l_\\\\infty. It would be interesting to compare it with MMR-l_1+AT-l_\\\\infty or with MMR-l_\\\\infty+AT-l_1.\", \"In eq(1), is c intended to be the true label at x?\", \"In Figure 2, the red curves are missing in the first and third plots.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Overview:\\n\\nThe paper is dedicated to developing a regularization scheme for the provably robust model. The author proposes the MMR-Universal regularizer for ReLU based networks. It enforces l1 and l infinity robustness and leads to be provably robust with any lp norm attack for p larger than and equal to one.\", \"strength_bullets\": \"1. Using convex hull to enforce robustness, it is very reasonable and straightforward which highly aligns our intuition.\\n2. The author provides detailed and conniving derivations and proof. And the results in table 2 do achieve the state-of-the-art provable robustness.\\n3. It is the first robust regularizer that is able to provide non-trivial robustness guarantees for multiple lp balls. The lp balls don't contain any other. It also is one of the most straightforward methods among all potential similar methods.\", \"weakness_bullets\": \"1. I am very curious about the landscape or decision boundary analysis and visualization. For the author's MMR-Universal regularization, it should give the model a very \\\"good\\\" decision boundary which has clear marginal between any two categories. In my opinion, it is necessary to evidence to convince more readers.\\n2. Try a few more classical CNN architectures.\", \"recommendation\": \"Due to the logical derivations and supportive experiment results, this is a weak accept.\"}" ] }

[ "Łukasz Kaiser", "Mohammad Babaeizadeh", "Piotr Miłos", "Błażej Osiński", "Roy H Campbell", "Konrad Czechowski", "Dumitru Erhan", "Chelsea Finn", "Piotr Kozakowski", "Sergey Levine", "Afroz Mohiuddin", "Ryan Sepassi", "George Tucker", "Henryk Michalewski" ]

[ "reinforcement learning", "model based rl", "video prediction model", "atari" ]

{ "note_id": [ "6-j3TrjYFk", "HJeGC9wFsH", "HJey_9DFjH", "H1lipKPYir", "ByxgJ6Jj5H", "BJxJlLRpFr", "H1lwnmopFH", "SJg3esx3tH", "Byg8y2bUKr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732527, 1573645001640, 1573644903273, 1573644739073, 1572695256323, 1571837414814, 1571824558692, 1571715827720, 1571326942396 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1787/Authors" ], [ "ICLR.cc/2020/Conference/Paper1787/Authors" ], [ "ICLR.cc/2020/Conference/Paper1787/Authors" ], [ "ICLR.cc/2020/Conference/Paper1787/Authors" ], [ "~Kacper_Piotr_Kielak1" ], [ "ICLR.cc/2020/Conference/Paper1787/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1787/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1787/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Spotlight)\", \"comment\": \"This paper presents a model-based RL approach to Atari games based on video prediction. The architecture performs remarkably well with a limited amount of interactions. This is a very significant result on a question that engages many in the research community.\\n\\nReviewers all agree that the paper is good and should be published. There is some disagreement about the novelty of it. However, as one reviewer states, the significance of the results is more important than the novelty. Many conference attendees would like to hear about it.\\n\\nBased on this, I think the paper can be accepted for oral presentation.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Answer to AnonReviewer2\", \"comment\": \"Thank you very much for the time taken to produce this high quality review. We appreciate the comments and have included them in the current version of the paper (in particular we included the mentioned references and made information about the asymptotic performance more visible in the main text, now in Section 6.2).\\n\\nThe aim of this work was to develop model-based methods using visual predictions. As pointed by the reviewer this is a huge gap in the current state-of-the-art. We consider our work to be the first step on a longer path. In particular, we consciously focused on the case of 100k frames, mostly for practical reasons of being computationally more feasible. On top of that, we provided results for other numbers of frames but without much tuning, which we believe would improve the performance. Having said that, most likely at the moment our method would not provide results matching state-of-the-art model free approaches. There are at least two reasons, one of a practical nature is that it is hard to compete with years of the model-free research in one step. The second is more fundamental, it is generally observed that model-based methods rarely compare with model-free ones as to the asymptotic performance [1].\", \"other_comments\": \"* \\u201cshould be also mentioned that there is a huge discrepancy between the maximum performance achieved by the proposed model-based method and the maximum performance achieved by the model-free baselines when training for 50,000,000 steps\\u201d\\nIn the summary of the submitted version of the paper, it was already stated that \\u201cthe final scores are on the whole lower than the best state-of-the-art model-free methods\\u201d. As mentioned above we have now added Sec. 6.2. discussing performance with more frames in the main text to make it more evident.\\n\\n[1] Tingwu Wang et al., Benchmarking Model-Based Reinforcement Learning, https://arxiv.org/abs/1907.02057\"}", "{\"title\": \"Answer to AnonReviewer3\", \"comment\": \"Thank you very much for the detailed review.\\n\\nWe updated the paper to be more clear about the computational cost of our proposed method (conclusion and Appendix C). In short, the computational resources required are higher than running a model-free algorithm (e.g. Rainbow) directly on the ALE environment. We aimed at developing a working model-based algorithm for the well-studied Atari domain. We ignored computational cost, though clearly it is to be addressed in the future. We also think that our method has additional benefits in environments where collecting real world experience is expensive or dangerous, such as robotics or autonomous driving.\\n\\nWe would like to clarify the differences between our work and Ha and Schmidhuber (2018). Their work is clearly important and presents a similar direction of using a world model for RL.\\nThe first difference is in the architecture of world models. We have introduced a novel architecture with a stochastic discrete latent variable, inspired by both VAE and pixel-RNNs; based on our extensive experiments the architecture was crucial for the ALE domain (please refer to the ablation section of our paper).\\nMoreover, we evaluated our approach on a broad set of Atari games coming from a much studied benchmark. These environments present a wide range of difficulties. In some cases, our method worked for games which require non-trivial exploration (as for example Freeway) and thus are particularly challenging for model-based methods (e.g. require the model to be consistent over an extended number of steps). Ha and Schmidhuber are training the world model only based on trajectories from random agents. In the case of ALE, we have verified that this is not enough and that better results are achieved when we repeat the loop of collecting experience, training the world model and training the agent multiple times (see. Figure 6(c) in the Appendix).\\n\\nThank you also for providing the interesting references to work not using deep learning. We have included them in the updated version of the paper.\"}", "{\"title\": \"Answer to AnonReviewer1\", \"comment\": \"We thank the reviewer for the valuable and detailed review.\\n\\nAs the reviewer mentioned, sample efficiency is most important when collecting samples is hard including the application interacting with the physical world. And we also agree with the reviewer that ALE is not a good example of such an environment since collecting new trajectories in ALE is quite cheap. However, we chose ALE to demonstrate the capability of our proposed method in a setting with: 1) relatively complex high dimensional observation space and 2) task variety. We believe this setting demonstrates the generality of the method which potentially can be employed in the real world, where collecting samples is expensive, as well. \\n\\n> - What is the impact on total wall-clock training time when using this approach? Given that the technique is centered on ALE, the characteristics of ALE compared to the learned world model are relevant (ALE executes very quickly and easily parallelizes whereas the learned world model presumably only runs where you have a GPU).\\n\\nThe wall-clock training time of SimPLe is increased over standard model-free training. We\\u2019ve updated the article with clear information about this in conclusions and Appendix C.\\nOur aim was to develop a working model-based RL algorithm of general applicability and decrease sample efficiency in the low-samples regime. Having said that, we chose the Atari domain as a well-established testing ground and consciously ignored the computation issues. In fact the wall-time of our algorithm is higher than using standard-model free. We feel it is the price to be paid for the above mentioned improvements, though admittedly making models faster (light-weight) is a tempting research area!\\nThe world model, in theory, doesn\\u2019t need GPU to be trained and evaluated, but in practice, the runtime would increase even further if run without an accelerator.\\n\\n> - Can this approach be stacked to benefit from training in a lighter-weight approximate model (env'') of the world model (env')?\\n\\nThis is an interesting idea, which we have not explored yet. However, we have observed that simplifications of our world model (e.g. not including stochastic discrete latent) produce significantly weaker predictions and agents trained inside of these world models don\\u2019t transfer well to the real environment.\"}", "{\"title\": \"Comparison with new model-free results and underlining our main contribution\", \"comment\": \"Thank you for pointing us to these articles. They are very interesting pieces of work, and we are happy that our research has \\u201cpartially inspired\\u201d that work [1]. In the final version of the paper, we will include stronger baseline results proposed in [2].\\n\\nMoreover, we would like to point out that:\\n1. In our opinion, even with the model-free results presented in [1] and [2] our method can still be considered state-of-the-art with regards to the sample complexity on Atari. [2] clearly states that out of the 26 games tried there is a tie between two compared methods (SimPLe is better in 13, so is OTRainbow). [1] claims that their optimized version of Rainbow beats SimPLe at 17 out of 26 games, but this is because they are using an outdated version of our result. With the best results presented in this paper (coming from SD Long model), we are again on pair with [1]. Detailed comparison using our current results can be found:\", \"https\": \"//docs.google.com/spreadsheets/d/1uf5C79LeaDZfOwt_4Pm2R1dOXPPTTjM3ijkM95nc3H8/edit?usp=sharing\\nPlease note that the only difference between our older and newer (better) results is the increased training time of the world model, which clearly shows that with further hyperparameter optimization SimPLe could score even better.\\n\\n2. It is no secret that RL methods behave quite differently with various random seeds and hyper parameters. We searched for optimal model-free baseline much more than we searched for optimal hyperparameter set for SimPLe. We believe both can be improved with a more rigorous search. We also would like to mention that the results in [2] are similar to the ones we reported in our paper. SimPLe has a hard time learning in environments that we consider hard and are discussed, in detail, in our analysis such as bankheist, games based on exploration: hero, private eye, james bond.\\n\\n3. Please note that one of the main focuses of our paper and our proposed method is to demonstrate the possibility of using model-based reinforcement learning for ATARI where observation space is huge and tasks vary significantly. We illustrated how such models can be scaled to this kind of problem, something that has not been shown before.\\n\\n\\n[1] https://arxiv.org/abs/1906.05243\\n[2] https://openreview.net/forum?id=Bke9u1HFwB\"}", "{\"comment\": \"Thank you for submitting your study. I really enjoyed reading about your SOTA world model architecture, especially that I was experimenting with MBRL at the time. It would be interesting to see how it impacts the performance of other existing algorithms that employ learnable world models.\\n\\nI'm curious about your thoughts on [1] and [2] though. Both papers conclude that utilizing the model in a way outlined by SimPLe does not provide an improvement over replay-based methods. Both of them employ quite simplistic tuning techniques to achieve at least equal performance to SimPLe in the low-data regime.\\n\\n[1] https://arxiv.org/abs/1906.05243\\n[2] https://openreview.net/forum?id=Bke9u1HFwB\", \"title\": \"What's your opinion on recent MFRL results in the low-data regime?\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper addresses sample-efficient learning (~2 hours of gameplay equivalent) for Atari (ALE) games. Building on the idea of training in a learned world model and the use of a u-net next-frame predictor, the approach is claimed to yield almost comparable performance to other models with only a fraction of the true-environment experience.\\n\\nSample efficiency is a major concern for DRL, particularly with an eye towards robotics and other physical domains. Although the approach is rather specific to the shapes and qualities of data in the ALE setting, the work is motivated at a high level, and the specific techniques for predicting the next frame explained in the past are explained.\\n\\nThis reviewer moves for a weak accept on account that the paper is well written (with quite thorough experiments explaining improvements in sample efficiency and possible limits in final task performance) but specifically targets ALE where execution is so cheap. The total number of PPO updates made in the new approach is not much reduced from before even if the number of trajectories evaluated in the true environment is very much reduced. On the problem of how much RL itself is sample efficient, not much progress is made.\", \"question\": [\"What is the impact on total wall-clock training time when using this approach? Given that the technique is centered on ALE, the characteristics of ALE compared to the learned world model are relevant (ALE executes very quickly and easily parallelizes whereas the learned world model presumably only runs where you have a GPU).\", \"Can this approach be stacked to benefit from training in a lighter-weight approximate model (env'') of the world model (env')?\"]}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper covers the author\\u2019s approach to learning a model of a game, which can then be used to train a reinforcement learning agent. The major benefit of the approach is that instead of requiring millions of training steps in the game, the model can be constructed with only 1-2 hours of footage, and then train in the simulated game for millions of training steps.\\n\\nThis is a well-written paper, and the results are very impressive. The approach builds upon prior work with the same general thrust, but broadly makes clear that it stands above these existing approaches. However, I would have appreciated some clarity in the comparison made to the work of Ha and Schmidhuber (2018). It is unclear if the difference given is just because of the environments employed by Ha and Schmidhuber or if the authors see the approach presented in this paper as fundamentally different or improved in some way. \\n\\nMy one major technical concern comes down to how this work is framed and what that implies about appropriate baselines. The authors state clearly that the benefit of this work is that it can learn a sufficient model of a game in only 1-2 hours of gameplay footage. As I said above that is very impressive. However, the agents then requires 15.2 million interactions in this environment to learn to play the game. I would have appreciated some clarity then in the computational resource cost in this approach as opposed to just training say Rainbow in the actual game environments with 15.2 million interactions. It\\u2019s also not clear if optimizing Rainbow\\u2019s performance on 1M steps is a fair comparison. Ideally I would have liked to have seen some variance in the amount of time Rainbow was trained for compared to the associated computational costs. Clarity on this especially in sections like 6.1 would help readers better grasp the tradeoffs of the approach.\\n\\nThe authors could have also included reference to non-DNN work on learning forward/engine/world models that were then used to play or reason about the game. For example Guzdial and Riedl\\u2019s 2017 \\u201cGame Engine Learning from Gameplay Video\\u201d on Super Mario Bros. or Ersen and Sariel\\u2019s 2015 \\u201cLearning behaviors of and interactions among objects through spatio\\u2013temporal reasoning\\u201d on a novel game.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary\\n\\nThis paper proposes a model-based reinforcement learning algorithm suitable for high-dimensional visual environments like Atari. The algorithmic loop is conceptually simple and comprises 1) collecting real data with the current policy 2) updating an environment model with old and newly acquired data and 3) updating the policy \\\"virtually\\\" inside of the environment model using PPO. The approach is evaluated on 26 games from the Atari benchmark and compared against the model-free baselines Rainbow DQN and PPO. The newly proposed model-based method clearly outperforms both model-free baselines in low training regimes (100,000 steps). Further ablation studies are provided, e.g. similar results are obtained in a more stochastic setting of ALE with sticky actions.\\n\\nQuality\\n\\nThis paper has a strong applied focus and needs to be judged based on its experiments. The quality of those are high. The method is evaluated on a suite of 26 games, compared to strong model-free baselines and results are averaged over 5 seeds. One concern I have is that the method is only evaluated in low training regimes. While I do understand that increasing the training horizon is computationally demanding, results in the appendix (Figure 11a) indicate that the proposed model-based method has worse asymptotic performance compared to the model-free baselines. After 500,000 training steps the effect of sample efficiency vanishes and the final performance results are far away from the final performance results of the model-free baselines after 50,000,000 training steps. Also, a plot similar to Figure 11a) from the appendix for Rainbow DQN would be good (but I do understand this might be difficult to obtain in the course of the review period should this require more experiments).\\n\\nClarity\\n\\nThe paper is clearly written and easy to follow. However, the authors could state in the main paper more clearly that their method excels in low training regimes and that the sample efficiency effect seems to vanish when increasing training iterations from 100,000 to 500,000 steps. In fact, Figure 11a) from the appendix should go into the main paper, and it should be also mentioned that there is a huge discrepancy between the maximum performance achieved by the proposed model-based method and the maximum performance achieved by the model-free baselines when training for 50,000,000 steps. Based on the experiments, it is not clear at all if the new method will eventually catch up with best model-free results from the literature when training time is increased, or stall in low-performance regimes indefinitely.\\n\\nOriginality\\n\\nThe originality of this paper is not very high since the proposed algorithm and its components are not novel (there might be some minor novelty in the environment model architecture). However, this paper should not be judged based on its originality but based on its significance. \\n\\nSignificance\\n\\nA working model-based RL algorithm for Atari is clearly a huge gap in the current literature and this paper takes an important step towards this direction. Demonstrating improved sample efficiency compared to strong model-free baselines in low training regimes is a significant result. The significance is however decreased by the fact that the paper does not answer the question how to obtain good asymptotic performance that matches (or comes close to) model-free state-of-the-art results. I therefore vote for weak accept at this stage.\\n\\nMinor details\\n\\nOn a side note, there are two citations missing related to model-based RL in visual domains:\\n- S. Alaniz. Deep Reinforcement Learning with Model Learning and Monte Carlo Tree Search in Minecraft. In the 3rd Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM), 2017.\\n- F. Leibfried and P. Vrancx. Model-Based Regularization for Deep Reinforcement Learning with Transcoder Networks. In NIPS Deep Reinforcement Learning Workshop, 2018.\\n\\nUpdate\\n\\nAfter the author response, my review remains the same. I think this paper is worthwhile publishing at ICLR.\"}" ] }

[ "Sanjeev Kumar Karn", "Francine Chen", "Yan-Ying Chen", "Ulli Waltinger", "Hinrich Schütze" ]

[ "interleaved posts", "abstractive summaries", "interleaved texts", "postings", "online chat conversations", "several conversations", "topic threads", "place", "difficulties", "readers" ]

{ "note_id": [ "Q0bcibzWTB", "S1gRaXDjoB", "ryxsnA9wiB", "Hyx2id9wjr", "ryxIU45wjB", "BklB71JS5H", "HklrJnbpFr", "Bye6JV5cYr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732499, 1573774278069, 1573527218559, 1573525668156, 1573524557772, 1572298524910, 1571785692650, 1571623908960 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1786/Authors" ], [ "ICLR.cc/2020/Conference/Paper1786/Authors" ], [ "ICLR.cc/2020/Conference/Paper1786/Authors" ], [ "ICLR.cc/2020/Conference/Paper1786/Authors" ], [ "ICLR.cc/2020/Conference/Paper1786/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1786/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1786/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes an end-to-end approach for abstractive summarization of on-line discussions. The approach is contrary to the previous work that first disentangles discussions, and the summarizes them, and aims to tackle transfer of disentanglement errors in the pipeline. The proposed method is a hierarchical encoder - hierarchical decoder architecture. Experimental results on two corpora demonstrate the benefits of the proposed approach. The reviewers are concerned about the synthetic nature of the datasets, limited novelty given the previous work, lack of clear explanation of whether disentanglement is actually needed for summarization, and simpler baselines in comparison to the state-of-the-art. Hence, I recommend rejecting the paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"updated experiments\", \"comment\": \"We would like to let the reviewer know that there was a mix up while running the disentangled text experiments: an incorrectly sampled corpus was used. Therefore, we reran the disentangled experiments with the correct 150k samples of the hard Pubmed corpus and updated the results (see Table 4). The results indicate that a sequential decoder is utilizing the ground-truth cluster information (concatenated disentangled threads), and therefore, is an upper-bound. However, hier2hier is better than seq2seq in utilizing ground-truth cluster information (see row 3 and 4 of Table 4).\"}", "{\"title\": \"performance on synthetic and real world dataset\", \"comment\": \"Thank you for your careful review. Regarding point 1 in the weakness, we would like to inform the reviewer that we have updated the hierarchical attention section, a key contribution, so, it is more clear and differentiable from other existing methodologies. To summarize, as ours is the first attempt in applying hierarchical decoding with hierarchical attention in the domain of text-summarization, we found that the previous hierarchical attention approach by Nallapati et al. 2016, i.e., using sentence (post) level attention (gamma) to scale word level attention (alpha) didn't improve very much, i.e., Rouge-1 of 31.3 and Rouge-2 of 10.2 against Rouge-1 of 30.6 and Rouge-2 10.0, in the hierarchical decoding of interleave text-summaries; see scores in rows 3 and 5 in Table 8. Therefore, in contrast to earlier work, we introduce another level of attention, i.e., high-level attentions (beta). Our newly introduced beta attention is also computed at the word-level with the primary objective being to obtain a thread representation and secondarily to rescale word-attention (alpha). Additionally, beta attention is rescaled by post attention (gamma). Thus, the major difference in our novel approach is attention rescaling, i.e., gamma->beta->alpha, where \\u2018->\\u2019 indicates rescaling, in contrast to previous approaches (Nallapati et al, Cheng et al. 2016, Tan et al. 2017 and Hsu et al. 2018), which use gamma->alpha only. The scores in rows 1 (Rouge-1 33.4) and 2 (Rouge-1 32.7) in Table 8 show it improves performance against previous approach (Rouge-1 31.3), and examples in Table 9 and the Appendix show beta attention brings in phrase-level information and also works in tandem with gamma.\\n\\nRegarding novelty in encoder and decoder, we would like to inform the reviewer that Nallapati et al. 2017 used a hierarchical encoder, they did not use a hierarchical decoder. We are the first to attempt hierarchical decoding in text-summarization while Li et al. 2015 applied it in the context of Autoencoders. Hierarchical decoding not only improves performance, but also computation speed which is equally important in case of longer summaries.\\nWe believe summarization is more challenging since it requires generating a new text distilling the original; it also forms a real test bed for hierarchical decoding. \\n\\nRegarding novelty in hierarchical attention, Thanks for letting us know the lacking, we have updated Table 1 with citations and also updated the hierarchical attention section. We refer the reviewer to page 6. Further, we draw inspiration from the See et al. 2017 pointer mechanism which allows modelling two distributions, i.e., over source tokens and vocabulary tokens together using a switch probability. Basically, the intuition is that in an end-to-end learning if there are two distributions that involves a common set of tokens then they must be modeled together by something, e.g., switch probabilities, but not in isolation. We also model two distributions, but in contrast to See et al., we use them to capture larger-scale and local context, i.e., high-level distribution (beta) and low-level distribution (alpha) over the same source tokens and we model them together by scaling one (alpha) by the other (beta). Further, the popularly held intuition of hierarchical attention, i.e., sentence attention scales word attention, is still upheld as gamma (post-attention) scales the beta. Therefore, all attentions are modeled together without any extra train labels unlike Hsu et al 2018.\\n\\nRegarding point 2 in the weakness, the reason for exclusion is that the disentanglement component of (Shang et al. 2018) requires an explicit number of clusters, which fixes the number of summaries, while the medium and hard corpus has a variable number of clusters, i.e., 2 to 3 (medium) and 2 to 5 (hard). We believe our hard corpus is a challenging task as the model has to detect highly variable threads, i.e., 2 to 5 from 500 source tokens. To justify this, we would also like to note that we mention in the Discussion section that we ran our model on the AMI corpus, a real-world meeting corpus, where the source size was limited to 600 and target to 150 tokens and hier2hier gives us Rouge-1 of 37.57 and Rouge-2 of 11.72, which outperforms the results of Shang et al. 2018 (approx Rouge-1 of 30) for the same target size (150 tokens). Additionally, we fine-tuned one of the Pubmed trained models to the AMI corpus and found a simple fine-tuning after 50k iterations lead to improvement of an additional 2-3 rouge scores. \\nAs these experiments were not part of the original submission, we can update the paper with them upon the acceptance. Regarding the last point in the weakness, we would like to inform reviewer that it is not only the novel hierarchical attention but also hierarchical decoding that is adding gain in the performance.\\n\\nFurther, we have reformatted images and tables, and reduced the draft size to a little over 8 page.\"}", "{\"title\": \"upper bound in interleaved text summarization\", \"comment\": \"Thanks for the valuable feedback. To address some of the reviewer concerns and questions, we ran the experiments to examine performance when disentangled threads are concatenated and used as input, e.g., posts-thread1->posts-thread2->posts-thread3->posts-thread4 => summary1->summary2->summary3->summar4, and for expediency, we used a 150k sampled Pubmed Hard corpus as the original Pubmed Hard corpus is twice the size and a seq2seq model takes about 5 days to converge. The results are presented in Table 4. Evidently, seq2seq doesn't gain much from ground-truth disentangled posts, which is coherent with the result for disentangling using clustering on the easy corpus (see Table 3). We believe this is due to the sequential decoder not being able to utilize the later cluster of posts. Consider a case where all the sentences in a channel are very similar, e.g., all discussing HIV (see example 2 in the Appendix), and when the seq2seq encoder runs over such disentangled posts sequentially, i.e., cluster1-cluster2-cluster3-cluster4 and the decoder generates summaries sequentially, i.e., summary1-summary2-summary3-summary4, then it might be good at generating summary1 and summary2; however, it may not be as good for summary3 and summary4, due to the long-term dependency problem.\\n\\nAdditionally, Table 4 shows that hier2hier still performs best when the posts are disentangled. Also, comparison of hier2hier results on interleaved and disentangled corpora shows that hier2hier is quite successful in disentangling on its own. see examples in the Table 9 and appendix. Overall, disentanglement is helpful for hier2hier models, however it\\u2019s a time-consuming process and can lead to error propagation. \\n\\nRegarding concerns on how tasks look like, we have rearranged Table 2 which illustrates an interleaved text in the corpus, and Table 9, which illustrates model output. We hope this makes the task further understandable . Additionally, we have put more examples in the appendix. Further, we also modified the draft to address reviewer concern on images and tables formatting, they are either enlarged, rearranged or combined to enhance comprehensibility.\"}", "{\"title\": \"elaborated the novelty in hierarchical attention\", \"comment\": \"Thanks for the valuable feedback.\\nRegarding the reviewers first concern on novelty in hierarchical attention, we agree that there are several pieces of literature on hierarchical attention for NLP tasks. Since we are focused on text-summarization, Table 1 (updated) summarizes some of the forms of hierarchical attention in the domain. We would also like to point out that previous ideas on hierarchical attention in the field were devised assuming sequential decoding, and top-down integration of attention, where sentence- (post-) level attention scale word-level attentions, e.g., Nallapati et al. and Hsu et al. As far as we know, ours is the first architecture of applying hierarchical decoding with hierarchical attention in text-summarization, and we found the previous approach on hierarchical attention, i.e., sentences level attentions scaling word level doesn't help much in hierarchical decoding of interleaved text; see scores in row 3 (hier2hier(+gamma-beta)) and row 5 (hier2hier(-gamma-beta)) in Table 7. Therefore, we introduce another mid-level attention, and refer to it as high-level attention (beta), which brings in phrase-level information and works in tandem with both post- and word-level attentions. We refer the reviewer to the updated hierarchical attention section on page 5. \\n\\nRegarding the second concern on weak baseline, we agree that SOTA summarization is superior to the base model, i.e., seq2seq; however, most of these SOTA are also built on seq2seq with advance features like a pointer mechanism and/or coverage and/or RL training. As future work, we aim to include those in our hier2hier setup, as it improved the performance in the news summarization task, we believe the similar improvement will follow in the interleaved case.\\n\\nFurther, we find a very good point is raised in the reviewer first question. At the moment, due to way seq2seq is trained, we are forcing our model to learn to generate tokens by recognizing the occurrence of thread (Table 5) or \\nthe density of threads (Table 7) as this is how token of summaries arranged; however, in the future, we will design or use a training loss that is computed at the summaries level rather than tokens. Further, we would also like to point out that the ROUGE-1 scores; however, doesn\\u2019t penalize the order of summaries, and the number of transitions between summaries is much smaller than the number of transitions between word pairs within a summary.\\nRegarding the second question, the answer is yes. We have compiled the corpus in a manner that the reference summary order follows the first new thread occurrence, e.g., \\npost1-thread1, post1-thread2, post-2-thread2, post1-thread3 => thread1-summary, thread2-summary, thread3-summary. Additionally, we evaluated models strength in recognizing threads where summaries are ordered by the location of each thread\\u2019s greatest density. Here, density refers to smallest window of posts with over 50% of posts belonging to a thread; e.g., post1-thread1, post1-thread2, post-2-thread2, post2-thread1, post3-thread1, post4-thread1 => thread2-summary, thread1-summary. In this example, although thread1 occurs early, as the majority of posts on thread1 occurs latter, therefore, its summary also occurs later. As shown in Table 7, the model recognizes threads even in these scenarios, which is indicated by its performance (see Table 7) and is similar to the above scenario (Table 5). \\nRegarding the concern on the implementation, we would like to inform reviewer that we indeed used one provided by TensorFlow, however, we modified hyperparameters and learning algorithms and therefore wanted to convey that the gains are not due to those changes. We will reference the TensorFlow model that we used as a base. Further, we have also modified the draft to fix typos pointed out by the reviewer and also changed images and tables formatting to enhance comprehensibility, now the draft size is a little over 8 page.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This work proposes an architecture to generate summaries for multi-participant postings as in chat conversations. Previous work tackled this problem as a pipeline of disentangling threads from mixed chat posts and then generating a summary for each post. The presented work proposes a hierarchical attention model to solve this task in an end-to-end fashion.\", \"the_proposed_architecture_is_divided_into\": [\"a hierarchical encoder: encodes each sentence to a vector using bi-lstm and average pooling which is then fed to a post-to-post encoder which generates a set of representations corresponding to those posts.\", \"a hierarchical decoder: number of disentangled posts are generated using a thread-to-thread decoder which given the max pooled representations from the post-to-post encoder generated a number of initial representations for each thread summary using hierarchical attention combining attention over paragraphs and over words at each decoding time step.\"], \"experiments\": \"Authors demonstrate the effectiveness of their approach by comparing against a set of baselines traditional seq2seq, as well as cluster-> seq2seq and the model from Shang et al. (2018). Evaluation datasets consisted of several datasets for chat summarization synthesized to mimic interleaved conversation with different difficulties. \\n\\nThrough an ablation study, authors also show the effectiveness of the hierarchical encoding and hierarchical decoding as well as the post level and high-level attention.\", \"overall\": \"Although this paper is well written and well-motivated. The work is a bit incremental and builds upon previous ideas from hierarchical attention literature to apply for interleaved text generation. The provided baselines are quite weak compared to the SOTA summarization methods at the moment, although none of them is directly modelled for interleaved text summarization through multi summaries.\", \"typos\": [\"Nallapati et al. Nallapati et al. (2016)\", \"See et al. See et al. (2017)\"], \"question\": \"Is there an intuition behind penalizing a model that generates correct summaries for each thread with different order? Do references summaries follow the same order as in the threads in the conversation?\\n\\nIn the paper you dedicated several paragraphs to reimplement basic models such as seq2seq or (See et al.) pointer generator network and make sure they work correctly. There are existing implementations and pre-trained models available such as in OpenNMT.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": [\"This paper proposes a technique for generating summaries of interleaved texts. Unlike previous work that first perform unsupervised clustering to extract ordered threads, the authors instead propose a hierarchical model that directly process interleaved threads. In particular, the authors propose using a hierarchical encoder that encodes words -> post -> threads as well as a hierarchical decoder that decodes threads -> words. On synthetic interleaved datasets composed from PubMed and StackExchange, the proposed method outperform seq2seq with attention as well as a pipeline system that applies unsupervised clustering followed by seq2seq. The authors present results for well-chosen ablations as well as baselines.\", \"My concerns stem from the synthetic nature of these tasks. In particular:\", \"is it difficult to disentangle the threads? Suppose we train a simple, supervised model instead of using unsupervised clustering and run the same pipeline experiment, how well would the model perform compared to hier2hier?\", \"due to the nature of these two tasks, is disentanglement even helpful for summarization? If not, heir2hier might be picking up on helpful signals that are not accessible to baselines. I don't have a good idea of what the tasks look like from reading the paper. I would ask that the authors put some examples of input output pairs in the paper. The algorithm in the appendix for interleaving is not very helpful.\", \"I would like to see some upperbound of running seq2seq on ground-truth disentangled threads.\"], \"some_comments_on_the_writing\": [\"Figure 4 and Table 2 are way too small\"]}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1786\", \"review\": [\"<Strengths>\", \"This paper addresses the problem of summarizing interleaved posts, which involves two subtasks of identifying the clusters of threads and summarizing each thread into a sentence.\", \"This is an understudied area of research in summarization research, only few previous works have been done; for example, (Shang et al. 2018). The proposed approach is end-to-end trainable, whereas (Shang et al. 2018)\\u2019s method tackles the problem in two separate stages.\", \"<Weakness>\", \"1. The technical novelty may need to be further justified.\", \"The proposed model (in Fig.2) consists of the encoder, the decode and hierarchical attention module.\", \"The encoder and the decoders are standard ones based on (Nallapati et al. 2017) and (Li et al. 2015), respectively. It is hard to find any technical novelty here.\", \"This paper argues that the proposed hierarchical attention is novel with no clear ground. It simply mentions the inspiration from (See et al. 2017), but does not explicitly what aspects are novel as Hierarchical attention has been popularly used in general seq2seq models for many NLP applications.\", \"Table 1 is rather cursory and incomplete as it compares with only three different models.\", \"2. Experiments have much room for improvement.\", \"The two datasets are simply mixture of text from existing datasets and rather artificial and far from actual interleaved posts, as shown in Table 2.\", \"In page 6, three different corpora are introduced with varying interleave parameters. However, the parameters a,b,m, and n are rather small (<= 5), so it is hard to say the configurations are challenging.\", \"The SOTA model (Shang et al. 2018) is only compared on the PubMed Easy Corpus in Table 3. Is there any special reason to exclude it from the comparison with more challenging corpora?\", \"In most results of experiments, the paper focus on showing that the proposed hierarchical attention model is better than basic seq2seq, which may not be strong results to show the effectiveness of the proposed model.\", \"<Minor comments>\", \"This draft does not seem to follow the ICLR format.\", \"Fig.2-4 are hard to see in details. They should be enlarged and re-organized.\", \"References should be added to Table 1.\", \"This paper exceeds the limit of 8 pages. I cannot find any good strong ground to have more additional pages than the commended length.\", \"<Conclusion>\", \"My rate is borderline with slightly leaning toward reject, mainly due to issues of technical novelty and evaluation. I will decide my rate finally, based on the authors\\u2019 response to the concerns above.\"]}" ] }

[ "Nimrod Segol", "Yaron Lipman" ]

[ "deep learning", "universality", "set functions", "equivariance" ]

{ "note_id": [ "fXxg3D9dK", "S1ljmyyjsH", "SJgNL-s_ir", "H1xgIgLWjr", "BJxuByU-oS", "S1gah0BboB", "rke3Frfx5S", "SyxwgLc0tB", "HylapLxoKH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732470, 1573740323387, 1573593420422, 1573113928196, 1573113663807, 1573113524854, 1571984772512, 1571886574650, 1571649220523 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1785/Authors" ], [ "ICLR.cc/2020/Conference/Paper1785/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1785/Authors" ], [ "ICLR.cc/2020/Conference/Paper1785/Authors" ], [ "ICLR.cc/2020/Conference/Paper1785/Authors" ], [ "ICLR.cc/2020/Conference/Paper1785/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1785/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1785/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper shows that DeepSets and PointNet, which are known to be universal for approximating functions, are also universal for approximating equivariant set functions. Reviewer are in agreement that this paper is interesting and makes important contributions. However, they feel the paper could be written to be more accessible.\\n\\nBased on the reviews and discussions following author response, I recommend accepting this paper. I appreciate the authors for an interesting paper and look forward to seeing it at the conference.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Revision uploaded:\", \"comment\": \"We thank the reviewers for their comments. We have uploaded a revision of our paper. The following are the changes made:\", \"reviewer1\": \"The metric in the Knapsack experiment was changed to percentage of sets for which the predicted solution satisfies all of the constraints and is within 10% of the optimal solution.\\nThe definition of the DeepSets and PointNetSeg models was edited. We also made slight changes to the proof of Corollary 1 in accordance to the notation in the new definition.\", \"reviewer2\": \"A precise reference to the result in Golubitsky&Stuart [2002] (equivatiant polynomials $\\\\mathbb{R}^n \\\\to \\\\mathbb{R}^n$) was added.\", \"reviewer3\": \"We added a conclusion section that summarizes our results and contributions.\"}", "{\"title\": \"Asking followup questions\", \"comment\": \"Thanks for the clarification.\\n\\n>> \\\"As such it can be approximated with the universal models considered in this paper. (e.g. PointNetST, DeepSets etc.)\\\"\\n\\nI would like to see some experimental results. How many layers of the DeepSet layers would be required to approximate one GraphOp. I am very interested in its capacity to approximate every continuous function and have some realistic estimate of the capacity of these networks.\"}", "{\"title\": \"Addressing Reviewer 2\", \"comment\": \"We thank the reviewer for the detailed and thoughtful review. We address the reviewers main concerns.\\n\\n\\u201cA more general function is $P(X)_i=Ax_i+\\\\sum_{j\\\\in N(x_i,X)} B(x_j,x_i) x_j + c$, where $N(x_n,X)$ is the set of index of neighbors within the set\\u2026 Then, can the function family the authors used in the paper approximate this function? No.\\u201c\\n>> We respectfully disagree. The function P(X) described by the reviewer can be approximated arbitrarily well using a continuous equivariant function (by using bump functions to approximate indicators functions of neighbors). As such it can be approximated with the universal models considered in this paper. (e.g. PointNetST, DeepSets etc.)\", \"reiterating_the_main_result_of_this_paper\": \"*Every* continuous equivariant function defined solely on a set of feature vectors can be approximated with PointnetST over a compact domain.\\nWe are happy to include a discussion about this in the revised paper. \\n\\n\\n\\u201cOn the Knapsack test, the metric of interest is not the accuracy of individual prediction. Rather, whether the network has successfully predicted the optimal solution, or how close the prediction is to the solution.\\u201d\\n>> Thank you for this comment. We will revise the Knapsack graph to show the success metric suggested by the reviewer. We remark that the goal here was not to construct a network that solves the Knapsack problem but to demonstrate the difference between universal and non-universal models.\\n\\n\\u201cSpare few lines of equations to define what are DeepSets and PointNetSeg in the paper and point out the difference\\u201d\\n>> The definition of DeepSets appears in Equations 7 and 8. The PointNetSeg model is described in detail in the discussion before the proof of Corollary 1. We will make the definitions clearer.\\n\\n\\n\\u201cLemma 3 is too trivial.\\u201d\\n>> We agree it is trivial (and indeed the proof is a one-liner). If the reviewers feel strongly about this, we can move it to appendix, however we feel it helps to provide a complete picture. \\n\\n\\u201cP.2 power sum multi-symmetric polynomials. \\\"For a vector and a multi-index ...\\\" I think it was moved out of the next paragraph since the same is defined again as again in the next sentence.\\u201d\\n>> Thank you. We will make this clear.\"}", "{\"title\": \"Replying to Reviewer 1\", \"comment\": \"We thank the reviewer for the review and the comments. Below we address the main concerns.\\n\\n\\u201cone may study the simple case of having single input channel, for which the output at index \\\"i\\\" of an equivariant polynomial is written as the sum of all powers of input multiplied by a polynomial function of the corresponding power-sum. This second part is indeed what is used in the proof of the universality of the permutation invariant version of DeepSets, making the connection more visible. Generalizing this to the multi-channel input as the next step could make the proof more accessible \\u201d\\n>> We will highlight the connection to the case of single input channel and DeepSets permutation invariance universality. \\n\\n\\u201cThe second issue I would like to raise is related to discussions around the non-universality of the vanilla PointNet model. ... if there were any other objectives beyond this in the experiments could you please clarify? \\u201c\\n\\n>> We agree it is trivial (and indeed the proof is a one-liner). If the reviewers feel strongly, we can move it to appendix, however we feel it helps to provide a complete picture. We included it in the experiments as a naive baseline and to show that adding a single transmission layer indeed provides a significant improvement.\\n\\n\\u201cFinally, could you give a more accurate citation (chapter-page number) for the single-channel version of Theorem 2.?\\u201c \\n>> This result is Proposition 2.27 in Golubitsky&Stuart(2002). We will update the paper to give a more accurate citation.\"}", "{\"title\": \"Addressing Reviewer 3 concerns\", \"comment\": \"We thank the reviewer for the detailed review. Below we address the main concerns.\\n\\n\\u201cI think the paper needs some changes in written style mainly, to make it friendlier to newcomers to the area... For instance, the omission of a results discussion section or a conclusion are clearly not reader friendly\\u201d\\n>> We discussed the results of the experiments in pages 7-8. We will revise to make this discussion easier to find.\\n\\n\\n\\n\\u201cWhat would have been nicer would be to perform some sort of ablation study, where the authors studied how the representational capacity of the network changed as a result of them introducing the universal linear transmission layer.\\u201d\\n>> We refer the reviewer to Figure 2 where one can see that the PointNet model underperforms across various tasks compared to PointNetST that is identical to PointNet except the addition of a single linear transmission layer. Furthermore, PointNetQT, PointNetSeg, and DeepSets can be seen as different versions of PointNet variations.\\n\\n\\n\\u201cA direct theoretical and experimental comparison between PointNet and PointNetST would have been useful for me to understand the impact of the change that the authors introduce..\\u201d\\n>> In Theorem 1 we state that PointNet is not equivariant universal, but PointNet with a single transmission (PointNetST) layer is. In the experiments section we compare these two models on three different learning tasks. \\n\\n\\n\\u201cWhere is the conclusion section?\\u201d\\n>>We felt it is unnecessary but see we were wrong, we will add one.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"TLDR: The function these deep set networks can approximate is too limited to call these networks universal equivariant set networks. Authors should scope the paper to the specific function family these networks can approximate. No baseline comparison with GraphNets.\\n\\n\\n\\nThe paper proposes theoretical analysis on a set of networks that process features independently through MLPs + global aggregation operations. However, the function of interest is limited to a small family of affine equivariant transformations.\\n\\nA more general function is\\n\\n\\\\begin{equation}\\nP(X)_i = Ax_i + \\\\sum_{j \\\\in N(x_i, X)} B_{(x_j, x_i)} x_j + c\\n\\\\end{equation}\\n\\nwhere $N(x_i, X)$ is the set of index of neighbors within the set $X$. It is trivial to show that this function is permutation equivariant.\\n\\nThen, can the function family the authors used in the paper approximate this function? No.\\nCan the proposed permutation equivariant function represent all function the authors used in the paper? Yes.\\n\\n1) If $B=0$, then the proposed function becomes MLP.\\n2) If $A=0, N(x_i, X) = [n]$ and $B_{(x_j, x_i)} \\\\leftarrow B$, then this is $\\\\mathbf{1}\\\\mathbf{1}^TXB$, the global aggregation function.\\n\\nAlso, this is the actual function that a lot of people are interested in. Let me go over few more examples.\\n\\n3) If $N(x_i, X) = $adjacency on a graph and $B_{(x_j, x_i)} \\\\leftarrow B$, then this is a graph neural network \\\"convolution\\\" (it is not a convolution)\\nExample adjacency $N(x_i, X) = \\\\{j \\\\;| \\\\; \\\\|x_i - x_j\\\\|_p < \\\\delta, x_j \\\\in X\\\\}$.\\n\\\\begin{equation}\\n\\\\text{GraphOp}(X)_i = Ax_i + \\\\sum_{j \\\\in \\\\{j \\\\;| \\\\; \\\\|x_i - x_j\\\\|_p < \\\\delta, x_j \\\\in X\\\\}} Bx_j + c\\n\\\\end{equation}\\n\\n4) If $x_i = [r,g,b,u,v]$ where $[r,g,b]$ is the color, $[u,v]$ is the pixel coordinate and $N(x_i, X) =$ pixel neighbors within some kernel size, $B(x_j, x_i)$ to be the block diagonal matrix only for the first three dimensions and 0 for the rest, then this is the 2D convolution.\\n\\n\\nAgain, the above function is a more general permutation equivariant function that can represent: a graph neural network layer, a convolution, MLP, global pooling and is one of the most widely used functions in the ML community, not MLP + global aggregation.\", \"regarding_the_experiment_metrics_and_plots\": \"On the Knapsack test, the metric of interest is not the accuracy of individual prediction. Rather, whether the network has successfully predicted the optimal solution, or how close the prediction is to the solution.\", \"for_example\": \"success rate within the epsilon radius of the optimal solution while satisfying all the constraints. Fail otherwise. If these networks can truly solve these problems, authors should report the success rate while varying the threshold, not individual accuracy of the items which can be arbitrarily high by violating constraints.\\n\\nAlso, the authors should compare with few more graphnet + transmission layer (GraphNetST) baselines with the graph layers: $P(X)_i = Ax_i + \\\\sum_{j \\\\in N(x_i, X)} Bx_j + c$ and the same single transmission layer $\\\\mathbf{1}\\\\mathbf{1}^TXB$ in PointNetST.\\nPointNet is a specialization of graphnets and GraphNetST should be added as a baseline with reasonable adjacency.\\n\\nAlso experiment figures are extremely compact. Try using log scale or other lines to make the gaps wider.\\n\\n\\n\\n\\n\\n\\nMinor\\n\\nI am quite confused with the name PointNetST. Authors claim adding one layer of DeepSet layer to a PointNet becomes PointNetST, but I see this as a special DeepSet with a single transmission layer. The convention is B -> B', not A + B -> A'. In this case, A: PointNet, B: DeepSet\\n\\nLemma 3 is too trivial.\\n\\nThe paper is not very self contained. Spare few lines of equations to define what are DeepSets and PointNetSeg in the paper and point out the difference since these networks are used throughout the paper extensively without proper mathematical definition.\\n\\nP.2 power sum multi-symmetric polynomials. \\\"For a vector $x \\\\in R^K$ and a multi-index ...\\\" I think it was moved out of the next paragraph since the same $x$ is defined again as $x \\\\in R^n$ again in the next sentence.\\nAlso, try using the consistent dimension for x throughout the paper, it confuses the reader.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper presents proof that the DeepSets and a variant of PointNet are universal approximators for permutation equivariant functions. The proof uses an expression for equivariant polynomials and the universality of MLP. It then shows that the proposed expression in terms of power-sum polynomials can be constructed in PointNet using a minimal modification to the architecture, or using DeepSets, therefore proving the universality of such deep models.\\n\\nThe results of this paper are important. In terms of presentation, the notation and statement of theorems are precise, however, the presentation is rather dry, and I think the paper can be significantly more accessible. For example, here is an alternative and clearer route presenting the same result: one may study the simple case of having single input channel, for which the output at index \\\"i\\\" of an equivariant polynomial is written as the sum of all powers of input multiplied by a polynomial function of the corresponding power-sum. This second part is indeed what is used in the proof of the universality of the permutation invariant version of DeepSets, making the connection more visible. Generalizing this to the multi-channel input as the next step could make the proof more accessible. \\n\\nThe second issue I would like to raise is related to discussions around the non-universality of the vanilla PointNet model. Given the fact that it applies the same MLP independently to individual set members, it is obvious that it is not universal equivariant (for example, consider a function that performs a fixed permutation to its input), and I fail to see why the paper goes into the trouble of having theorems and experiments just to demonstrate this point. If there were any other objectives beyond this in the experiments could you please clarify? \\n\\nFinally, could you give a more accurate citation (chapter-page number) for the single-channel version of Theorem 2.?\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I made a quick assessment of this paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"*CAVEAT*\\nI must caveat that this paper is out of my comfort zone in terms of topic, so my review below should only be taken lightly. It also explaina the brevity of my review. My apologies to the authors and other reviewers.\\n\\n*Paper summary*\\n\\nThe authors design a set architecture, which is equivariant to permutations on the input. They show the simplest such set architecture, which preserves equivariance, while being a universal approximator. Nicely this architecture relies on a correction to PointNet, called PointNetST, which they show is not equivariant universal. Furthermore, they run experiments on a few toy examples demonstrating that their system performs well.\\n\\n*Paper decision*\\n\\nI have decided to give this paper a weak accept, since it contains both theory and nice experiments. To change to a firm accept, I think the paper needs some changes in written style mainly, to make it friendlier to newcomers to the area, which can easily be implemented in the camera ready stage of paper preparation. For instance, the omission of a results discussion section or a conclusion are clearly not reader friendly.\\n\\n*Supporting arguments*\\n\\n- The paper is written clearly. This said, it requires a great deal of effort to follow the maths if you are not already fluent in a lot of the ideas used in the paper (this includes myself). \\n- I think the structure of the paper is fine for this sort of work. Perhaps at the beginning it would be more useful to spend more time on a roadmap of the results presented in the paper and to explain the exact significance of why the reader should want to continue reading.\\n- I think the selection of experiments is nice, containing both regression and classification. What would have been nicer would be to perform some sort of ablation study, where the authors studied how the representational capacity of the network changed as a result of them introducing the universal linear transmission layer.\\n- A direct theoretical and experimental comparison between PointNet and PointNetST would have been useful for me to understand the impact of the change that the authors introduce.\\n\\n*Questions/notes for the authors*\\n\\n- Please answer my concerns in the support arguments\\n- Where is the conclusion section?\"}" ] }

[ "Hadi Pouransari", "Oncel Tuzel" ]

[ "Binary Neural Networks", "Quantization" ]

{ "note_id": [ "317S0eKGo_", "rXDkzMNLua", "SkgXuMtDnB", "SkxFoCbjjH", "HkxLCbMcsS", "rJeVGrbcir", "Syg3lE-qir", "rylLGm-9iB", "SygD-H4pKH", "SklAmT2cKS", "BkxkiR7o_r" ], "note_type": [ "official_comment", "decision", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1579129946096, 1576798732440, 1574568555505, 1573752481159, 1573687757533, 1573684492423, 1573684212149, 1573683981935, 1571796223253, 1571634469652, 1570614935220 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Paper1783/Authors" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1783/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1783/Authors" ], [ "ICLR.cc/2020/Conference/Paper1783/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1783/Authors" ], [ "ICLR.cc/2020/Conference/Paper1783/Authors" ], [ "ICLR.cc/2020/Conference/Paper1783/Authors" ], [ "ICLR.cc/2020/Conference/Paper1783/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1783/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1783/AnonReviewer3" ] ], "structured_content_str": [ "{\"title\": \"Response to Review 4\", \"comment\": \"We would like to thank the reviewer for their time and valuable comments. However, we would like to note that, the comments would be more helpful if were posted before the end of discussion period to find a constructive solution.\", \"comment_1\": \"(\\\"the authors should explicitly state in the paper that the optimality is in terms of what...\\\")\\nThe optimality refers to the least squares solution of quantization of isolated tensors (both weights and activations). It is mentioned at the beginning of page 4 (Section 3) in the paper: \\u201cAll optimal solutions are with respect to this error and hold for an arbitrary distribution p(x). \\u201d\\n\\nSimultaneous quantization of all layers , or loss aware quantization techniques may result in better model accuracy compared to independent quantization. However, In this paper we study the effect of quantization error in isolation and propose provably optimal (in L2 sense) 2-bits Xnor quantization.\", \"comment_2\": \"(\\\"... it is not appropriate to formulate quantization error the weights and activations in the format of continuous distribution...\\\")\\nWe respectfully disagree with the reviewer. The number of weights and activations for networks of interests (e.g., resnet18) are statistically large, make the continuous approximation accurate enough. We have verified this by sub-sampling tensors (i.e., use a fraction of values in a tensor instead of all its elements) and observe negligible change in the resulted statistics. Moreover, as mentioned in the paper continuous notation is used only for ease of derivation and explanation of the methodologies.\", \"comment_3\": \"(\\\"It is also not clear to me why this paper begins with rank-1 quantization...\\\")\\nPlease see our answer to Reviewer 2 regarding this question.\", \"comment_4\": \"(\\\"One of my main concerns is the novelty of this paper...\\\")\\nThe main novelty of the paper is on rank-1 analysis and optimal 2-bits quantization solution. The optimal 1-bit quantization is provided for completeness, and optimal ternary is mentioned as a specific case of 2-bits quantization. We thank the reviewer to point to [2], which unfortunately we missed in our previous works review. As the reviewer mentioned, in [2] the optimal ternary quantization is obtained when Hessian is replaced by scalar times identity matrix; similarly, we obtain optimal ternary quantization when add a constraint on scaling factors in our general 2-bits solution: v1 and v2 to be equal.\", \"comment_5\": \"(\\\"Yet another concern is about the experiments...\\\")\\nThe performance boost in our implementation of BWN comes from using full precision short-cuts. It is mentioned in Appendix E in training details.\\n\\nPACT uses uniform quantization with 2-bits for weights and activations. In this paper we focus on networks with binary weights, 1-bit per value (therefore maximum size compression), and improve model accuracy by reducing the quantization error of activations. Hence, results are not directly comparable. Please note that the adaptive clipping mechanism introduced in PACT is orthogonal to our optimal quantization techniques and could be used together.\\n\\nWe propose provably optimal 2-bits quantization and support optimality of the method empirically as well. The final accuracy depends on other factors as well (e.g., learning rate schedule and batchsize as shown in Table 2), that we did not tune for. LQ-Net provides a suboptimal solution to the optimization problem that we solve in this paper. Few other differences: LQ-Net uses quantization with {0,1} encoding for activations and {-1,1} for weights (while we use {-1,1} for both), and uses a different order of operations in a layer for network than Xnor-Net (that we adopted in this paper). Our final accuracy results for ImageNet-ResNet18 are close (62.4% vs 62.6%).\"}", "{\"decision\": \"Reject\", \"comment\": \"This paper proposes to quantize the weights of neural networks that can minimize the L_2 loss between the quantized values and the full-precision ones. The paper has limited novelty, as many of the solutions presented in the paper have already been discovered in the literature. During the discussion, the reviewers agree that it is an incremental contribution. Parts of the paper can also be clarified, particularly on the optimality of the solution, assumptions used in the approximation, and some of the experimental results. Experimental results can also be made more convincing by adding comparision with the more recent quantization methods.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper proposes to quantize the weights of neural networks that can minimize the L_2 loss between the quantized values and the full-precision ones. The authors propose solutions for optimal 1-bit/ternary/2-bit quantization, as well as a greedy algorithm to approximate the optimal k-bit quantization. Experiments are performed on image classification data set ImageNet using ResNet-18.\\n\\nFirst of all, the authors should explicitly state in the paper that the optimality is in terms of what? Indeed the authors obtain the solution of (1) quantization with a scaling parameter, and (2) in terms of minimizing quantization error (L_2 loss between the quantized value and the full-precision one), which is quite restricted to be a universally optimal one.\\n\\nSince the number of weights and activations are limited in the network, it is not appropriate to formulate quantization error the weights and activations in the format of continuous distribution in (4) and (8). \\n\\nIt is also not clear to me why this paper begins with rank-1 quantization but ends up with scaled quantization. What kind of assumption are used in this approximation, and can the optimality still be guaranteed?\\n\\nOne of my main concerns is the novelty of this paper. Many of the solutions in the paper have already been discovered in literature. For instance, the optimal 1-bit solution in (5) was already obtained in Binary-Weight-Network [1] in 2016. The optimal ternary solution (i.e., the ternarization threshold should be 1/2 of the scaling parameter) in (12) was also already obtained in Corollary 3.1 in [2] in 2018, as a special case when curvature information is dropped.\\n\\nYet another concern is about the experiments. Since the proposed optimal binarization has the same solution as BWN, where does the performance gain in Table 2 come from? Moreover, some of the recent quantization methods are not compared. For instance, in PACT [3], 2-bit weight&acitvation quantization already achieves 64.4% top-1 accuracy of Resnet18 on ImageNet, while the proposed method achieves the same accuracy with full-precision activation and 2-bit weight (Table 2). In addition, according to Table 2, the proposed method also can not beat LQ-Net. \\n\\n\\n[1] Rastegari, Mohammad, et al. \\\"Xnor-net: Imagenet classification using binary convolutional neural networks.\\\" ECCV, 2016.\\n[2] Hou, Lu, and James T. Kwok. \\\"Loss-aware weight quantization of deep networks.\\\" ICLR 2018.\\n[3] Choi, Jungwook, et al. \\\"Pact: Parameterized clipping activation for quantized neural networks.\\\" arXiv preprint arXiv:1805.06085 (2018).\"}", "{\"title\": \"On the connection between rank-1 binary quantization and scaled binary quantization\", \"comment\": \"Thank you for your prompt feedback.\", \"we_make_this_conclusion_by_the_following_steps\": \"1) For any matrix $X$ we prove in Appendix A that the best (in Frobenius norm sense) rank-1 binary quantization is the truncated SVD of its element-wise absolute value matrix, shown by truncated-SVD-1($|X|$), times (Hadamard) its element-wise sign.\\n\\n2) We then consider a case where entries of $X \\\\in \\\\mathbb{R}^{m \\\\times n}$ are i.i.d. $\\\\sim \\\\mathcal{N}(0,1)$, a case that is close to the actual distribution of weights and activations (after batch normalization) in neural networks.\\n\\n3) With the assumption (2), we empirically show (in Figure 4 in the revised manuscript) that the truncated-SVD-1($|X|$) $\\\\approx \\\\mu {\\\\bf 1} {\\\\bf 1}^T$ where $\\\\mu = \\\\sqrt{2/\\\\pi}$. We did not present a rigorous proof since it involves asymptotic analysis of largest singular values of a random matrix with non-zero mean which we believed to be not very relevant to the contributions of the paper and would unnecessarily complicate it. The sketch of this proof is as follows: In [1] and [2] the authors show that the largest singular value of a random matrix with i.i.d. entries from a distribution with mean $\\\\mu$ and bounded 4th moment (which is the case for standard folded normal distribution) asymptotes to $\\\\sqrt{mn}\\\\mu$ as $m$ and $n$ are increased (with $m/n \\\\to$constant). Note that $\\\\mathbb{E}[\\\\frac{{\\\\bf 1}^T}{\\\\sqrt{m}} |X|~\\\\! |X|^T \\\\frac{{\\\\bf 1}}{\\\\sqrt{m}} ] = \\\\mathbb{E}[\\\\frac{{\\\\bf 1}^T}{\\\\sqrt{n}} |X|^T ~\\\\! |X| \\\\frac{{\\\\bf 1}}{\\\\sqrt{n}} ] = mn\\\\mu^2$ (with convergence given by the central limit theorem), and therefore, for large matrices the first left and right singular vectors are expected to be almost constant. If the reviewer believes that this proof adds to the paper, we will be happy to include it.\\n\\nFrom (3), we conclude that: (a) The optimal rank-1 binary quantization captures $2/\\\\pi \\\\simeq 0.64$ of the total energy of $X$, making it a good approximation, (b) optimal rank-1 binary quantization can be written as a scalar times a binary matrix.\\n\\n\\n[1] Silverstein, Jack W. \\\"The spectral radii and norms of large dimensional non-central random matrices.\\\" Stochastic Models 10.3 (1994): 525-532.\\n[2] Bryc, Wlodek, and Jack W. Silverstein. \\\"Singular values of large non-central random matrices.\\\" arXiv preprint arXiv:1802.02960 (2018).\"}", "{\"title\": \"From rank-1 quantization to scaled quantization\", \"comment\": \"Thank you for the revision, which has improved the presentation.\\n\\nI still don't understand why rank-1 quantization can be approximated by scaled quantization. In your approximation, why u_1, v_1, the singular vectors can be approximated as a vector of all 1's?\"}", "{\"title\": \"Response to Review 3\", \"comment\": \"We would like to thank the reviewer for their time and helpful feedback. We believe that the concern of the reviewer is due to a misunderstanding (convexity of the optimization problem) which we clarify below. We would appreciate if the reviewer would reevaluate the contributions based on the provided clarifications.\", \"comment_1\": \"(\\u201cFinding the best quantization ends up in solving a program, which is convex and quite straightforward for k=1 and 2, and unfortunately (apparently) non-convex for k>2 ...\\u201d)\\nThe optimization problem discussed in this paper is not convex for any $k$. In (8), the optimization is over both $s_i$\\u2019s (functions with range in {-1,1}) and $v_i$\\u2019s (non-negative scalars). This optimization domain is a non-convex set. For this non-convex problem, we derive the optimal solution when $k$ is equal to 1, 2 and for ternary quantization. The solution to this problem requires finding a solution that satisfies equations (10) and (11). An efficient algorithm is explained in Section 3.2.2 to solve equations (10) and (11). We suppose the convexity discussion after equation (8) might have led to the misunderstanding that the problem is convex for $k=1$ and 2, therefore, in the revised draft we modified this section to be more clear.\", \"comment_2\": \"(\\u201cThe motivation of this work is DNN, arguing that quantized vectors should improve computations cost, hence some experiments are provide on DNN ...\\u201d)\\nThis work is a follow-up on a sequence of papers on Binary Neural Networks (BNNs). In this work, we specifically focussed on improving the quantization accuracy to improve the model performance, while maintaining a similar computational saving as previous works (for example in [1] the authors show 23x speed-up in matrix-matrix multiplication when binary quantization is used, in [2] the authors report 58x speed up when using BNNs). Therefore, we did not discuss latency and power saving when XNOR based computations are used. We clarified this in the revised draft.\", \"comment_3\": \"(\\u201cThe main criticisms is therefore the lack of concrete evidences that those schemes are actually helpful and, on the other hand, the relative simplicity ...\\u201d)\\nWe respectfully disagree this criticism.\", \"experiments\": \"Our goal in this paper is to improve the quantization accuracy, while keeping the same computational budget as original BNN/XNor papers. In Section 5, we present a comprehensive set of experiments: (1) quantization error of activations, (2) post-training quantization, and (3) during-training quantization. All experiments show that the proposed optimal quantization algorithms significantly improve the accuracy.\", \"theory\": \"Introduced rank-1 binary quantization analysis is a unified framework to analyze different scaling strategies that were proposed in the literature (e.g. scaling used in XNOR-net [2]). Using this analysis, we show that the scaled binary quantization is a good approximation to a full precision tensor. Moreover, as discussed above, the optimization problem (8) is non-convex. For this non-convex problem, we derive the provably optimal solution when $k$ is equal to 1, 2 and for ternary quantization for an arbitrary data distribution.\\n\\n[1] Hubara, Itay, et al. \\\"Quantized neural networks: Training neural networks with low precision weights and activations.\\\" The Journal of Machine Learning Research 18.1 (2017): 6869-6898.\\n[2] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In European Conference on Computer Vision, pp. 525\\u2013542. Springer, 2016.\"}", "{\"title\": \"Response to Review 2\", \"comment\": \"We would like to thank the reviewer for their time and constructive feedback. We are glad that the reviewer appreciated the significance of the presented theory. The following are our responses to reviewer\\u2019s comments.\", \"comment_1\": \"(\\u201cThe writing is of good quality. The length is a bit larger than the recommended length of 8 page\\u201d)\\nAs suggested by the reviewer we compressed the manuscript to 8 pages.\", \"comment_2\": \"(Relation of rank-k quantization in Section 2 to the proposed method)\\nAs reviewer mentioned computing the optimal low-rank binary quantization through SVD is computationally expensive. Our goal to introduce low rank binary quantization is providing a unified framework to analyze different scaling strategies used in the literature. For example, we show that the scaling used in XNOR-net is a special rank-1 binary quantization. Moreover, using this framework we show that \\u201cscaled binary quantization\\u201d is a good approximation to a full precision tensor with i.i.d. Normal entries (which is approximately the case in layer activations or weights of a full precision neural network). In our derivation we first approximate a full precision tensor by the optimal rank-1 binary quantization. We then show that the optimal rank-1 binary quantization captures most of the energy (~0.64) of the full precision matrix, and can be approximated as a scalar times a binary matrix.\\n\\nThanks for the suggestion to simplify Section 2. In the revised manuscript, we removed the rank-k quantization and only discuss the rank-1 binary quantization to make the above points clearer. This also helped us to compress the paper to 8 pages as suggested by the reviewer.\", \"comment_3\": \"(\\u201cSomewhere in the text, the authors have to explain the optimality is with respect to the L_2 loss\\u201d)\\n In the revised draft, we clarified the choice of L2 loss and the optimal solution depends on the loss (in the first paragraph of Section 3).\", \"comment_4\": \"(\\u201cFigure 4, could you explain why the angle trends up as the layer index increases\\u201d)\\nIn Figure 4 (Figure 3 in the revised manuscript), we investigate the accuracy of different quantization schemes discussed in the manuscript for activation maps of a trained (non-quantized) network. The network is a trained image classifier, therefore, as the layer index increases the feature maps become more linearly separable. This is done through the composition of several layers of non-linearities (ReLU in this case). Application of ReLU in each layer results in concentration of many inactive feature maps at zero. Hence, after batch-normalization we observe more skewed distribution in later layers, in contrast to early layers that are more similar to normal distribution. The family of binary quantization schemes discussed in this paper (that enables efficient implementation using bit-wise XNOR operations) are symmetric, hence, are less accurate for more skewed distributions. Figure 4 (Figure 3 in the revised manuscript) shows that when the introduced optimal 2-bits quantization is used, the sensitivity of error (angle) to layer index is significantly less. We clarified this in the revised draft.\", \"comment_5\": \"(\\u201cAfter equation(4): mention the choice of p(x)\\u201d)\\nThe optimal quantizations introduced in Section 3 (1-bit, 2-bits, and ternary) hold for any distribution. We do not make any specific choice for distribution p(x). We clarified this in the revised draft in the first paragraph of Section 3.\", \"comments_6\": \"(Typos)\\nThanks for pointing out to the typos. We corrected them in the revised manuscript.\"}", "{\"title\": \"Response to Review 1\", \"comment\": \"We would like to thank the reviewer for their time and helpful feedback. We particularly appreciate checking the proofs in the Appendix. The following are our responses to reviewer\\u2019s comments.\\n\\nComments 1&2: (Clarification on $v_1 \\\\ge v_2 \\\\ge \\\\ldots \\\\ge v_k \\\\ge 0$ constraint)\\nAny permutation of ($v_i$, $s_i$)\\u2019s results in the same k-bits quantization. To remove ambiguity we assume that $v_1 \\\\ge v_2 \\\\ge \\\\ldots \\\\ge v_k \\\\ge 0$. Note that this constraint only removes the ambiguity but we still consider all the possible k-bits quantizations. Thanks for pointing to a potential confusion. We clarified this in the revised draft after equation (6) and before equation (9).\", \"comment_3\": \"(\\u201cFeels like the draft can be compressed into 8 pages, even if the work looks nice\\u201c)\\nAs suggested we compressed the draft into 8 pages.\\n\\nThe network architecture we use in the paper is a standard binary network (slightly improved implementation of [1]) and close to the state-of-the-art (see Table 2). Our goal in this work is on improving quantization for binary neural networks. In terms of quantization, the proposed methods are provably optimal. We empirically demonstrate that optimal quantization leads to improved accuracy on this strong baseline. We note that, there are other directions to further improve the accuracy of BNNs (architecture optimization, training methods, hyper parameters, etc.), all complementary to the proposed optimal quantization. \\n\\n[1] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In European Conference on Computer Vision, pp. 525\\u2013542. Springer, 2016.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This work proves seval binary quantization guarantee for the 1 or 2 bit cases, and shows some empirical error analysis.\\n\\nI am not an expert on neural network compression so I am not quite sure how the proposed method compares with the state-of-the-art algorithms. On the other hand, I checked several proofs provided by the authors for the 1-bit and 2-bit quantization cases. The proofs look good to me.\", \"some_minor_comments\": \"1. For the definition (9) can the authors make it clear that it is for all v_j s.t. v_1>= v_2>=.... instead of \\\\exits v_j?\\n2. Can authors provide some explanation why in (8) we want to have v1>=v2>=vk in the constraint? I understand we need that in the proof, but is there any reason this is also the case in empirical evaluation? To me we may also have cases such that v1 < v2, is there any guarantee for those cases?\\n3. Feels like the draft can be compressed into 8 pages, even if the work looks nice.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a new family of quantized neural networks, where the weights are quantized based on fixed precisions. The authors proposed the optimal 1-bit/2-bit/ternary quantization schemes based on minimizing the L_2 loss. The authors proposed a greedy algorithm to approximate the optimal k-bit quantization. The authors showed through extensive experiments on real DNNs that the proposed optimal quantization (Opt in the tables) can give better generalization as compared to several state-of-the-art alternative methods.\\n\\nThe reviewer appreciates that this quantized DNN-paper has a theory. This theory is based on minimizing the L_2 loss between the original data and quantized data. It unifies several quantization schemes into one framework. This is in contract with technical papers in the area which propose a single quantization method. The proposed quantization can further be implemented efficiently using XNOR operations and bit-counting.\\n\\nThe writing is of good quality. The length is a bit larger than the recommended length of 8 pages.\\n\\nThe reviewer has the following concerns,\\n\\n- This paper motivates from rank-k quantization but implemented as a scaled quantization. At the end of page 3, the authors showed that the scaled quantization can somehow approximate the rank-k quantization. However, the approximation is loose without any guarantee. Overall, I don't understand how the low-rank quantization in section 2 is related to the proposed method and fits in the overall picture. Ideally, in section 2, the authors can have some theoretical statements to state the optimality of the rank-k quantization. Then, the authors can say that, because of practical difficulties to implement the SVD, the use the scaled binary quantization instead. Or, the authors can simply remove section 2, and add a paragraph to introduce rank-k quantization, which is in contract to the scaled quantization they used in the paper.\\n\\nSomewhere in the text, the authors have to explain the optimality is with respect to the L_2 loss. There can be alternative quantization based on different losses.\\n\\nFigure 4, could you explain why the angle trends up as the layer index increases?\\n\\nAfter equation(4): mention the choice of p(x)\\n\\nIn conclusion, analyz -> analyze, introduc->introduce\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"this paper looks at different quantization schemes (replace values of vector in \\\\R to, basically, plus or minus v, for well chosen v).\\n\\nDifferent schemes are considered, namely the low rank binary quantization (a matrix is approximated by the componentwise product of a low rank matrix and a +/-1 matrix) and k-bit binary quantization.\\nFinding the best quantization ends up in solving a program, which is convex and quite straightforward fo k=1 and 2, and unfortunately (apparently) non-convex for k>2. So the authors suggest a greedy approach for k>2.\\n\\nThe motivation of this work is DNN, arguing that quantized vectors should improve computations cost, hence some experiments are provide on DNN, yet they only illustrate the fact that quantization does not deteriorate too much the learning & test error.\\n\\nThe main criticisms is therefore the lack of concrete evidences that those schemes are actually helpful and, on the other hand, the relative simplicity (so the theoretical part of the paper are not sufficient by itself)\"}" ] }

[ "Li Tangqing", "Wang Zheng", "Liu Siying", "Daniel Lin Wen-Yan" ]

[ "deep", "unsupervised anomaly detection", "novel", "anomalies", "large datasets", "unsupervised setting", "key idea", "algorithm", "representation" ]

{ "note_id": [ "5bgvZ4BKJ", "SyxMGkB2or", "rylzfReniH", "rylFrTg3sH", "rkxuJ6x3sS", "SklXH3l2jB", "SJx1fc7aKr", "B1g4zUaotB", "SJlQMrZjtS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732411, 1573830409870, 1573813770449, 1573813569215, 1573813472376, 1573813306595, 1571793415111, 1571702283800, 1571652875185 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1782/Area_Chair1" ], [ "ICLR.cc/2020/Conference/Paper1782/Authors" ], [ "ICLR.cc/2020/Conference/Paper1782/Authors" ], [ "ICLR.cc/2020/Conference/Paper1782/Authors" ], [ "ICLR.cc/2020/Conference/Paper1782/Authors" ], [ "ICLR.cc/2020/Conference/Paper1782/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1782/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1782/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors propose an approach for anomaly detection in the setting where the training data includes both normal and anomalous data. Their approach is a fairly straightforward extension of existing ideas, in which they iterate between clustering the data into normal vs. anomalous and learning an autoencoder representation of normal data that is then used to score normality of new data. The results are promising, but the experiments are fairly limited. The authors argue that their experimental settings follow those of prior work, but I think that for such an incremental contribution, more empirical work should be done, regardless of the limitations of particular prior work.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Reviewers, any comments on the author response?\", \"comment\": \"Dear Reviewers, thanks for your thoughtful input on this submission! The authors have now responded to your comments. Please be sure to go through their replies and revisions. If you have additional feedback or questions, it would be great to know. The authors still have one more day to respond/revise further. Thanks!\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"Thank you for your valuable time and comments\\n\\n1. Comments on formatting and typographical errors.\\n\\nThank you for your careful reading. We have revised our paper per your suggestions. \\n\\n2. Is the DAGMM method SOTA in anomaly detection with deep autoencoder? \\n\\nTo our best knowledge, it is the SOTA for UNSUPERVISED anomaly detection. We have compared against GAN methods as recommended. However, available GAN methods, such as IGMM-GAN and GPND mentioned by the reviewer, require labels of normal data for training discriminator. Our method, despite being fully UNSUPERVISED, performed competitively. \\n\\nFor IGMM-GAN, unfortunately we did not find publicly available code, so we cannot compare the performance. \\n\\n For GPND, when rerun on MNIST dataset with \\u20184\\u2019 as normal data, we found out that the source code uses training label information to search for the optimal threshold for F1/AUROC calculation, so they can report very good result. This trick gives the method unfair advantage. For 20% of anomaly, they achieved F1 score of 0.966 and AUROC score of 0.9737. For 50% of anomaly, they achieve even higher AUROC of 0.9748. This makes the GPND results dubious. \\n\\n We also evaluate another semi-supervised GAN method done by Zaneti2018*. This paper, is claimed to be the improved version of \\u201cEfficient GAN-based anomaly detection(https://arxiv.org/abs/1802.06222), mentioned by the reviewer. Please note that it is still semi-supervised because it requires training with normal data only. Their code is publicly available. The re-run on KDD dataset shows that the best AUC-ROC and AUC-PRC achievable is 0.9925 and 0.9325; whereas for our method, the AU-ROC on KDD dataset is 0.935. Despite being unsupervised, our method performs competitively.\\n\\n*Adversarially Learned Anomaly Detection(Zaneti, ICDM2018) (https://arxiv.org/abs/1812.02288) \\n\\n3. Why was the only selected digit for analysis is 4? \\n\\nAs 2 of published works*, that we are comparing with, only use digit 4 (as normal data) for analysis, we believe this is an established benchmark. We therefore follow the protocol specified in Zhou2017 and Chalapathy2018. We tested our method by replacing digit 4 with other digits and it achieved similar results.\\n \\n We note that in some literatures, such as \\u201cEfficient GAN-based anomaly detection\\u201d(https://arxiv.org/abs/1802.06222) and \\u201cA survey on GANs for anomaly detection\\u201d(https://arxiv.org/abs/1906.11632), each figure in MNIST dataset is treated as anomaly data, while the rest of the digits are treated as normal ones. We did not adopt this setup, because it is more suitable for semi-supervised methods who are given the labels of normal classes, such that discriminators from normal data can be learned. In this setup, when there\\u2019s a subset in the test data the network never \\u201csees\\u201d, it is deemed \\u2018anomaly\\u2019. We do not think it is truly unsupervised.\\n \\nTherefore, in our evaluation we follow the methodology of the 2 afore-mentioned UNSUPERVISED works, Zhou2017 and Chalapathy2018, for a fairer comparison on the MNIST performance.\\n\\n*Anomaly Detection with Robust Deep Autoencoders(Zhou, KDD2017)\\n(https://www.eecs.yorku.ca/course_archive/2017-18/F/6412/reading/kdd17p665.pdf)\\n*Anomaly Detection using One-Class Neural Networks(Chalapathy, KDD2018)(https://arxiv.org/pdf/1802.06360v1.pdf)\\n\\n\\n4. Other metrics like AUPRC, F1-scores will better reflect the work of the algorithms in comparison.\\n\\nThank you for the advice. We have added evaluation based AUPRC metric in our revision. Please refer to Table 3 in the revision.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We appreciate your constructive comments.\\n\\n1. How is performance affected by very low ratio of anomalies? Results on 2%, 3% anomalies?\", \"we_assessed_our_model_on_datasets_with_smaller_anomaly_ratios\": \"1%, 2%, 3%, 4% and 5%. Both averaged AUROC and AUPRC scores indicate consistent and robust performance.\\n\\n| Anomaly ratio | AUROC | AUPRC |\\n| ------------------- | ---------------|--------------- |\\n| 1 | 98.2 +- 0.1 | 90.0 +- 1.0 |\\n| 2 | 98.1 +- 0.1 | 91.8 +- 0.2 |\\n| 3 | 98.4 +- 0.2 | 92.4 +- 0.1 |\\n| 4 | 98.2 +- 0.1 | 91.3 +- 0.9 |\\n| 5 | 99.0 +- 0.3 | 94.5 +- 0.1 |\\n\\n\\n2. Comparison with Zhou 2017 paper?\\n\\nSimilar to Zhou\\u2019s paper, our method does not require any noise-free training data for model training. Both Zhou\\u2019s method and our method does not need any data label. However, Zhou\\u2019s paper use L1 or L2,1 as penalty terms, while our method combined reconstruction error and the learned low-dimensional representation (bottleneck feature from AE) in the training process. In Zhou\\u2019s paper, it has been reported that the best F1-Score (based on threshold that gave the best trade-off between precision and recall) for MNIST dataset where \\u201c4\\u201d is treated as normal data (remaining digits are anomaly) is 64%(Figure 5). In our paper, the ROC figure reported is (70.5+/-2.1)%, which validates the performance of our method. Unfortunately, Zhou\\u2019s paper does not evaluate on datasets other than MNIST.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We thank the reviewer for the comments. We feel there might be misunderstanding that needs to be clarified.\\n\\n1. \\u201cNovelty of this paper is in doubt. .. reconstruction error of AE has been explored thoroughly. Extension seems to be straightforward. \\u201c\\n\\nThe novelty of this paper lies in the fusion of data selection (based on clustering) and representation learning for anomaly detection under a fully unsupervised setting. By incorporating clustering to provide supervisory signals, we iterate between hypothesizing normal candidate subset and representation learning. This framework is carefully engineered to distill out anomalous data and improve the learned representation of normal data. \\n\\nWe would like to clarify that our method did not use reconstruction error alone. Instead, it combined reconstruction error and the learned low-dimensional representation (bottleneck feature from AE) in the training process. Using reconstruction error alone is insufficient for classifying normal/anomalous samples in the training process. The reasons are two-fold: \\n\\nFirst, in the early iterations, as our algorithm adopts a tight cut-off to guess the normal candidates for auto-encoder training, some excluded but true normal samples will see large reconstruction errors. Second, during the intermediate steps, it is hard to draw a decision boundary based on the reconstruction errors (scalar values) alone. Empirically, we have tested using only the reconstruction error for all our experiments and it was confirmed to perform poorly. When combining cosine error and bottleneck feature, we re-normalized the vector after concatenation. We found such combined feature more discriminative for training set selection. The bottleneck feature, working as a low-dimensional embedding, improves the similarity measure in a clustering process. We will add this discussion to the revision.\\n\\n2. Dependency on initialization\\n\\nWe discussed this issue in the submission. Our method starts with a tight variance cut-off to ensure the initial training set is as pure as possible. This means some of the true normal samples could be discarded in this initialization. Our assumption is that given partial normal data, the autoencoder would be able to generalize well on the \\u201cunseen\\u201d normal data that was discarded, and progressively recover them as iterations go on. We also presented in Fig. 4, Appendix section that empirically, low-variance clusters correspond well to normal samples, while higher cluster variances suggest the presence of anomalous samples. In Fig. 5, we showed the convergence behavior, with initial threshold p_0 (assumed anomaly percentage) set to 30%, while ground truth is 20%. \\n\\n3. Problematic refining process\\n\\nWe would like to address that the re-clustering of data in each iteration, is not solely based on reconstruction error. Our error handling mechanism which combines reconstruction error and the learned low-dimensional representation (bottleneck feature from AE), improves the performance of method from pure clustering, as validated in Tab 3, right-most column.\\n\\n4. Poor result on MNIST dataset \\n\\nWe explained in our submission that this is due to the MNIST data being innately not high-dimensional. The NetVlad feature may have been an overkill. We re-run our experiment using raw images, and it achieved AUROC 0.824 and F1-score 0.97 (based on threshold that gave the best trade-off between precision and recall). In terms of AUROC, this result is ~30% better than state-of-the-art unsupervised method, DAGMM. Also, we would like to bring to your attention that our results on more complex dataset are significantly better compared to other UNSUPERVISED methods, and comparable with semi-supervised methods. As far as we know, we are the first paper to obtain such a good performance on unsupervised anomaly detection on video data.\"}", "{\"title\": \"General Response to All Reviewers\", \"comment\": \"We would like to thank the reviewers for their valuable time and comments. Our reviews are mixed, with R1=Weak Accept, R2=Reject, and R3=Weak Reject. We believe we can address the concerns raised by the reviewers that have led to the lower scores, especially R2 who has raised concerns on the novelty of the proposed method. We have re-run our KDD experiments on the suggested algorithms and added the AUPRC evaluation metric. In this regard, our method demonstrates the best performance among unsupervised methods, and comparable performance to semi-supervised counterparts. We would like to emphasize that our proposed method is an UNSUPERVISED method which did not use any label information, even in the training phase. Some methods, although claimed as unsupervised, require label information, because only normal data is used in the training. For example, in the GAN-based methods, normal data is fed into the discriminator during training. Hence, methods such as IGMM-GAN or GPND (mentioned by R3) should be classified under semi-supervised. It should be noted that while comparing with semi-supervised methods can put our method into perspective, it is UNFAIR to compare our method directly with semi-supervised methods.\\n\\nMoreover, we would like to highlight that our model works with the least assumption on the data itself. Our ONLY assumption is that the anomalies are not statistically dominant in the entire dataset (and for this exact reason they are anomalies by nature). We do not make assumptions on how many classes are present in the normal data, unlike the semi-supervised methods, which require pre-isolation/identification of classes of normal data. The superiority of our method is more evident in finding undefined anomaly in video data, as demonstrated in section 4.3 of the paper.\\n\\nWe sincerely hope the reviewers to reconsider their scores after our rebuttal, we believe we can address all concerns in the final revision. \\n\\nWe have incorporated the comments in our revision (please refer to revised.pdf). The changes are highlighted in red.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper proposed an unsupervised anomaly detection method for the scenarios where the training data not only includes normal data but also a lot of anomaly data. The basic idea of this paper is to iteratively refine the normal data subset, selected from the whole training data set. Specifically, the paper first train an auto-encoder (AE) and determine which are the normal data samples according to the reconstruction errors. Then, using the normal data to retrain the AE again, and re-select the normal samples. Repeat the above two steps until convergence.\\n\\nOverall, the novelty of this paper is in doubt. Detecting anomaly by reconstruction error of AE has been explored thoroughly, and this paper only extends it to iteratively select the normal samples. The extension seems to be very straightforward. \\n\\nAlso, the refining process is also problematic. It will highly depend on the initial selection, and the error will be propagated to subsequent detections. How to determine which data samples are anomalous is a key to the success of the model, but the proposed method based on the variance assumption is too intuitive and not convincing.\\n\\nIn addition, the experimental results on the very simple MNIST task is very poor, putting the effectiveness of the proposed model in doubt.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": [\"In this paper the authors propose a framework for anomaly detection. The method is based on autoencoders and reconstruction error, but instead of training the autoencoder using all the data-points, the method iteratively uses some form of clustering to determine the points which presumably belong to the normal set, and uses them for training the autoencoder. This helps make the method robust when the portion of anomalous data-points is high.\", \"The paper is well-written and clear and the proposed architecture is novel to my knowledge. Nevertheless, I have the following concerns about the paper. Given clarifications in the author response, I would be willing to increase the score.\", \"In terms of novelty, separating the dataset into normal + outliers/noise is not novel (Zhou 2017 cited in the paper). The novel part here is perhaps using variance and clustering for making the separation. However, using variance is not well motivated and it is referred to Figure 4, in which the argument is not clear. Similarly, why the reconstruction error is included in the latent representation (eq 5) is not clear.\", \"Given the similarity of the idea to Zhou 2017, the comparison seems important (if the code is available), and also proper discussion of such similar works is required, which currently is not presented.\", \"How is the performance affected by very low ratio of anomalies? This can be shown by including %2,%3 of anomalies in Table 2.\", \"The sensitivity of the results to the choice of hyper-parameters: p0, p, and r is not clear, and how these parameters are chosen is not discussed. It would be interesting to see how the performance is affected by different choices of the hyper-parameters.\"]}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #3\", \"review\": \"In the article, the authors solve the problem of anomaly detection using a fully unsupervised approach. They try to deal with the main challenge of anomaly detection: a lack of certainty on what defines normal data and anomaly one. For this purpose, the authors iteratively use: 1) autoencoders to learn the representation of the data; 2) applying in the latent space clustering to get a new training set and retrain autoencoders. The experimental results show that the author\\u2019s method performed better results than such a baseline model as one-class SVM and one-class NN.\\n \\nThe proposed algorithm looks robust and well-motivated, but the text of the article and the experiments can be improved. As the proposed approach is a heuristic, the experiments should be done more persuasively, including more metrics used and more alternative algorithms considered.\", \"the_key_comments_are_the_following\": \"1. The formatting of the article needs to be improved e.g.:\\n2. there is no comma between rows in the equation (1) ;\\n<<to be accepted into the \\u201ctraining\\u201d set, .>> - there is an extra comma;\\n3. round brackets in the equation (6) should be bigger;\\n4. Table 3 is bigger than the page sizes.\\n5. The quality of the pictures should be improved:\\n- Increase the captions font size in Figure 2;\\n- The captions and the legends in Figure 3 are practically not visible;\\n6. Is the DAGMM method SOTA in anomaly detection with deep autoencoder? There are many other methods with similar ideas. We expect that we should provide a comparison with other methods:\", \"https\": \"//arxiv.org/pdf/1906.11632.pdf\\n8. Class imbalances can affect the value of the AUROC metric. Possibly, the other metrics like AUPRC, F1-scores will better reflect the work of the algorithms for comparison. Also, AUROC is not representative when it comes to the selection of the threshold for anomaly detection. Precision and Recall can help to get more insights.\\n9. In Table 3, the result of applying the proposed algorithm presented with standard deviation, but other methods are represented by one metric value. Why? The explanation is required.\"}" ] }

[ "Timothée Lacroix", "Guillaume Obozinski", "Nicolas Usunier" ]

[ "knowledge base completion", "temporal embeddings" ]

{ "note_id": [ "SW4GeNFjG_", "BJlXSxDVsH", "Ske_fgwVsH", "rye601PVoB", "r1eF91vVjB", "SJeB8K5AYS", "SkeVp-x0FS", "BJeL9wg4YB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732382, 1573314618726, 1573314575652, 1573314516845, 1573314448671, 1571887437466, 1571844540370, 1571190670001 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1781/Authors" ], [ "ICLR.cc/2020/Conference/Paper1781/Authors" ], [ "ICLR.cc/2020/Conference/Paper1781/Authors" ], [ "ICLR.cc/2020/Conference/Paper1781/Authors" ], [ "ICLR.cc/2020/Conference/Paper1781/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1781/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1781/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The authors propose a new algorithm based on tensor decompositions for the problem of knowledge base completion. They also introduce new regularisers to augment their method. They also propose an new dataset for temporal KB completion.\\n\\nAll the reviews agreed that the paper addresses an important problem and presents interesting results. The authors diligently responded to reviewer queries and addressed most of the concerns raised by the reviewers. \\n\\nSince all the reviewers are in agreement, I recommend that this paper be accepted.\", \"title\": \"Paper Decision\"}", "{\"title\": \"detailed comments\", \"comment\": \"1 The technical novelty is a bit limited in compared with those 3-order tensor factorization based methods.\\n\\u2192 Despite a simple modification in the model, we believe our study of regularizers and their effects is non-trivial. Our model provide a large improvement over the previous state of the art for temporal knowledge base completion on all benchmarks considered.\\n\\n2 For the time-aware link prediction problem, some deep learning based methods may also perform well, especially those capturing sequential signals, e.g., RNN and their variants. The authors are encouraged to include those works in the related work and in the empirical studies.\\n\\u2192The work we include from Garcia-Duran et al. 2018 uses RNN. Its performances are included in Table 2 as TA. The other works we were able to find that use RNN are Lei et al. 2019 and Chen et al. 2018. They focus on much smaller graphs than ours: maximum number of node 256 for Lei et al., 274 for Chen et al., while we work with up to 400k entities.\\n\\nLei et al. 2019 GCN-GAN: A Non-linear Temporal Link Prediction Model for Weighted Dynamic Networks\\nChen et al. 2018 GC-LSTM: Graph Convolution Embedded LSTM for Dynamic Link Prediction\"}", "{\"title\": \"Detailed comments\", \"comment\": \"1. The work seems to assume that the time is discretized by year. However, it is unclear how one would deal with a KB where the temporal relations can change on the order of weeks (example popular movies). For that matter, how would the model be modified to deal with heterogenous time scales in the evolution of the relations in the KB? Can the authors add some clarification as well as explanations for how this would be addressed in this formulation?\\n\\u2192 Changing the discretization of time is not an issue unless the number of timestamps to be considered becomes too large. For the exemple of movies, with 52 weeks per year, this would lead to 5200 timestamps for 100 years worth of movies, which is completely manageable. As mentioned in our answer to Reviewer #2 however, the temporal regularizer becomes more important when the granularity along time increases.\", \"regarding_heterogeneous_timescales\": \"in the temporal regularizer, we penalize the difference of embeddings of successive timestamps, irrespective of the actual timestamps. In preliminary experiments, we also tried weighting by the time difference: ||T_i - T_{i+1}||_2^2 / (ts_{i+1} - ts_i), where T_i is the embedding corresponding to timestamp ts_i. This reduces the temporal regularization for embeddings of timestamps that are far from their neighbors. This did not lead to improvements on the datasets considered.\\n\\n2. While there is a section devoted to different regularization, from Table 3, the impact of choosing the right regularizer seems minimal in terms of MRR. Also, it seems that the results in Table 3 are comparable to the \\u201cranks multiplied by 10\\u201d setting in Table 2. What is the reason for this choice?\\n\\u2192 Impact of regularizer: the impact is e.g., 2 points of MRR absolute on ICEWS-5, which is substantial. We consider our results with higher ranks as the new state-of-the-art (since performances are significantly better than previous methods), so we compare the regularizers in Table 3 in this most challenging and important setup.\\n\\n3. For Table 4, why is there a performance difference between static and non-static relations? It would help if the authors could provide some more error analyses to dive into this performance difference - is it just data imbalance or inherent difficulty in the task for temporal relations.\\n\\u2192 The difference in performance is due to a discrepancy in temporal vs non-temporal data. Notably, 93% of temporal tuples (subject, predicate) have two or more valid objects. This proportion is only 47% for non-temporal tuples. One-to-many relations lead to lower MRR than one-to-one relations (see table 3, Lacroix et al. 2018 Canonical Tensor Decomposition for Knowledge Base Completion) \\n\\n4. Section 6 talks about enforcing another auxiliary loss, however, these results are not part of the table. I would urge the authors to add this to Table 4. Also, the loss in MRR mentioned is it the average loss or does enforcing this auxiliary loss also influence performance on T-MRR as well? If so, might it be that the auxiliary loss is too strong and might need to be penalized? Did the authors try penalizing the auxiliary loss?\\n\\n\\u2192 The auxiliary loss has no influence on the T-MRR, the point of MRR is lost over non-temporal triples. We did not try reducing the weight of this loss in the overall loss to obtain the best (MRR, time AUC) operating point as this section is meant to be an example of new metrics that could be interesting on this dataset.\\n\\nFinally, the graph in Figure 2 is hard to follow when printed in black and white. What would the plot look like in comparison to a model that does not enforce the auxiliary function for the same example?\\n\\n\\u2192 We will add different strokes for each presidents in the final version of the paper. On this example, the plot for TNTComplEx trained without the auxiliary loss gives similar scores to the presidents over time.\\n\\nHaving some representative examples would make it easy to understand where these models differ in their performance.\\n\\u2192 The breakdown we provide shows that the difference comes from non-temporal triples, which is expected when comparing a temporal and non-temporal model. Since the non-temporal triples account for 90% of the link prediction loss, it is natural that adding a new loss will impact in majority the MRR of these triples and leave the MRR of temporal triples intact, which is what we observe. \\n\\n5. In the results, can the authors include metrics like filtered MRR as well as Hits@10, similar to the one suggested by Bordes 2013? This would make it easier to compare against previous literature results which all seem to be reporting on these metrics.\\n\\u2192 We report the filtered MRR similarly to previous state of the art on these datasets. This is specified in the experimental set-up section. To avoid clutter, we added hits@k in the supplementary materials (Appendix 8.5).\"}", "{\"title\": \"detailed comments\", \"comment\": \"1. For novelty, the paper does not clearly point out that Ma et al. (2018) already augmented the ComplEx model with time embedding for completing temporal knowledge graphs.\\n\\u2192 We already cite Ma et al. in the related work section. Please note that they *do not* extend the Complex model. To the best of our understanding, this paper proposes an extension to Tucker in which a new core tensor is learned for every timestamps. This leads to many more parameters than this work. Their results are below the other results we include in our state-of-the-art comparison.\\n\\n2. The paper does not point out whether the units of the timestamps affects the model and how to adjust the model accordingly. For example, the time granularity of the ICEWS dataset is 24 hours. If we switch the unit of timestamps from hours to days, do the results of the proposed model change? If yes, how to peak the best time unit for a given dataset? \\n\\u2192 At higher granularities, we expect the temporal regularizer to become even more useful, since it should mitigate the lack of data per timestamps. In practice, several granularities could be cross-validated. For ICEWS, we use the granularity of the evaluation proposed in previous work.\\n\\n3. For the Wikidata, the author reports the filtered Mean Reciprocal Rank of the conducted experiments, where the author provides not only the overall score but also the temporal MRR. However, the paper does not provide information about error bars as well as the unfiltered version of the experiment results. Since the results of TComplEx and TNTComplEx are only slightly better than ComplexE, the reviewer doubts whether the proposed model can really beat the ComplEx model when considering the error bars. Also, NT-MRR and T-MRR are not clearly defined.\\n\\n\\u2192 NT (NonTemporal-MRR) is the MRR computed over test triples that have no timestamps associated. T-MRR is the MRR computed over test triples that have at least 1 timestamp associated. We do not provide unfiltered metrics because these tend to saturate more quickly than filtered ones. The standard deviation for non-temporal queries is 0.0035 for TNTComplEx on non-temporal relations, making the 0.02 MRR difference we observe significant.\\n\\n4. The submission proposes a new large-scale temporal event dataset, and states that this dataset is more realistic than the current benchmarks. However, the reviewer does not find any argument in the paper to support this statement. \\n\\u2192 We rephrased the conclusion of the paper. A knowledge base such as wikidata, freebase or yago all have timestamps naturally attached to the triples they represent. We believe this temporal data to be an essential part of the knowledge base and thus it should be included in knowledge base completion benchmarks. Moreover, we believe that the scale of this dataset is a good step in the direction of completion at the scale of full knowledge bases such as wikidata.\"}", "{\"title\": \"General comments\", \"comment\": \"We thank the reviewers for their comments. We will address general concerns here and go into more details in separate comments. We added the standard deviation of TNTComplEx on each datasets, computed over 5 runs, to the supplementary materials (Appendix 8.5). The standard deviation is of the order of 1e-3, which is consistent with our use of boldface in results tables.\", \"regarding_the_performances_of_tntcomplex_compared_to_complex_on_wikidata\": \"we show that adding a new loss, our temporal model can answer queries that a non-temporal model cannot such as \\\"When was Jacques Chirac the president of France ?\\\". We further show that our TNTComplEx model is capable of answering these queries, while maintaining an overall MRR that is competitive with ComplEx.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Review:\\nThis paper extends the ComplEx model (Trouillon et al., 2016) for completing temporal knowledge bases by augmenting it with timestamp embeddings. Besides, based on the assumption that these timestamp representations evolve slowly over time, the paper introduces this prior as a regularizer. Also, the paper adds a non-temporal component to the model to deal with static facts in knowledge bases. The proposed model has been evaluated using the current benchmark temporal event datasets, showing state-of-the-art performance.\\nThis paper could be weakly accepted because the paper introduces new regularization schemes based on the tensor nuclear norm for tensor decomposition over temporal knowledge graphs, which could be a significant algorithmic contribution. Additionally, the submission empirically studies the impact of regularizations for the proposed model, and weight different regularizers according to the joint marginal of timestamps and predicates, which achieves good experimental results.\", \"feedback_to_improve_the_paper\": \"1. For novelty, the paper does not clearly point out that Ma et al. (2018) already augmented the ComplEx model with time embedding for completing temporal knowledge graphs.\\n2. The paper does not point out whether the units of the timestamps affects the model and how to adjust the model accordingly. For example, the time granularity of the ICEWS dataset is 24 hours. If we switch the unit of timestamps from hours to days, do the results of the proposed model change? If yes, how to peak the best time unit for a given dataset?\\n3. For the Wikidata, the author reports the filtered Mean Reciprocal Rank of the conducted experiments, where the author provides not only the overall score but also the temporal MRR. However, the paper does not provide information about error bars as well as the unfiltered version of the experiment results. Since the results of TComplEx and TNTComplEx are only slightly better than ComplexE, the reviewer doubts whether the proposed model can really beat the ComplEx model when considering the error bars. Also, NT-MRR and T-MRR are not clearly defined.\\n4. The submission proposes a new large-scale temporal event dataset, and states that this dataset is more realistic than the current benchmarks. However, the reviewer does not find any argument in the paper to support this statement.\", \"references\": \"The\\u0301o Trouillon, Johannes Welbl, Sebastian Riedel, E\\u0301ric Gaussier, and Guillaume Bouchard. Com- plex embeddings for simple link prediction. In International Conference on Machine Learning, pp. 2071\\u20132080, 2016.\\nYunpu Ma, Volker Tresp, and Erik A Daxberger. Embedding models for episodic knowledge graphs. Journal of Web Semantics, pp. 100490, 2018.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The work in this paper is focused on the task of knowledge base completion, dealing specifically with temporal relations, which is quite important in practice, and also not as well studied in literature. Specifically, the authors have 3 main contributions in this paper:\\n1. They present an order 4 tensor factorization for dealing with temporal data, which is a nice extension of the work in Lacroix 2018. The authors introduce different forms of regularization to extend to the order 4 tensors. Inspired by previous work, they produce different regularization strategies for order 4 tensor by unfolding the modes to reduce it ot an order 3 formulation. They also describe a regularization term for smoothing the temporal embeddings. \\n3. Finally, the authors mine wikidata for temporal relations and contribute a dataset based on wikidata that is much larger than existing datasets for this task.\\n\\n\\nOverall, I think this paper is interesting as it provides an incremental extension of ComplEx to the temporal case, and the experiments to support the formulation show improvements in MRR. However, the experiments around standard benchmarks as well as the data produced by the authors do not always support the hypothesis that modeling temporal dimension in the proposed formulation is a win for the KB completion task. For example, the use of auxiliary losses for enforcing temporal constraints makes the overall performance worse. This is mentioned in a section, but I think it deserves a more thorough explanation. Also, there is no mention about statistical significance of the results, and so it is hard to judge the claim of these being SOTA as made by the authors. For example, on the wikidata dataset produced, the ComplEx model outperforms the proposed models.\", \"strengths\": \"1. The work in this paper is quite well motivated and the modeling formulation is clear and easy to follow. The authors did a great job in citing relevant work and walking through the model formulation as well as the various regularization terms introduced.\\n2. The authors compare their work to the previous SOTA - the ComplEx models for multiple datasets, including their own released datasets. They present a clear set of experiments to demonstrate the effectiveness of their approach both on non-temporal and temporal relations.\\n3. There is a dataset being released and also code, which should aid in reproducibility of the results (though I have not tested the code).\", \"areas_to_be_addressed\": \"1. The work seems to assume that the time is discretized by year. However, it is unclear how one would deal with a KB where the temporal relations can change on the order of weeks (example popular movies). For that matter, how would the model be modified to deal with heterogenous time scales in the evolution of the relations in the KB? Can the authors add some clarification as well as explanations for how this would be addressed in this formulation?\\n2. While there is a section devoted to different regularization, from Table 3, the impact of choosing the right regularizer seems minimal in terms of MRR. Also, it seems that the results in Table 3 are comparable to the \\u201cranks multiplied by 10\\u201d setting in Table 2. What is the reason for this choice? \\n3. For Table 4, why is there a performance difference between static and non-static relations? It would help if the authors could provide some more error analyses to dive into this performance difference - is it just data imbalance or inherent difficulty in the task for temporal relations. \\n4. Section 6 talks about enforcing another auxiliary loss, however, these results are not part of the table. I would urge the authors to add this to Table 4. Also, the loss in MRR mentioned is it the average loss or does enforcing this auxiliary loss also influence performance on T-MRR as well? If so, might it be that the auxiliary loss is too strong and might need to be penalized? Did the authors try penalizing the auxiliary loss? Finally, the graph in Figure 2 is hard to follow when printed in black and white. What would the plot look like in comparison to a model that does not enforce the auxiliary function for the same example? I would recommend re-working Section 6 to provide some more details about the performance of the models both across the various datasets as well as error analyses of a few examples compared across TNTComplex, TNTComplex + auxiliary loss + Baseline Complex. Having some representative examples would make it easy to understand where these models differ in their performance.\\n5. In the results, can the authors include metrics like filtered MRR as well as Hits@10, similar to the one suggested by Bordes 2013? This would make it easier to compare against previous literature results which all seem to be reporting on these metrics.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"In this paper, the authors study an important problem, i.e., time-aware link prediction in a knowledge base. Specifically, the authors focus on predicting the missing link in a quadruple, i.e., (subject, predicate, ?, timestamp). In particular, the authors design a new tensor (order 4) factorization based method with proper regularization terms shown in Eqs.(4-6).\\n\\nThe authors also prepare a dataset which may be helpful for further study on this topic, which is highly appreciated.\\n\\nEmpirical studies on three public datasets show the effectiveness of the proposed model over the state-of-the-art tensor factorization based methods that are published very recently.\", \"some_comments\": \"1 The technical novelty is a bit limited in compared with those 3-order tensor factorization based methods.\\n\\n2 For the time-aware link prediction problem, some deep learning based methods may also perform well, especially those capturing sequential signals, e.g., RNN and their variants. The authors are encouraged to include those works in the related work and in the empirical studies.\", \"minor\": \"\", \"typo\": \"'ICEWS05-15'\"}" ] }

[ "Chaoyun Zhang", "Marco Fiore", "Iain Murray", "Paul Patras" ]

[ "spatio-temporal forecasting", "point cloud stream forecasting", "recurrent neural network" ]

{ "note_id": [ "UQM2-EKkyu", "H1goK1W3jS", "H1gC3MZcsr", "H1eRY1JvjS", "S1eohztSjH", "rJgfGGKrsB", "HkgqkZFBjB", "r1l5J0OSsS", "rJxT5QF55S", "ByeeBVfN9S", "BkgVBUphtH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732352, 1573814146941, 1573683894364, 1573478278112, 1573388978948, 1573388810088, 1573388513710, 1573387746475, 1572668309507, 1572246584176, 1571767867574 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1780/Authors" ], [ "ICLR.cc/2020/Conference/Paper1780/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1780/Authors" ], [ "ICLR.cc/2020/Conference/Paper1780/Authors" ], [ "ICLR.cc/2020/Conference/Paper1780/Authors" ], [ "ICLR.cc/2020/Conference/Paper1780/Authors" ], [ "ICLR.cc/2020/Conference/Paper1780/Authors" ], [ "ICLR.cc/2020/Conference/Paper1780/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1780/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1780/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper presents an approach to forecasting over temporal streams of permutation-invariant data such as point clouds. The approach is based on an operator (DConv) that is related to continuous convolution operators such as X-Conv and others. The reviews are split. After the authors' responses, concerns remain and two ratings remain \\\"3\\\". The AC agrees with the concerns and recommends against accepting the paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thanks for your reply, more experiments were added\", \"comment\": \"Thanks very much for taking time to read our response. We have updated the paper to address your concerns, as follows:\", \"q\": \"I think I'd like to see a more controlled experiment, where there is some outlier with opposite response to any neighbor, and we see how its prediction changes as its distance grows.\", \"a\": \"To further show that our CloudLSTM is robust to outliers, we re-run the experiments on the air quality dataset, where we push some weather stations away from the center with different distances, so as to construct artificial outliers. Our experiments show that our CloudLSTM performs equally well when forecasting over inliers and outliers, irrespective of the moving distance of the outliers. Importantly, CloudLSTM achieves significantly better performance over its counterpart PointLSTM (PointCNN + LSTM). New results can be found in Appendix K of the revised paper, and the synthetic dataset will be also released publicly for the sake of reproducibility.\"}", "{\"title\": \"Thanks for a thorough rebuttal; some concerns remain\", \"comment\": \"Overall, I'm still a bit borderline on this paper, but I'm leaning more positive than before based on the thorough rebuttal. If others are voting for acceptance I won't stand strongly against it. I'm glad authors are thinking about a reproducible version of the urban-air dataset and were willing to implement a few baselines I suggested.\\n\\nRE experiments not considering day-of-week / seasonality effects\\n\\nI am not quite satisfied with this explanation. I find it a bit hard to believe that storing (sparser) traffic data long enough to capture day of week effects wouldn't be worth it. Surely there is more Saturday activity in downloading game apps than weekday activity, why throw this away? Plus, I think the dataset you have would let you quantify the value of using longer windows, etc. I think at least showing some concrete comparison of the current window vs a 7 day window would be needed. \\n\\nI looked thru the plots in J, I'm not too convinced the current solution is \\\"good enough\\\" and would not be substantially improved with day-of-week effects.\\n\\nRE baselines with nearest neighbors\", \"one_lingering_issue\": \"it seems weird to use K=3 for your method (which seems to consistently do best), but only let the baseline use K=1 or K=9 or higher.\\n\\nRE outliers\\n\\nI don't think just using the sigmoid to rescale things solves my concern... even if there is some lone data point way far from the existing \\\"normal\\\" data, your approach will still force its predictions to be based off (potentially very unrelated) neighbors, rather than say falling back on some simpler heuristic.\\n\\nI am happy that there is some experimental evidence that the method is somewhat robust to outliers, though.\\n\\nI think I'd like to see a more controlled experiment, where there is some outlier with opposite response to any neighbor, and we see how its prediction changes as its distance grows (and how this differs from simpler approaches).\"}", "{\"title\": \"Paper has been revised\", \"comment\": \"We appreciate the valuable feedback from all reviewers. We have revised the manuscript to address the concerns\"}", "{\"title\": \"Response to Blind Review #1 (2)\", \"comment\": \"Continues from the `````````````\\u201cResponse to Blind Review #1 (1)\\u201d\", \"q\": \"I would suggest avoiding calling the method \\\"$\\\\mathcal{D}$-Conv\\\", and instead use just \\\"DConv\\\", since this is easier to type into search engines and easier to search for in a PDF document\", \"a\": \"Thanks for the suggestion. We change $\\\\mathcal{D}$-Conv to DConv in the revised paper. Regarding the graph embedding, we agree this will have potential and will consider this for future work.\", \"related\": \"Point clouds could be represented as graphs, and then use graph embeddings as feature representations\"}", "{\"title\": \"Response to Blind Review #1 (1)\", \"comment\": \"Thank you for acknowledging the contribution of our paper and offering insightful comments. We address the concerns raised below.\", \"q\": \"Is there a good reason to not try to compare on publicly available datasets like those used in the PointCNN paper (focusing only on the non-temporal versions of the model)? Using proprietary datasets makes following up on this work a bit hard, would be nice to have some reproducible experiment.\", \"a\": \"The datasets in the PointCNN paper are dedicated to point-cloud classification and segmentation, while our work focuses on point-cloud stream forecasting, hence the scope is different. Since the D-Conv, along with the CloudLSTM, are designed for the task of temporal forecasting, we do not test the model on those datasets.\\n\\nWe fully agree however that it would be good to have a reproducible experiment with publicly available datasets. While the mobile traffic dataset can not be released, the air quality datasets are publicly available and can be found at https://www.microsoft.com/en-us/research/project/urban-air/.\\nWe will release the code and processed dataset upon final decision, so as to support the reproducibility of our results.\\n\\nto be continued...\"}", "{\"title\": \"Response to Blind Review #2\", \"comment\": \"Thank you for your feedback and for indicating which aspects can be improved on. We note that some of the comments made are already addressed in the original manuscript.\", \"q\": \"Furthermore, the std is quite large, which makes one wonder if the improvement is statistically significant.\", \"a\": \"We note that the large std exists only in the Cluster A of the air quality dataset. We checked this dataset more carefully, and found that the level of noise therein is more severe than in other case studies. We believe that is the root cause of the larger std. In addition, we note that all models achieve similar std in this dataset (except for MLP), while our proposal obtains the best mean for all metrics, which proves that the improvement is statistically significant.\"}", "{\"title\": \"Response to Blind Review #3\", \"comment\": \"We appreciate the detailed and positive comments, which clearly reflect many of the essential contributions of our work. We answer the points raised in what follows.\", \"q\": \"Appendix B, they claim that the complexity of finding the K nearest neighbors (in dimension L for n points) is close to O(K L log(n)) if using KD trees. I vaguely recall issues in high dimension though (in particular that the above complexity is only valid for specific distributions of points in low dimension). where it is mentioned that L $<<$ log(n) is normally needed to guarantee efficiency. The claim should properly be nuanced.\", \"a\": \"This is correct, and, in fact, in real life the dimensions of a point-cloud dataset are normally 2 or 3. Also, we usually have much more than 3 points in the dataset. Therefore, L $<<$ log(n) should hold for most applications. We clarified this aspect as a footnote in Appendix B.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"=========== Update after rebuttal\\n\\nThanks for the clarifications and the update. I recommend acceptance of the paper and updated to 8.\", \"last_comment\": \"please still improve the appearance of Figure 4 by using a more diverse set of marker shapes as well as overlay and offset tricks -- see https://www.cs.ubc.ca/~schmidtm/Software/prettyPlot.html for an example.\\n\\n============\\n\\nThis paper introduces a new convolution operator (D-conv) specifically tailored to model point-cloud data evolving over time, i.e. a set of n points with features and localization-coordinates that can evolve over time. The main idea is to use the k-nearest neighbor structure for each point to get a fixed size k window to use in the convolution to determine the new location and feature values of a point (and a permutation-invariant operation). The D-Conv operator is included in a LSTM architecture (CloudLSTM) to enable the spatio-temporal modeling of point-cloud data, and can be combined in standard neural network architectures such as a Seq2Seq with attention. This is in contrast to previous approaches which modeled the data on a grid through preprocessing, or did not include the temporal component for cloud data. Experiments is conducted on 4 benchmark datasets, covering two point-cloud stream forecasting applications, showing how CloudLSTM give lower prediction error than numerous baselines and alternatives.\\n\\nWhile the D-Conv idea seems fairly simple and natural, it is novel AFAIK and fairly appropriate to model point-cloud data streams. The approach is well situated in the literature, and the experiments are indicative that this method can improve on the current approaches. I am thus leaning towards accept.\\n\\nThe paper is fairly clear, though the notation is a bit confusing and somewhat sloppy (see detailed comment below).\", \"important_clarification_requested\": [\"the current notation suggests that each channel could have a different location for a point p_n, the K nearest points seem to be defined irrespective on the channel. So is the location fixed across channels; or does this paper allow the neighborhood structures to vary across channel?\", \"== Other detailed comments ==\", \"p.3 Q_n^K -- it seems it would be more appropriate to define it as an ordered list of k points (rather than a set, as this would loose all information about the order); unless you append a new dimension to each point where you put the ordering information there for the purpose of defining the k points in Q_n^K.\", \"(2) the notation is a bit weird and overloaded for the summation (without being defined). Examples include \\\"i in U1\\\" (when U1 is an integer, not a set); \\\"p_n^k in Q_n^K\\\" when p_n^k does not appear in the summation (a clearer alternative would be using the notation v(p_n^k)_i^h for the h^th value of channel i of point p_n^k, e.g.; now p_n^k would indeed appear in the expressoin); \\\"v_n^h in v_n\\\" -> why not just summing over h as it is really doing? Etc.!\", \"(2) S_out^j: each p_n^' should be a *tuple* (not a set like currently written).\", \"Figure 4: the lines are really hard to distinguish just by the similar colors -- please use different markers for the different lines (and offset the marker so that they can be seen)\", \"Several neighborhood sizes are experimented with. Note though that smaller neighborhood sizes are just *special cases* of bigger neighborhood sizes (by using zero weight on the last few neighbors in the convolution). Wouldn't it make sense to use a big neighborhood size and regularize in some way the weights for the further neighbors?\", \"Table 2: for SSIM, there are two rows with 0.69 +/- 0.07 (minimal value) -- they could be both bolded.\", \"Appendix B, they claim that the complexity of finding the K nearest neighbors (in dimension L for n points) is close to O(K L log(n)) if using KD trees. I vaguely recall issues in high dimension though (in particular that the above complexity is only valid for specific distributions of points in low dimension). E.g. see https://en.wikipedia.org/wiki/K-d_tree#High-dimensional_data where it is mentioned that L << log(n) is normally needed to guarantee efficiency. The claim should properly be nuanced.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper proposed a new convolution operator, named dynamic post-cloud convention over spatiotemporal data, and the convolution operator can be embedded in different neural network architectures, like recurrent neural networks. In order to achieve the convolution over point-clouds by using both value features and the spatial-features, given a data point, the convolution is conducted over its k-nearest neighbors generated by CNN. They compared the proposed convolution method by embedding it into RNN, GRN, and LSTM against a number of existing methods on two datasets, in terms of MAE, RMSE, PSNR, and SSIM. Overall, this paper is interesting but needs some clarifications on\\n\\n1. Given that the proposed convolution operator use KNN to choose the nearest neighbors. It would be good to empirically to study how K would affect the performance, does it data-dependent\\n2. Is it possible to study the time complexity for various models?\\n3. Table 1 and table 2 seem to show that the proposed convolution operator contributes to the performance in terms of the mean of each metric. It might be good to do further oblation test to study which mechanism actually contribute to the performance. The choice of RNN, attention, or the new operator? Furthermore, the std is quite large, which makes one wonder if the improvement is statistically significant.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Review Summary\\n--------------\\nOverall this is almost above the bar for me to accept, but I think there's enough concerns about the method and experiments that I'm hesitant. Strengths include the invariance to point cloud order and the relative simplicity of the architecture (compared to PointCNN). Weaknesses include a vulnerability to outliers, experiments that don't seem to think about practical effects like day-of-week in forecasting, and experiments that leave out baselines to help directly assess the impact of neighbors.\\n\\n\\nPaper Summary\\n-------------\\nThe paper develops a new neural net architecture for processing data structured as spatial point clouds that vary over time (e.g. hourly traffic at several antennas spread throughout a city).\", \"the_core_of_the_approach_is_a_new_neural_net_unit\": \"the \\\"D-Conv\\\" operator (See Eq. 2). The output value at each point is obtained via a weighted combination of nearby coordinates and features, using only the K-nearest neighbors (stored in ranked order) to maintain invariance to the original order of points. This layer can be included in modern convolutional (CloudCNN) or recurrent (CloudLSTM) or attention-based architectures in a straightforward way.\\n\\nUnlike many previous methods that require converting point clouds to quantized regular grids, the present approach directly consumes point cloud data. Unlike some existing methods like PointCNN, it avoids information loss (does not reduce dimension from input to output layer).\", \"two_experimental_evaluations_are_conducted\": \"forecasting mobile app traffic across 2 European cities (given past 30 min, predict next 30 min), and air quality across several regions in China (given last 12 hrs, predict next 12 hrs). In both experiments, the locations of the sensors are fixed across time. Fig 4 further looks at traffic forecasting as a function of the lookahead time, from 0-3 hours ahead.\\n\\n\\nNovelty & Significance\\n-----------------------\\n\\nThe paper definitely tackles an important problem (point cloud forecasting). \\n\\nThe present paper's new \\\"D-Conv\\\" operator appears new, though it looks really like a simplification of the PointCNN's \\\"X-Conv\\\" operator rather than a brand new operator. \\n\\nThe most similar work seems to be the PointCNN (Li et al NeurIPS 2018). This work's contribution was a new \\\"X-Conv\\\" operator, which also consumes point clouds and produces learned representations. X-Conv, like the present paper's D-Conv, computes K-nearest neighbors of each point p, but performs first an embedding of each neighbor to a learned \\\"local\\\" feature space and then performs convolution on this embedding. Perhaps the biggest practical difference is that D-Conv has fewer parameters (does not perform the embedding) and does not reduce dimensionality from input to output. \\n\\nTechnical Concerns\\n------------------\\n\\nMy biggest concerns are that the D-Conv has a strong reliance on nearest neighbors. This means the D-Conv has not much accomodation for \\\"outlier\\\" points that are far from others. The X-Conv operator has some nice properties in this regard (it changes coordinate systems so neighbor locations are centered around the current point), but I don't see this in the D-Conv operator, as in Eq. 2, where the coordinate locations are fed directly into the weighted sum after global rescaling to (0,1). I would imagine that data with outliers (whose values are unlike most others) would dramatically hurt performance, as the weights of D-Conv would need to be shared equally by outliers and inliers.\\n\\n\\nExperimental Concerns\\n---------------------\\n\\nIs there a good reason to not try to compare on publicly available datasets like those used in the PointCNN paper (focusing only on the non-temporal versions of the model)? Using proprietary datasets makes following up on this work a bit hard, would be nice to have some reproducible experiment.\\n\\nIt's not clear to me that the experiments here consider realistic scenarios. Why would I predict mobile app traffic using only the past 30 minutes of data? Why predict air quality using only the last 12 hours? Certainly there are time-of-day, day-of-week, and seasonal effects that are all important. At a minimum, I'd think that for the mobile traffic case you could at least look at consuming the last 48 hr of data and predicting the next 30-90 minutes. I suspect that would make even simpler models do much better. \\n\\nFurther, I think the experiments are missing some key simple baselines (or I misunderstand something). For example, rather than the complicated CNN/LSTM architectures, why not try to directly see how much value there is in \\\"neighbors\\\" in this 2d space? At each point, you can make predictions using only the K nearest neighbors' data, with K swept from 1 to 100 or something. I would expect with these features, using just a simple MLP or RNN would do quite well. I'd like to see a stronger qualitative case made for why we expect the complicated DConv weighting operator here to do better than this baselines.\\n\\nOverall, the results tables appear promising (for app traffic forecasting in Table 1, the proposed CloudLSTM achieves 3.66 MAE compared to 4.95 for PointCNN and 4.8 for an MLP). However, it's not clear why and I'd like to understand why. Is it that the other approaches are overfitting? \\n\\n\\nMinor Concerns\\n--------------\\nI would suggest avoiding calling the method \\\"\\\\mathcal{D}-Conv\\\", and instead use just \\\"DConv\\\", since this is easier to type into search engines and easier to search for in a PDF document\", \"related\": \"Point clouds could be represented as graphs, and then use graph embeddings as feature representations\"}" ] }

[ "YoungJoon Yoo", "Sanghyuk Chun", "Jaejun Yoo", "Sangdoo Yun", "Jung Woo Ha" ]

[ "Neural approximation method", "Auto-regressive model", "Sequential sample generation" ]

{ "note_id": [ "HbBlKsMRoZ", "ryxp5HFpFr", "ByxpVK-jtB", "H1l8W_ndtS" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732322, 1571816853186, 1571653940921, 1571502077939 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1779/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1779/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1779/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper presents a technique for approximately sampling from autoregressive models using something like a a proposal distribution and a critic. The idea is to chunk the output into blocks and, for each block, predict each element in the block independently from a proposal network, ask a critic network whether the block looks sensible and, if not, resampling the block using the autoregressive model itself.\\nThe idea in the paper is interesting, but the paper would benefit from\\n- a better relation to existing methods\\n- a better experimental section, which details the hyper-parameters of the algorithm (and how they were chosen) and which provides error bars on all plots (and tables)\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper addresses the sequential limitation of autoregressive model when doing sampling. Specifically, instead of sampling next observations in a sequential fashion, this paper generates future observations in parallel, with the help of i.i.d. future priors. With the help of learned confidence model, the model is able to get trade-off between speed and approximation accuracy. Experiments on synthetic data and image generation with PixelCNN++ show the comparable results while being significantly faster than baseline.\\n\\nOverall the paper is well motivated, with an interesting design of the variational distribution to approximate the true autoregressive distribution. The design of the confidence model looks a bit heuristic, but the trade-off ability between efficiency and quality it brings is also quite interesting.\", \"below_are_some_minor_comments\": \"1. The theoretical analysis is basically comment about the objective which is less interesting. However more interesting guarantees would be: 1) with the additional correction term added, how would it help with reducing the variance; 2) As the q_{\\\\theta, \\\\phi} is always in a limited form due to the parallelism requirement, how bad the approximation could be in the worst case ---- I\\u2019m not asking for these results, but any form of discussion would be helpful.\\n\\n2. The author only compared with the raw PixelCNN++. Would any of the existing AR-speedup method be applicable for a comparison?\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #3\", \"review\": \"The authors consider the problem of sampling time series.\\nTo solve the problem they propose a method that is based on the autoregression model. The novelty here lies in the proposed sampling methods: we start with a sampling of a prior and then try to generate data according to the restored distribution. We learn two functions: signal recovery and confidence prediction.\\nThe main hyperparameter of the algorithm $\\\\varepsilon$ identifies how much samples we accept.\\nThe distinctive feature of the algorithm is speed-up for the sample generation process.\\n\\nWeak reject\\n\\nThere are a significant number of works on video generation, see e.g. [1, 2], references therein and articles that cite these two articles. The problem setting seems pretty similar. It seems like a good idea to compare to these methods (and it seems that video generation is a very resource-demanding procedure, and they don't use parallel applications similar to proposed in the paper. What is the reason?) Most of the approaches use only one frame to generate video, but it seems that LSTM in these methods will benefit from using of multiple frames as input (and will be able to transfer information in autoregression manner by transferring all they need in a hidden state).\\nThe article, in my opinion, will benefit from comparison to these approaches or at least by using some benchmarks from these works to demonstrate feasibility of the considered approach, also it seems that these works are good for demonstration of parallelization capabilities (as in many cases the same idea applies).\", \"not_minor_comments\": \"1. In Figure 2 (a) it is not clear how the data and prediction were generated. According to the procedure in Figure 1 and text we use the same input for all approaches. However solid lines for different epsilons are different.\\n2. The effect of the dependence of recovery of quality for Figure 2 (b) is not explained and is controversial: we get the smallest error for intermediate acceptance ratio, however, there is also a decrease of error if we further increase the gauge threshold (btw the term gauge threshold is new to machine learning community, consider replacement of it)\\n3. More simpler examples will benefit the paper, as we'll be able to know more fundamental properties of the proposed approach.\", \"minor_technical_comments\": \"1. s. 3.1. predictor predicts\\ncommas in equation (8)\\n2. Figure 2: no axis labels for the left plot, use for label \\\"acceptance ratio\\\" red color font & for label \\\"L1 error\\\" blue color font\\n3. Table 1 bracket after l_1 is missing\\n4. Maybe $\\\\sigma$ is not the best designation of confidence, as it can be confused with the variance\\n5. Figure 1: some indexes should be not $x_{i + 2}$, but $x_{i + j}$, $x_{i + M}$. Also, some \\\">\\\" before \\\\epsilon should be \\\"<\\\"\\n6. \\\"a auto-encoder architecture\\\" -> \\n\\\"an auto-encoder architecture\\\"\\n\\n[1] J. He et al. Probabilistic Video Generation using Holistic Attribute Control. 2018. ECCV\\n[2] E. Denton, R.Fergus. Stochastic Video Generation with a Learned Prior. 2018.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper presents a technique for approximately sampling from autoregressive models using something like a a proposal distribution and a critic. The idea is to chunk the output into blocks and, for each block, predict each element in the block independently from a proposal network, ask a critic network whether the block looks sensible and, if not, resampling the block using the autoregressive model itself.\\n\\nIn broad strokes the approach makes sense. It assumes, essentially, that parts of the sequence are hard to predict and parts are easy and, if there are enough easy parts, this procedure should lead to faster inference.\\n\\nThe paper's writing is not ideal. There are some grammatical mistakes that harm reading (for example, the second paragraph of the introduction says \\\"However, these models must infer each element of the data x \\u2208 RN step by step in a serial manner, requiring O(N) times more than other non-sequential estimators\\\", where it is unclear what is O(N) more than what, how is this measured, etc). That said I was mostly able to follow all key points.\\n\\nThe authors do not point out the obvious connection to GANs, which also rely on a critic network to decide whether a sample looks like it comes from the correct distribution, except in GANs the critic is jointly trained with the generator (as opposed to here where it's trained after) and in GANs the critic is only used at training time, while here the critic is used to accelerate sampling (the better the critic the faster this method can sample).\\n\\nI wish the experimental results were a little more explicit about the time vs quality tradeoff; I expected to see more plots with pareto curves, since as-is it's hard to judge the magnitude of the tradeoffs involved. I'd also like a more thorough analysis on why there is a non-monotonic tradeoff in some experiments (table 1, figure 2(b)) between the amount of approximation and the sample quality; this makes me think something else is going on here as this approximate inference method should just decrease quality, never increase it.\\n\\nOverall I lean towards accepting the paper, but I encourage the authors to revise the writing and to add a few plots explicitly showing the time vs quality tradeoff both in likelihood (wrt the full model) and in downstream metrics like FID.\"}" ] }

[ "Kimin Lee", "Kibok Lee", "Jinwoo Shin", "Honglak Lee" ]

[ "Deep reinforcement learning", "Generalization in visual domains" ]

{ "note_id": [ "HRXoeWfG6B", "S1gKJLqijS", "SyxUfr9ojB", "S1lX_V5ijS", "rJePkVciir", "Hke62XQY9H", "Hyl49oREcH", "BJlOAwOpYr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732294, 1573787105197, 1573786893985, 1573786730723, 1573786591418, 1572578228878, 1572297611824, 1571813328429 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1778/Authors" ], [ "ICLR.cc/2020/Conference/Paper1778/Authors" ], [ "ICLR.cc/2020/Conference/Paper1778/Authors" ], [ "ICLR.cc/2020/Conference/Paper1778/Authors" ], [ "ICLR.cc/2020/Conference/Paper1778/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1778/AnonReviewer5" ], [ "ICLR.cc/2020/Conference/Paper1778/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This submission proposes an RL method for learning policies that generalize better in novel visual environments. The authors propose to introduce some noise in the feature space rather than in the input space as is typically done for visual inputs. They also propose an alignment loss term to enforce invariance to the random perturbation.\\n\\nReviewers agreed that the experimental results were extensive and that the proposed method is novel and works well.\\n\\nOne reviewer felt that the experiments didn\\u2019t sufficiently demonstrate invariance to additional potential domain shifts. AC believes that additional experiments to probe this would indeed be interesting but that the demonstrated improvements when compared to existing image perturbation methods and existing regularization methods is sufficient experimental justification of the usefulness of the approach.\\n\\nTwo reviewers felt that the method should be more extensively compared to \\u201cdata augmentation\\u201d methods for computer vision tasks. AC believes that the proposed method is not only a data augmentation method given that the added loss tries to enforce representation invariance to perturbations as well. As such comparisons to feature adaptation techniques to tackle domain shift would be appropriate but it is reasonable to consider this line of comparison beyond the scope of this particular work.\\n\\nAc agrees with the majority opinion that the submission should be accepted.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Common response to all reviewers: short summary of rebuttal\", \"comment\": \"Dear reviewers,\\n\\nFirst of all, we appreciate your efforts in providing valuable comments on our manuscript.\\n\\nTo best respond to the questions and concerns raised by the reviewers, we have carefully revised and enhanced our manuscript with a substantial amount of additional experiments following suggestions from the review comments, including:\\n- Adding oracle performances of PPO agents (Section 3 and 4)\\n- Ablation study (Appendix N)\\n- Discussion on potential failure cases of the proposed method (Appendix M)\\n- Extension of the proposed method to domains with different dynamics (Appendix L)\\n- More clarification on dogs vs. cats experiments (Section 3)\\n\\nRevised parts in the new draft are highlighted in red (in particular, we updated Section 3, 4 and 5, Appendix A, L, M, and N, and Figure 5, 7, 16, and 17).\\n\\nThank you very much.\\nAuthors.\"}", "{\"title\": \"Response to R3\", \"comment\": \"We appreciate your valuable comments, efforts and times on our paper. As you and R4 mentioned, we introduce a simple, yet powerful solution for generalization in deep RL, and show that it achieves significant performance gains in comparison to many known regularization and data augmentation techniques in various environments. Our responses to all your questions are provided below. Revised parts in the new draft are colored by red (in particular, we updated Section 3, 4 and 5, Appendix A, L, M, and N, and Figure 5, 7, 16, and 17).\\n\\nQ1. Related work \\n\\nA1. We cited relevant works that leveraged feature matching loss in Section 3.1, including [1, 2] in semi-supervised learning. Thank you very much for your suggestion.\\n\\nQ2. Random networks on computer vision tasks \\n\\nA2. Our method is designed for improving generalization across different data distributions (between training and testing). Namely, the purpose of our method is different from conventional data augmentation methods. Hence, its effect may not be significant for standard computer vision tasks (e.g., image classification tasks on CIFAR-10 and ImageNet) which consider the generalization in the same data distribution. We remark that \\\"bright vs dark cat/dogs\\\" resembles our problem setting in that the training and test datasets are sampled from a different distribution, so this is the reason why we put this experiment in our paper. Nevertheless, an extension of our method to other computer vision tasks should be an interesting future direction to explore.\\n\\nQ3. State-of-the-art methods \\n\\nA3. In our setting where a domain-specific simulator with various control parameters is not available, the cutout-variant method proposed in [3] (published in ICML 2019) is the state-of-the-art method (denoted as PPO+CO in our paper) and our method outperforms it by a large margin (see Figure 5). In addition, domain randomization [4] that can also be regarded as another state-of-the-art method when the domain-specific simulator is available. Our method outperforms it even without utilizing the simulator (see Table 3). We think this is strong evidence that the proposed method is significantly more effective than any prior method. Furthermore, following R4's suggestion, we included the performance of an agent trained directly on unseen environments to Figure 5 and 7 of the revised draft, which shows an upper bound performance of any method for our purpose. Here, our method is shown to perform very well quite close to this upper bound performance. \\n\\n[1] Miyato, T., Maeda, S.I., Koyama, M. and Ishii, S., Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8), pp.1979-1993, 2018.\\n\\n[2] Xie, Q., Dai, Z., Hovy, E., Luong, M.T. and Le, Q.V., Unsupervised data augmentation. arXiv preprint arXiv:1904.12848, 2019.\\n\\n[3] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In ICML, 2019.\\n\\n[4] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, Pieter Abbeel. Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World. In IROS, 2017.\"}", "{\"title\": \"Response to R5\", \"comment\": \"We appreciate your valuable comments, efforts and times on our paper. As R3 and R4 mentioned, we introduce a simple, yet powerful solution for generalization in deep RL, and show that it achieves significant performance gains in comparison to many known regularization and data augmentation techniques in various environments. Our responses to all your questions are provided below. Revised parts in the new draft are colored by red (in particular, we updated Section 3, 4 and 5, Appendix A, L, M, and N, and Figure 5, 7, 16, and 17).\\n\\nQ1. Potential failure case and beyond visual changes \\n\\nA1. Thank you for the suggestion. As R4 also pointed out, some color (or texture)-conditioned RL tasks can be difficult for our methods to work. However, we remark that our method would not always fail for such tasks (please see our response A1 for R4, Section 5 and Appendix M of the revised draft). Furthermore, to address your concern about limitations, we evaluate our randomization idea on domains with different dynamics. Specifically, we conduct an experiment on the CartPole and Hopper where an agent takes proprioceptive features (e.g., positions and velocities) by changing its dynamics (see Appendix L of the revised draft for more details). Similar to the randomization method we applied to visual inputs, we introduce a random layer between the input and the model. As a natural extension of the proposed method, we consider performing the convolution operation by multiplying a $d \\\\times d$ diagonal matrix to $d$-dimensional input states. For every training iteration, the elements of the matrix are sampled from the standard uniform distribution $U(0.8,1.2)$. One can note that this method can randomize the amplitude of input states while maintaining the intrinsic information (e.g., sign of inputs). Following [1, 2], we measure the performance of the trained agents on unseen environments with a different set of dynamics parameters such as mass. Our simple randomization improves the performance of the agents in unseen environments while achieving performance comparable to seen environments. We believe that this provides evidence for the wide applicability of our high-level idea of using randomization for achieving better generalization beyond visual changes. We added related results and discussions in Appendix L and M of the revised draft. We believe more comprehensive investigation on other domains is beyond the scope of this work, and we leave as future work.\\n\\nQ2. Clarification on dogs vs. cats experiments \\n\\nA2. As shown in [3], it is well-known that grayscaling cannot remove all undesired visual biases such as brightness. However, in dogs vs. cats dataset, we found that our randomization technique can outperform other baselines because it makes DNNs capture more meaningful features such as the shape of objects, rather than texture or color. In other words, our randomization method captures an important property by maintaining a spatial (or temporal) structure of input observation. Due to this, policies trained by our methods can be more robust to environments with different observation spaces in many RL benchmarks. We believe that such invariance to low-level transformations in the observation space features is really important in deep RL as R4 mentioned. We clarified this in Section 3.1 of the revised draft.\\n\\nQ3. Fairness on evaluation \\n\\nA3. Our inference method is not an ensemble method that requires multiple networks, but a stochastic method based on a single neural network with randomized layers. Hence, we believe that the comparison with the baselines and ours is fair because they are all based on a single neural network, i.e., have the same number of learnable parameters. Furthermore, our methods with a single MC sample (i.e., with no aggregation) still outperform all baseline methods as shown in Figure 3(d) and Table 2. \\n\\n[1] Wenxuan Zhou, Lerrel Pinto, and Abhinav Gupta. Environment probing interaction policies. In ICLR, 2019.\\n\\n[2] Charles Packer, Katelyn Gao, Jernej Kos, Philipp Kr \\u0308ahenb \\u0308uhl, Vladlen Koltun, and Dawn Song. Assessing generalization in deep reinforcement learning. arXiv preprint arXiv:1810.12282, 2018\\n\\n[3] Byungju Kim, Hyunwoo Kim, Kyungsu Kim, Sungjin Kim, and Junmo Kim. Learning not to learn: Training deep neural networks with biased data. In CVPR, 2019.\"}", "{\"title\": \"Response to R4\", \"comment\": \"We appreciate your valuable comments, efforts and times on our paper. As you and R3 mentioned, we introduce a simple, yet powerful solution for generalization in deep RL, and show that it achieves significant performance gains in comparison to many known regularization and data augmentation techniques in various environments. Our responses to all your questions are provided below. Revised parts in the new draft are colored by red (in particular, we updated Section 3, 4 and 5, Appendix A, L, M, and N, and Figure 5, 7, 16, and 17).\\n\\nQ1. Handling potential failure case of our methods\\n\\nA1. Thank you for the suggestion to acknowledge the failure cases of our method. As you pointed out, some color (or texture)-conditioned RL tasks can be difficult for our methods to work. For example, consider an extreme seek-avoid object gathering setup, where the agent must learn to collect good objects and avoid bad objects which have the same shape but different colors. However, we remark that our method would not always fail for such tasks if other environmental factors (e.g., the shape of objects in Collect Good Objects in DeepMind Lab [1]) are available to distinguish them. As shown in Appendix M of the revised draft, our method can perform well on the modified CoinRun environment with good and bad coins by capturing the other factors such as the location of coins to perform this task. As another example, consider color-matching tasks such as the keys doors puzzle in DeepMind Lab [1] where the agent must collect colored keys to open matching doors. Even though this task is color-conditioned, a policy trained with our method can perform well because the same colored objects will have the same color value even after randomization, i.e., our randomization method still maintains the structure of input observation. Finally, we remark that our method can handle such corner cases by adjusting the fraction of clean samples during training ($\\\\alpha$ on page 4, in the paragraph \\\"Details of the random networks.\\\"). In summary, we believe that the proposed method covers a broad scope of generalization across low-level transformations in the observation space features as you mentioned. We added related discussions in Section 5 and Appendix M of the revised draft. \\n\\nQ2. \\u201cOptimal\\u201d performance (i.e. upper bound) \\n\\nA2. Following your suggestion, we included the performance of an agent trained directly on unseen environments to Figure 5 and 7of the revised draft. In summary, our method approaches this hypothetical upper bound performance. \\n\\nQ3. The choice of $\\\\alpha$\\n\\nA3. We would like to clarify that the chosen fraction $\\\\alpha=0.1$ of clean samples is set empirically based on controlled experiments. To support this, we added controlled experimental results by varying the fraction of clean samples in Figure 17(d) of the revised draft. \\n\\n[1] Charles Beattie et al. Deepmind lab. arXiv preprint arXiv:1612.03801, 2016.\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper proposes applying random convolutions to the observation space to improve the ability of deep RL agents to generalize to unseen environments. To encourage the learning of invariant features, the authors further include a loss term to align features of perturbed and unperturbed observations. Thorough experiments on multiple generalization benchmarks show that this method outperforms many previously used regularization and data augmentation techniques.\\n\\nAlthough the proposed method is simple, it represents a useful contribution. The need to generalize across low level transformations in the observation space features prominently in several environments, including DeepMind Lab and CoinRun. The clear need for agents to be invariant to these low level transformations well motivates the proposed approach, as does the failure of many existing methods to provide this invariance.\\n\\nThe authors could more explicitly discuss the main drawbacks of this approach. As with any data augmentation, there is an assumption that the applied transformation generally won\\u2019t destroy information pertinent to the task. While this is true for the MDPs investigated here, it is easy to imagine slight variants of these MDPs for which this approach would fail. If an optimal policy must condition on color or texture information from observations, then using these random convolutions would render training impossible. Encountering such MDPs is not farfetched, so this weakness seems worth acknowledging.\\n\\nIn Figure 5 it would be useful to visualize performance of an agent trained directly on these unseen environments, as this presumably serves as an upper bound for the zero-shot performance of \\u201cPPO + ours\\u201d. How close does \\u201cPPO + ours\\u201d come to closing this gap? Without any context on the reward scale, it\\u2019s hard to infer how well this method is generalizing, beyond seeing that it beats some (possibly weak) baselines. Admittedly some closely related curves can be found in Appendix Figures 9 and 14, though they\\u2019re a bit out of the way.\\n\\nSection 3.1 mentions that using alpha = 0 complicates training. It is somewhat surprising that using alpha > 0 is necessary or significant and yet the value used (alpha = .1) is relatively small. Any further comments on this choice?\\n\\nI appreciate the discussion in Appendix F. It\\u2019s natural to wonder about alternative injection sites for the random network, and it\\u2019s good to see how the proposed method compares to these alternatives.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #5\", \"review\": \"This work proposes using a randomly parameterized convolutional layer as additional processing of the input observation to provide data augmentation to make policies more robust to environments with different observation spaces. The empirical results are thorough, comparing with other regularization techniques, including dropout, L2 regularization, and batch normalization with the same policy gradient method, PPO on a variety of generalization in RL benchmarks. There are additional experiments of this method to check that it actually removes visual bias in a computer vision problem better than other methods.\\n\\nThey also incorporate a feature matching loss that explicitly forces the learned representations of equivalent states to be close in L2 distance. While the empirical results are impressive, it is hard to feel excited about this work, which relies on the inductive bias of a randomly parameterized convolution layer to modify the texture of the observation and show it works in certain settings. I'd like more discussion and showcasing of failure modes, it seems that this wouldn't work for settings where the train and test environments are different in ways beyond texture and changes in small objects, and additional analysis in terms of the dogs and cats database about why it performs so much better than other methods. What exactly is the desired and meaningful information in images that a random convolution layer can keep while removing something that is able to generalize to different shades of cats and dogs? Why would this perform better than grayscaling? What about grayscaling and additive Gaussian noise?\\n\\nThe comparison of PPO is also unfair in that the author's method uses an ensemble of policies to act, which other methods do not. A more fair comparison would use ensembles in all other baselines as well or results showing how their method performs without this ensemble.\\n\\nOverall, the presentation, analysis, and writing can all be improved to match the strong empirical results produced.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes methods to improve generalization in deep reinforcement learning with an emphasis on unseen environments. The main contribution is essentially a data augmentation technique that perturbs the input observations using a noise generated from the range space of a random convolutional network. The empirical results look impressive and demonstrate the effectiveness of the method. The experiments are thorough (includes even adversarial attack) and the core method is novel as far as I am aware.\\n\\nThat said, I have a couple of concerns regarding this paper and I would be willing to change my score if authors can address these.\\n\\n1) Feature matching loss (Eq 2) is presented as a novel contribution without referring to related work in semisupervised learning literature. This is essentially consistency training. See:\\na) Miyato, Takeru, et al. \\\"Virtual adversarial training: a regularization method for supervised and semi-supervised learning.\\\" IEEE transactions on pattern analysis and machine intelligence 41.8 (2018): 1979-1993.\\nb) Xie, Qizhe, et al. \\\"Unsupervised data augmentation.\\\" arXiv preprint arXiv:1904.12848 (2019).\\n\\n2) The main contribution appears to be a data augmentation technique where we add a random neural net based perturbation to the state. My question is:\\n\\n*Why don't you first evaluate this on computer vision tasks given that the core idea is data augmentation for images?*\\n\\nIf this technique is so powerful, shouldn't this do a great job in CIFAR10, Imagenet etc? Instead authors only provide a niche example (bright vs dark cat/dogs).\\n\\nIf this can compete with top augmentation techniques on Imagenet (e.g. autoagument), then it can explain the RL performance. Otherwise, please provide some intuition on why this works so well on RL but not as well on computer vision tasks. Is it the unseen environment diversity of RL challenges?\\n\\n3) While proposed method performs well on the benchmarks, it is not clear whether authors compare to the state-of-the-art algorithms. For each task (CoinRun, DeepMind Lab, etc), please explicitly state the best prior result (e.g. Espeholt et al, Tobin et al, Cobbe et al etc) so that proposed method's performance can be better assessed.\\n\\n-------------------------\", \"after_rebuttal\": \"Authors addressed most of my comments. I also found the new experimental results (Fig 5 and 7) very insightful. I increase my score to Weak Accept.\", \"for_future_improvement\": \"More realistic experiments on computer vision tasks (besides cats and dogs) would be welcome. Otherwise, please justify why proposed strategy is particularly good for RL (rather than traditional computer vision benchmarks) in boosting robustness to new domains.\"}" ] }

[ "Jean-Yves Franceschi", "Edouard Delasalles", "Mickael Chen", "Sylvain Lamprier", "Patrick Gallinari" ]

[ "stochastic video prediction", "variational autoencoder", "residual dynamics" ]

{ "note_id": [ "a8Whva4r58", "kxgDTmQY3", "HJxrNhLnjS", "rygBpsU2oB", "BJePei82sr", "r1xmnYInjH", "S1efBB1AYS", "r1xt2GQpFH", "r1gsleMpKS" ], "note_type": [ "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1583400833946, 1576798732264, 1573837869200, 1573837757010, 1573837550632, 1573837227152, 1571841338374, 1571791536871, 1571786739013 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Paper1777/Authors" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1777/Authors" ], [ "ICLR.cc/2020/Conference/Paper1777/Authors" ], [ "ICLR.cc/2020/Conference/Paper1777/Authors" ], [ "ICLR.cc/2020/Conference/Paper1777/Authors" ], [ "ICLR.cc/2020/Conference/Paper1777/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1777/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1777/AnonReviewer3" ] ], "structured_content_str": [ "{\"title\": \"New Version & Changes\", \"comment\": [\"A new version of our article was accepted to ICML 2020 (https://proceedings.icml.cc/book/3249.pdf), with the following main improvements:\", \"we included new results on the challenging Human3.6M dataset, showing that we outperform the state-of-the-art StructVRNN [1];\", \"we computed FVD scores for all tested models;\", \"we clarified the experimental settings and interpretation of the experiment regarding generation at different frame rates;\", \"we improved our results and updated their analysis for the BAIR dataset;\", \"we publicly released the fully documented code at https://github.com/edouardelasalles/srvp;\", \"we publicly released the pretrained models at https://data.lip6.fr/srvp/.\", \"[1] Minderer et al. Unsupervised Learning of Object Structure and Dynamics from Videos. In NeurIPS 2019.\"]}", "{\"decision\": \"Reject\", \"comment\": \"The paper proposes a method for learning a latent dynamics model for videos. The main idea is to learn a latent representation and model the dynamics of the latent features via residual connection motivated by ODE. The architectural choice of residual connection itself is not new as many prior works have employed \\\"skip connections\\\" in hidden representations but the notion of connecting this with ODE and factoring time as input into the residual function seems a new idea. The experimental results show the promise of the proposed method on moving MNIST, KTH, and BAIR datasets. The experiments on different frame rates are also nice. In terms of weakness, the evaluation is performed on relatively simple domains (e.g., moving MNIST and KTH) with static backgrounds and the improvement on BAIR dataset (which is not considered as a difficult benchmark) in terms of FVD is not as clear. For the BAIR dataset, it's unclear how the proposed method will handle the interactions between the robot arm and background objects due to the modeling assumption (i.e., static background). In this sense, content swap results on BAIR dataset look quite anecdotal, and the significance is limited. For improvement, I would suggest adding evaluations on other challenging domains, such as Human 3.6M (where human motions are much more uncertain compared to KTH) and other Robot datasets with more complex robot-object interactions. Overall, the paper proposes an interesting architecture with promising results on relatively simple datasets, but the advantage over existing SOTA methods on challenging benchmarks is unclear yet.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Answer to Reviewer 3\", \"comment\": \"We would like to thank you for your supportive feedback. We updated the submission following your remarks and answer your comments below.\\n\\n - Decoupling appearance and dynamic:\\n\\n Our model manages to decouple appearance and dynamics in the following way.\\n Firstly, the dynamics is made independent from the previously generated frames, in a state-space manner, as described in the manuscript. \\n Secondly, the content variable w is designed to remove as much visual information from y as possible. This relies on the fact that the dynamic variables y and z are regularized by KL, while the content variable w is not regularized in the loss function. w is instead prevented from containing any temporal information as it is inferred during training with randomly sampled frames and a permutation-invariant network. w is otherwise free to contain any non temporal information. The dynamic variables y and z are encouraged, because of the KL penalty, to contain only necessary information, and therefore should only contain temporal information that could not be captured by w.\\n We clarify this point in Section 3.2 of the revised submission.\\n\\n\\n - Human3.6M:\\n\\n We thank the reviewer for suggesting this relevant dataset. We would have been pleased to show results of our method on this dataset. Unfortunately, it is not publicly available and we were only granted access only a couple of days ago. Please note, however, that the KTH dataset that we consider is somewhat similar to Human3.6M, as both datasets consist in videos of actions performed by subjects in front of a camera.\\n\\n\\n - Related work:\\n\\n We thank the reviewer for pointing out the recent missing reference [1] that we included in the manuscript as related work along with discussion regarding differences with our work. \\n It is orthogonal to our work as they improve the VRNN dynamic model by using frame-wise key-point representations instead of raw frames, while we focus on improving the dynamic model itself.\\n\\n\\nWe hope that we were able to answer your questions.\\n\\n\\n[1] Minderer et al. Unsupervised Learning of Object Structure and Dynamics from Videos. In NeurIPS 2019.\"}", "{\"title\": \"Answer to Reviewer 2\", \"comment\": \"We would like to thank the reviewer for his encouraging feedback and helpful suggestions. We updated the submission accordingly and provide answers and comments below.\\n\\n - Impact and difference with SVG:\\n\\n The difference between our model and SVG is twofold. Firstly, unlike SVG, our model decouples frame synthesis and dynamics, thanks to its state-space nature, a different prediction model and the use of a content variable. Secondly, our model introduces a stochastic residual update that is shown to significantly improve the performance of our model compared to a regular recurrent network.\\n\\n Combined, these improvements allow our model to reach the state of the art of stochastic video prediction, besides obtaining desirable properties for its latent space that could not be achieved by prior models.\\n We think that these contributions could have substantial impact in the community, as the general principles of our model (state-space, residual dynamic, static content variable) can be generally applied to other models as well, even though this is out of the scope of this paper. For instance, replacing the VRNN model in [1] to model the evolution of key-points could bring additional performance gains. Moreover, the state-space nature of our model allows us to learn meaningful dynamic representations, typically used in model-based reinforcement learning [2], and illustrated in Figure 9.\\n\\n We emphasized these points in the revised version of the paper in Sections 2, 4 and 5.\\n\\n\\n - Origin of improvements:\\n\\n Our model shares the same encoder / decoder architecture as SVG, showing that its performance is due to our inference and dynamic systems rather than our choice of neural architecture.\\n\\n To further study the origin of improvements, we presented, in the original submission, two baselines (GRU and MLP). The only difference between those baselines and our model is dynamic-related (they either use recurrent networks such as those of VRNN [3] or recurrent MLPs as an update method, instead of our residual updates). From our experiments, we reach two conclusions. First, all three versions of our method (residual, MLP, GRU) outperform prior methods. Therefore, this improvement is due to their common inference method, latent nature and content variable. Secondly, the residual update provide an additional gain of performance compared to the other methods, that can only be explained by the difference in the update method.\\n\\n This analysis, that was originally performed on both deterministic and stochastic versions of Moving MNIST, have been confirmed by *new experiments on KTH*.\\n\\n We clarified these points and included the new results in the revised version of the submission in Section 4 and Table 2. \\n\\n\\n - FVD metric: \\n\\n We would like to thank the reviewer for suggesting this additional metric. We provide additional FVD scores for all tested methods on KTH and BAIR, as well as an analysis of these results in Table 2 and 3, and Section 4. These results experimentally confirm our previous findings with respect to PSNR, SSIM, and LPIPS, i.e., we outperform the state of the art on KTH and are on par with state-of-the-art methods on BAIR. We believe that FVD is complementary to previously considered metrics, and we include an additional discussion on this matter in Section 4.\\n\\n\\n - PlaNet:\\n\\n We thank the reviewer for pointing out the missing reference [4], that we added in the related work section. The differences with our model are the following. It is a control model that focuses and is evaluated on planning tasks which are out of scope of our paper. Also, [4] uses a dynamic model that is close to the one of [5] (with additional latent overshooting) based on recurrent neural networks, unlike our residual model.\\n\\n\\n - Website:\", \"we_provide_an_anonymized_website_at\": \"https://sites.google.com/view/srvp/. It hosts generated videos with qualitative comparisons, along with the corresponding anonymized code.\\n\\n\\nWe hope that we were able to answer the questions adequately and to alleviate raised concerns.\\n\\n\\n[1] Minderer et al. Unsupervised Learning of Object Structure and Dynamics from Videos. In NeurIPS 2019.\\n[2] Gregor et al. Temporal Difference Variational Auto-Encoder. In ICLR 2019.\\n[3] Chung et al. A Recurrent Latent Variable Model for Sequential Data. In NIPS 2015.\\n[4] Hafner et al. Learning Latent Dynamics for Planning from Pixels. In ICML 2019.\\n[5] Fraccaro et al. Sequential Neural Models with Stochastic Layers. In NIPS 2016.\"}", "{\"title\": \"Answer to Reviewer 1\", \"comment\": \"We would like to thank the reviewer for his comprehensive and useful feedback. We updated the submission accordingly and provide answers and comments below.\\n\\n - Missing references and discussions:\\n\\n We thank the reviewer for pointing out missing references [1] and the recent [2]. We included them in the manuscript as related work along with discussions regarding differences with our work.\\n\\n [1] tackles video prediction in a deterministic setting, while we focus on stochastic prediction instead. As its architecture doesn\\u2019t include stochastic components, it cannot produce a diversity of predictions by design.\\n\\n [2] is orthogonal to our work as they improve the VRNN dynamic model by using frame-wise key-point representations instead of raw frames, while we focus on improving the dynamic model itself. Specifically, we pointed out that re-using predicted frames to perform future frame predictions leads to prediction error accumulation over time and should be avoided. As [2] uses VRNN as a backbone dynamic model, it is autoregressive, albeit in the key-point space instead of the pixel space. While this change could mitigate the above-mentioned problem, the extent of such mitigation is unclear. We believe both approaches to be complementary. It would be an interesting future work to study the behavior of our model in the setting of [2] by replacing their VRNN-based temporal component by ours. This is, however, outside the scope of this paper.\\n\\n\\n - Generation at arbitrary frame rates:\\n\\n We would like to bring to the reviewer\\u2019s attention that our model was not meant to serve as an interpolation method and is not trained as such. Instead, we presented experiments regarding the ability to generate at higher frame rates in order to analyze the dynamics learned by our model. The fact that our model maintains the same performance when generating at a higher framerate, without further training, shows that it learned a continuous dynamic driven by a piecewise ODE (i.e., the learned dynamic of each interval between two consecutive frames is driven by an ODE). We emphasized this point in Sections 3.1 and 4 of the manuscript revision.\\n\\n\\n - Further study of the update rule:\\n\\n In addition to the previous experiments on deterministic and stochastic Moving MNIST, we performed further experiments on the KTH dataset to complement our analysis of the effects of the update rule. Results are presented in Table 2 and further discussed in Section 4 of the revised manuscript. They confirm the structural advantage in terms of performance of residual updates compared to other considered update methods (GRU and MLP). The MLP version better captures dynamics than GRU (in terms of PSNR and SSIM), but drops in terms of realism (captured by LPIPS and FVD). On the other hand, the residual version slightly improves this gain in terms of dynamics but significantly pushes further the realism of the generated videos. This observed advantage comes from the well documented structural prior that the residual update establishes in the network [3, 4].\\n\\n\\n - Interpolation of dynamics:\\n\\n The experiment showing interpolation of dynamics in Figure 9 is designed to assess the interpretability and representation quality of the learned latent space. Such experiments are standard in the literature of generative models [5], but could not have been proposed in the stochastic video prediction community before, as it takes advantage of the state-space nature of our model. We believe that these considerations are another interesting argument in favor of state-space generative models.\\n\\n\\n - FVD metric:\\n\\n We would like to thank the reviewer for suggesting this additional metric. We provide additional FVD scores for all tested methods on KTH and BAIR, as well as an analysis of these results in Table 2 and 3, and Section 4. These results experimentally confirm our previous findings with respect to PSNR, SSIM, and LPIPS, i.e., we outperform the state of the art on KTH and are on par with state-of-the-art methods on BAIR. We believe that FVD is complementary to previously considered metrics, and we include an additional discussion on this matter in Section 4.\\n\\n\\n - We changed the wording from \\u201cresidue\\u201d to \\u201cresidual\\u201d.\\n\\n - The error bars correspond to the 95%-confidence interval of the plotted mean, similarly to [6]. We added this information in the figure captions.\\n\\n\\nWe hope that we were able to alleviate raised concerns.\\n\\n\\n[1] Wichers et al. Hierarchical Long-term Video Prediction without Supervision. In ICML 2018.\\n[2] Minderer et al. Unsupervised Learning of Object Structure and Dynamics from Videos. In NeurIPS 2019.\\n[3] Chen et al. Neural Ordinary Differential Equations. In NeurIPS 2018.\\n[4] He et al. Deep Residual Learning for Image Recognition. In CVPR 2016.\\n[5] Kingma et al. Glow: Generative Flow with Invertible 1x1 Convolutions. In NeurIPS 2018.\"}", "{\"title\": \"Rebuttal and Revision\", \"comment\": [\"We would like to thank the reviewers for their comprehensive, insightful and encouraging comments. We answered individually to each of them. Below, we summarize the main changes made in the revision of the manuscript.\", \"We added FVD scores for all compared methods in Tables 2 and 3, as well as an additional analysis in Section 4. We outperform other methods on this metric on the KTH dataset, and are on a par with the state of the art on BAIR.\", \"We extend the study of our dynamic model to the KTH dataset on Section 4, and reported new results on Table 2. They confirm our findings on the Moving MNIST dataset showing the structural advantage of residual dynamics.\", \"We emphasized the role of the content variable in decoupling dynamic model and visual features in Section 3.2 with an additional explanation.\", \"We further clarified the differences between our method and SVG in Sections 2 and 4.\", \"We added references [1, 2, 3] to the related work with additional discussions.\"], \"we_would_like_to_highlight_the_anonymized_project_webpage_containing_the_code_corresponding_to_our_method_as_well_as_video_samples\": \"https://sites.google.com/view/srvp/.\\n\\n[1] Minderer et al. Unsupervised Learning of Object Structure and Dynamics from Videos. In NeurIPS 2019.\\n[2] Wichers et al. Hierarchical Long-term Video Prediction without Supervision. In ICML 2018.\\n[3] Hafner et al. Learning Latent Dynamics for Planning from Pixels. In ICML 2019.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary:\\n\\nThe paper proposes a video prediction method based on State-Space Models. The paper describes two main contributions:\\n\\n1. By learning dynamics in the latent state space, the method avoids the high computational cost and accumulating image reconstruction errors of autoregressive models that condition on generated frames.\\n\\n2. To model dynamics, the paper proposes a residual update rule inspired by Euler\\u2019s method to solve ODEs. According to this rule, the update to the latent state y_t is modeled as an additive residual f(y_t, z_{t+1}). This has the advantage that the step size of the discretization can be adjusted freely, e.g. between training and inference. \\n\\nThe paper provides extensive experimental comparison of their model to the SVG and SAVP models on several standard datasets. The paper further contains experiments illustrating features of the model such as disentangling dynamics and content, and interpolation of dynamics in the latent space.\", \"decision\": \"The paper is written clearly and the mathematical treatment and experiments appear rigorous. The idea of predicting video using state-space models is interesting and promising. However, as described below, the paper overstates its novelty and falls short of showing the advantages of the method beyond incremental improvements on frame-wise image quality metrics. I therefore suggest rejection in its current version.\", \"supporting_arguments_and_suggestions\": \"1. The idea to use fully latent models for video prediction, to untie frame synthesis and dynamics, is not new and the paper does not fully cite this literature. For example, [1] and [2] perform unsupervised, non-autoregressive video prediction. The differences to these models should be discussed.\\n\\n2. The advantages of the residual update rule are not made clear enough. The parallels to the ODE literature seem tenuous. The main advantage described in the paper is the ability to synthesize videos at different frame rates, but interpolation over such short time horizons is not a hard problem. At least, the paper should compare to existing methods for frame interpolation. Apart from interpolation (variable step size), it appears that the update rule could be changed from y_{t+1} = y_t + f(y_t, z_{t+1}) to y_{t+1} = f(y_t, z_{t+1}) without impact to the model. How is it different from the standard VRNN formulation [3]? More experiments to show the advantage of the proposed update rule would be helpful.\\n\\n3. Some of the experiments seem like interesting starting points but do not support general claims. For example, Fig 2 (b) shows that the proposed dynamics model is better than an MLP or GRU on deterministic Moving MNIST, but is this also true on real datasets, which have much more complex dynamics? Similarly, the interpolation in Figure 9 is intriguing, but it would be helpful to describe and test how this ability is useful for applications of the predictive model.\\n\\n4. The comparisons use frame-wise metrics of image quality (PSNR, SSIM, LPIPS). Even though they are common in the literature, these metrics are unsuitable for comparing long video sequences due to their stochasticity. The metrics are probably dominated by relatively uninteresting features such as the quality of the static background. Metrics for comparing entire videos exist (e.g. FVD [4]) and should be used. Even better, the paper should demonstrate the usefulness of the model for downstream tasks such as reinforcement learning, although I understand that this may be out of scope.\", \"minor_comments\": \"- As far as I know, the correct term for error terms is residual, not residue.\\n- What do the error bars in the figures show? Please add this information to the figure legends.\\n\\n[1] Wichers et al., 2018, https://arxiv.org/pdf/1806.04768.pdf\\n[2] Minderer et al., 2019, https://arxiv.org/abs/1906.07889\\n[3] Chung et al, 2015, https://arxiv.org/abs/1506.02216\\n[4] Unterthiner et al., 2018, https://arxiv.org/abs/1812.01717\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper proposes a video prediction model which explicitly decouples frame synthesis and motion dynamics. This is a very subtle change (compared to the current models) that can result in higher quality predictions.\\n\\nFirst of all, the paper is extremely well written. It provides clear motivations and goals, as well as an impressively comprehensive related work that discusses their shortcomings. The experiments are comprehensive and provide good support for the claims. And finally, the appendix presents additional visualization and information. \\n\\nOn the main proposed method, it is a very subtle but reasonable change. Therefore, my suggestion to the authors is to provide a more thorough comparison with existing methods specifically SVG (Denton 2018) since the models share a lot of similarities. It is also quite similar to PlaNet (Hafner 2019). This is where the paper can be improved.\\n\\nFor the experiments, although they are quite comprehensive, there is still room for improvement. First, none of the metrics used are good evaluation metrics for frame prediction (I know they are quite common but that doesn't make them good) as they do not give us an objective evaluation in the sense of the semantic quality of predicted frames, specially for long videos. It really helps if the authors present additional quantitative evaluation to show that the predicted frames contain useful semantic information with metrics such as FVD and Inception score. Second, a pretribulation study is required to see where the improvements are coming from. Is it from a different architecture or the separation of dynamics? Finally, a website with generated videos really helps for qualitative comparison! \\n\\nOverall, this is a well-written paper with clear motivations and goals. I find the impact of the paper to be marginal (given the quality difference with already existing models) which can be improved by emphasizing more on other aspects such as disentanglement.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Contributions: this submission proposes a video pixel generation framework with the goal to decouple visual appearance and dynamics. The latent dynamics are modeled with a latent residual dynamics model. Empirical evaluations on moving MNIST show that the proposed residual dynamics model outperform MLP or GRU. On more challenging KTH and BAIR datasets, the proposed method achieves on par or better quantitative performance with previous methods, and have nice qualitative results on content \\\"swap\\\" and dynamics interpolation.\", \"assessment\": [\"To my knowledge, the proposed model is novel for video generation.\", \"The proposed method is supported with strong quantitative results and qualitative analysis, ablation on Moving MNIST shows that the proposed latent residual dynamics model outperforms MLP and GRU baselines.\", \"The authors might be interested in related work on video generation with decoupled appearance and dynamics models, such as [1]. It would also be interesting to see evaluation on more challenging datasets, such as Human3.6M.\", \"Question: how does the proposed inference framework make sure to decouple appearance with dynamics? Can y_i not encode the appearance information?\", \"[1] Minderer et al., Unsupervised Learning of Object Structure and Dynamics from Videos. NeurIPS 2019.\", \"-----------------------------\"], \"post_rebuttal\": \"Thank you for your answers to my questions and the updated manuscript. My questions have been addressed and the additional results further confirm the performance of the proposed method. Therefore I recommend weak accept of the submission.\"}" ] }

[ "Antonia Creswell", "Luis Piloto", "David Barrett", "Kyriacos Nikiforou", "David Raposo", "Marta Garnelo", "Peter Battaglia", "Murray Shanahan" ]

[ "Graph networks", "alignment", "objects", "relation networks" ]

{ "note_id": [ "gxAihoNTWM", "r1x9jiPKiS", "SJx1DjvYjr", "BygXLcPKiB", "B1x_gbizcr", "ByxlfiUntH", "B1gLTsN2FB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732235, 1573645218362, 1573645142936, 1573644875396, 1572151536158, 1571740423729, 1571732413559 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1776/Authors" ], [ "ICLR.cc/2020/Conference/Paper1776/Authors" ], [ "ICLR.cc/2020/Conference/Paper1776/Authors" ], [ "ICLR.cc/2020/Conference/Paper1776/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1776/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1776/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes a network architecture which labels object with an identifier that it is trained to retain across subsequent instances of that same object.\\n\\nAfter discussion, the reviewers agree that the approach is interesting, well-motivated and written, and novel. However, there was unanimous concern about the experimental evaluation, so the paper does not appear to be ready for publication just yet, and I am recommending rejection.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to reviewer 3:\", \"comment\": \"The authors would like to thank reviewer three for their feedback and for appreciating our contributions. We agree that experiments on real-world datasets would be interesting, however the motivation for using a symbolic dataset was that we were better able to focus on the alignment problem, rather than the problem of extracting and representing entities.\\n\\nYour suggestion of cross-modal and cross-lingual is really interesting and definitely something we will be considering for future work, thank you.\"}", "{\"title\": \"Response to reviewer 2:\", \"comment\": \"The authors are grateful for reviewer two\\u2019s interest in our work and for their thoughtful feedback.\", \"we_have_addressed_the_comments_of_reviewer_two_in_detail_below\": \"\", \"r2\": \"The appearance of the same object can change due to lighting and distance, and it is unreasonable to assume that their features would remain static (apart from simple uniform noises).\", \"response\": \"We appreciate this comment, however, we have chosen to decouple the problem of representing objects from the problem of aligning objects over time. There are already many excellent models for representing objects; for example, MONet and Iodine, but the problem of aligning objects is still an open one.\"}", "{\"title\": \"Response to reviewer 1:\", \"comment\": \"The authors would like to thank reviewer one for taking the time to read and understand our paper. However, we realise that we did not sufficiently emphasise the extent of our contribution and would like to take this opportunity to persuade the reviewer that our contribution is both original and substantial. In our paper we provide a novel framework in which a neural network is able to learn a similarity measure. The AlignNet defines a formal setting, which the reviewer refers to as \\u201cjust a set of rules\\u201d, in which a network may be used to learn an implicit similarity measure. Indeed, we are the first to formulate the alignment problem in such a way that a neural network may be trained to solve the problem. Thank you for raising this concern, to make our contribution more clear we have added the following to our list of contributions in the revised paper:\\n\\n\\u201cWe are the first to formulate the alignment problem such that a neural network may learn a similarity measure for aligning entities. We propose the self-supervised AlignNet, a model capable of finding entities in one set that correspond with entities in another set, while also being able to deal with the adding and dropping of entities.\\u201d\\n\\nObtaining this formulation was non-trivial and required many iterations.\", \"we_will_now_address_the_reviewers_comments_in_more_detail\": \"\", \"r1\": \"Comparisons are made only with hand-crafted features. It is quite evident that a learned similarity measure will be better than hand-crafted distances.\", \"response\": \"We would like to reiterate that the AlignNet provides a novel framework in which a neural network is able to learn a similarity measure. Before this work, it was not clear how to construct a network and loss by which a similarity measure may be learned. It is for this reason that we compare to hand-crafted baselines.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper presents a model that is able to associate objects seen in a new time step with the objects seen in previous frames and therefore consider the cases of adding new objects, dropping object no longer visible, but keeping them in memory in case they reappear in the future. The model assumes that the location of the objects is given and the only task to perform is the correct association of instances. Results shows that the proposed approach is better than hand-crafted features on different simulated tasks. Additionally the propose model is shown to help to reduce the computational cost for question answering task.\\n\\nI lean to reject this paper because in my opinion the proposed method is just a set of predefined rules to train a network to be able to perform object association between a new time step and a memory. Additionally the experimental evaluation is very weak in several points (see below). \\n\\n- Contribution: this approach proposes a set of rules to train a network to be able to learn the correct association between two set of features. Additionally, the network is used together with a memory to keep track of previously seen objects. In my understanding the proposed work presents a set of hard-coded rules to train a network for tracking. However, the connection with tracking is presented only in related works (at the end of the paper) and no comparison with other tracking approach is presented. Additionally, the network model that is actually used for the experiments (graph network) is not presented in the main paper, which makes more difficult to understand how the network works for the given task.\\n\\n- Experimental evaluation: Experiments are performed on a very simple, simulated environment. Objects are very simple to recognize and their location (which is often a difficult component of the problem) is given. As the task is quite straight forward, the model obtains almost 100% on all tests. Comparisons are made only with hand-crafted features. It is quite evident that a learned similarity measure will be better than hand-crafted distances. \\n\\nAuthors should explain more clearly that the proposed contribution is a set of rules used to learn a distance between features in order to associate object instances. \\nA cleared connection with tracking should be provided from the introduction. In the evaluation there should be a comparison on multiple and more challenging tracking datasets. In this way we can compare the proposed training technique with other tracking approaches that also learn to a distance in order to associate object at different time steps.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"With the assumption of object persistence and inspired by the sticky indices, this paper proposed a novel object alignment method for matching arbitrary entities in different sets. The motivation is clear and easy to follow. The simulation experiment with the symbolic dataset gives impressive results.\\n\\nI like this idea and have one concern about the experiment. To be specific, all the reported results are obtained in the symbolic dataset to simulate the real-world case. Have you performed the entity alignment in the real-world data, such as the cross-lingual knowledge graphs, or cross-modal analysis? It's expected to experiment with more challenging datasets. By the way, have you released your code? This is important to an ICLR submission.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes AlignNet, a bipartite graph network that learns to match to sets of objects. AlignNet has a slot-wise object-based memory that associates an index with each unique object and can discover new and re-appearing objects. Experiments are conducted on a symbolic dataset.\\n\\nI do not think the paper meets the acceptance threshold, and recommend for weak rejection. While the paper proposes an interesting architecture to address the alignment problem, it has noticeable flaws in its experimental designs.\\n\\nFirst, all the experiments are conducted on toy symbolic datasets, where the alignment problem is rather easy to solve. On the other hand, real-world scenarios can be far more complicated. For example, the appearance of the same object can change due to lighting and distance, and it is unreasonable to assume that their features would remain static (apart from simple uniform noises). In addition, the paper only compares against hand-crafted similarity measures (MSE and cosine). It is unfair to compare learned methods only to hand-crafted methods. As a reasonable and fair comparison, the paper should also compare AlignNet against learned similarity measures (such as a neural network supervised with ground-truth labels for alignments).\\n\\nThe toy dataset and simple baselines in this paper raise doubts on whether the proposed method is applicable to more complex scenarios (such as aligning two sets of objects in natural images through their appearance features).\"}" ] }

[ "Masahiro Kato", "Yoshihiro Fukuhara", "Hirokatsu Kataoka", "Shigeo Morishima" ]

[ "Learning with Rejection", "Adversarial Examples" ]

{ "note_id": [ "wudeaXlvBi", "rJelEpUniH", "BkeLyHI3oH", "rkgA_QI3sr", "S1xcdBLAtB", "Sye1YXKatH", "SylCPVKnFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732204, 1573838119996, 1573835997766, 1573835638041, 1571870066042, 1571816310860, 1571751013588 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1775/Authors" ], [ "ICLR.cc/2020/Conference/Paper1775/Authors" ], [ "ICLR.cc/2020/Conference/Paper1775/Authors" ], [ "ICLR.cc/2020/Conference/Paper1775/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1775/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1775/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper addresses the setting of learning with rejection while incorporating the ideas from learning with adversarial examples to tackle adversarial attacks. While the reviewers acknowledged the importance to study learning with rejection in this setting, they raised several concerns: (1) lack of technical contribution -- see R1\\u2019s and R2\\u2019s related references, see R3\\u2019s suggestion on designing c(x); (2) insufficient empirical evidence -- see R3\\u2019s comment about the sensitivity experiment on the strength of the attack, see R1\\u2019s suggestion to compare with a baseline that learns the rejection function such as SelectiveNet; (3) clarity of presentation -- see R2\\u2019s suggestions how to improve clarity.\\nAmong these, (3) did not have a substantial impact on the decision, but would be helpful to address in a subsequent revision. However, (1) and (2) make it very difficult to assess the benefits of the proposed approach, and were viewed by AC as critical issues.\\nAC can confirm that all three reviewers have read the author responses and have revised the final ratings. AC suggests, in its current state the manuscript is not ready for a publication. We hope the reviews are useful for improving and revising the paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to AnonReviewer2\", \"comment\": \"Dear Reviewer,\\n\\nThank you for your comments. Our replies to the comments are listed below.\\n\\nQ1. The definition of \\u201csuspicious example\\u201d in Sec.3.1 has no relationship with adversary examples. Does the paper focus on adversary examples? If the definition has no relationship, it is classical learning with rejection. \\nA1. This definition is the standard definition of adversarial examples and adversarial attacks. For example, [1] introduced the definition as the traditional definition of adversarial examples. We can derive the definition from the formulation of adversarial training, min_f max_{x'\\\\in B(x)} \\\\ell(f(x), y) [2]. The definition of suspicious samples is based on the traditional definition of adversarial examples, and we extended the traditional definition to introduce the proposed algorithm. \\n[1] Theoretical Framework for Robustness of (Deep) Classifiers, ICLR workshop 2017\\n[2] Rademacher Complexity for Adversarially Robust Generalization, ICML 2019\\n\\nQ2. In the last equation of Page 3, there is no definition of \\\\tilde L. Actually, according to Figure 1, x\\u2019s is more close to the decision boundary, it is an example more hard to classify, which could also be \\u201csuspicious\\u201d. \\nA2. In page 3, to introduce the idea of the proposed algorithm, we did not explain the definition of \\\\tilde L, and we explained it in Eq.(8). However, as you pointed out, it is confusion and we will revise it in the next division.\\n \\nQ3. In the definition of \\u201csuspicious example\\u201d at the beginning of Sec.3.1, is both x and x\\u2019 defined as suspicious examples in this way?\\nA3. Yes, both x and x\\u2019 can be suspicious examples. However, as we explained in section 4, we can select a specific class that should be protected. If we are only afraid of the misclassification of a class of x, we define x only in learning with protection.\\n\\nQ4. In the last equation of page 2, there is a rejection function, so minimizing this loss is a \\u201cseparation-based approach\\u201d. However, at the end of Sec.2 the paper states they \\u201cfollow a confidence-based approach\\u201d. Any comment on the inconsistency?\\nA4. It is consistent. The last equation of page2 is not only for \\u201cseparation-based approach\\u201d, but also \\\"confidence-based approach\\u201d.\\n\\nQ5. The motivated problem is not new. It is called cost-sensitive learning in machine learning and can date back to 2001.\\nA5. We think that the motivation for cost-sensitive learning is different from that of our research. In cost-sensitive learning, they do not care about the existence of adversarial examples. \\n\\nQ6. \\u201can estimator \\\\hat h might return result that differ greatly from h^* in a case with finite samples\\u201d. Actually there are rigorous theoretical results describing how the number of finite samples will impact the estimator \\\\hat h on unseen data. So inaccurate/unclear statements that will mislead readers should be avoided. \\nA6. To motivate our research, we described that the adversarial examples are caused by estimation error. However, as you pointed out, the sentence will confuse readers. We change the expression in the next revision.\\n\\nQ7. In the last equation of Page 3, there is no definition of \\\\tilde L. In writing, the paper also lacks the necessary references in many places. The organization is also problematic. For example, in the second half of Sec.2 introducing two kinds of learning with rejection models, it should be included in a \\u201crelated work\\u201d part. \\nA7. Thank you for pointing out them. We will revise our paper by reflecting your opinions.\"}", "{\"title\": \"Response to AnonReviewer3\", \"comment\": \"Dear Reviewer,\\n\\nThank you for your inquiry about the proposed method and the experimental results. Our replies to the questions are listed below.\\n\\nQ1. My understanding of classification with a reject option is that the rejection cost c(x) is a design choice that can depend on the specific application. \\nA1. In this paper, we proposed a new framework of applying learning with rejection to adversarial examples. As far as we know, there is no method based on a similar idea. As you pointed out, we also think that we should design our algorithm to enable us to control c(x). However, there are other difficulties in the theoretical justification for it. Therefore, based on this paper, we tackle the problem of designing c(x) in the next research.\\n\\nQ2. The rejection function r(x) relies on z^* which essentially amounts to computing an adversarial perturbation for a given testing point, see (2). The authors state that they use a 30-step of the PGD algorithm to find z^*. The attacks on which the method is tested only uses 10 PGD steps. What if the attacker is stronger in the sense that it runs significantly more PGD iterations than used to compute the rejection function? How sensitive is the rejection function to different initializations?\\nA2. We also think that we need to increase the number of iterations of PGD when an attacker uses a stronger PGD attack. Following your opinion, in the next revision, we will show the sensitivity by experiment.\\n\\nQ3. Similarly, what happens if the attacker and the rejection function rely on different norms to compute the attack and the rejection score, respectively? I think it is important to investigate this aspect.\\nA3. We have shown the results in Table 2. In the experiments, we used l_inf norm to reject suspicious samples and l_2 norm to attack the model. We confirmed that the proposed method is effective against an unexpected attack.\\n\\nQ4. In Tables 1 and 2, I don\\u2019t understand why the precision is the same for all rows corresponding to the same attack (strength), while the other metrics vary.\\nA4. The proposed algorithm rejects suspicious samples. Therefore, it does not change the numbers of samples that are predicted as positive. It means that the numbers of true positive and false positive do not change. Because precision = TP/(TP+FP), precision also does not change. Even though precision does not change, we showed the result to evaluate the algorithm.\"}", "{\"title\": \"Response to AnonReviewer1\", \"comment\": \"Dear Reviewer,\\n\\nThank you for your insightful comments. Our replies to the comments are listed below.\\n\\nQ1. The proposed approach is not novel. For example, [1] uses adversarial example style detection to augment their training data and improve their end-to-end model. \\nA1. The idea of our paper is different from that of Wang et al. [1]'s paper. They augmented their samples using adversarial examples. On the other hand, we reject some samples that will mislead the learner's decision making. The idea of Wang et al. [1]'s paper is similar to adversarial training. However, it is well known that adversarial training can be a cat\\u2010and\\u2010mouse game, i.e., a model trained by adversarial training is still weak against a different attack that is not assumed in the training process. Because training a robust model is a difficult task as many existing works pointed out, our idea is to propose a new way of decision making to deal with adversarial examples by introducing a rejection option. \\n\\nQ2. There have been approaches which attempt to learn rejection function [2], so it would have been good to at least do a comparison of the proposed approach with such methods.\\nA2. We agree with your opinion. In fact, in the next research, we should consider learning a rejection function and classifier together. However, because the paper [2] does not consider the existence of adversarial examples, we consider that the method proposed in [2] is weak against adversarial examples. Therefore, we did not compare our method with their method that considers standard learning with rejection.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": [\"There is still no universal method to deal with adversarial examples, and introducing a reject option to flag potential attacks seems a sensitive choice for many applications. The considered problem is well-motivated and introduced, and I\\u2019m unaware of prior work studying classification with reject option in the context of adversarial examples. However, I think there are different dimensions along which the paper could be improved:\", \"My understanding of classification with a reject option is that the rejection cost c(x) is a design choice that can depend on the specific application. While c(x) is introduced as part of the framework it is then derived in a very specific way, removing the design aspect or at least not explaining very clearly how one would design it. Also, Algorithm 1 doesn\\u2019t have corresponding parameters.\", \"The rejection function r(x) relies on z^* which essentially amounts to computing an adversarial perturbation for a given testing point, see (2). The authors state that they use a 30-step of the PGD algorithm to find z^*. The attacks on which the method is tested only uses 10 PGD steps. What if the attacker is stronger in the sense that it runs significantly more PGD iterations than used to compute the rejection function? How sensitive is the rejection function to different initializations?\", \"Similarly, what happens if the attacker and the rejection function rely on different norms to compute the attack and the rejection score, respectively? I think it is important to investigate this aspect.\", \"In Tables 1 and 2, I don\\u2019t understand why the precision is the same for all rows corresponding to the same attack (strength), while the other metrics vary.\", \"Overall, I think the paper explores an interesting direction, but would greatly benefit from a revision along the lines outlined above.\"], \"minor_comments\": [\"P3 3.1 2nd paragraph: \\u201cLet B^p...\\u201d rather than \\u201cLet B^\\\\infty...\\u201d\", \"P4 bottom: \\u201c)\\u201d is missing before \\u201c..., where\\u201d\", \"P6 bottom: (FR) should be (TR). The acronyms FA and FR don\\u2019t seem to be introduced.\", \"###\"], \"reply_to_rebuttal\": \"I thank the reviewers for their detailed reply. It seems that the authors agree that some of the issues I raised should be addressed and improved. My assessment that the paper should be revised (and accordingly the rating) therefore remains unchanged.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposes a framework for learning with rejection using ideas from adversarial examples. The essential idea is, while predicting on a point x, we can reject classifying the point if it has an adversarial example very close to it. So, the algorithm can be simply summarized as,\\n1. Learn a classifier function f\\n2. On the test set, predict on a point, only if it doesn't have an adversarial example close by.\", \"i_am_inclined_to_reject_the_paper_for_the_following_reasons\": \"1. The proposed approach is a variation of a fairly well-known heuristic. Having a close adversarial example is same as saying that the current point is very close to the decision boundary. Being close to the decision boundary is a heuristic that has been applied in multiple scenarios in machine learning.\\n2. The proposed approach is not novel. For example, [1] uses adversarial example style detection to augment their training data and improve their end-to-end model. \\n3. There have been approaches which attempt to learn rejection function [2], so it would have been good to at least do a comparison of the proposed approach with such methods.\\n\\n[1] Adversarial Examples For Improving End-to-End Attention-based Small-Footprint Keyword Spotting, ICASSP 2019\\n[2] SelectiveNet: A Deep Neural Network with an Integrated Reject Option, ICML 2019\\n\\n---\\n\\nThanks for the rebuttal. I have raised my scores, but I still believe that this paper falls short of acceptance.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper wants to study the problem of \\u201clearning with rejection under adversarial attacks\\u201d. It first naively extends the learning with rejection framework for handling adversarial examples. It then considers the classical cost-sensitive learning by transfer the multi-class problem into binary classification problem through one-vs-all and using the technique they proposed to reject predictions on non-important labels, and name such technique as \\u201clearning with protection\\u201d. Finally, they do some experimental studies.\\n\\nThe paper does not show any connection between \\u201clearning with rejection\\u201d and \\u201cadversarial learning\\u201d. The method it proposes is also a na\\u00efve extension of existing methods. Both the problem setting and the technique does not have novelty. The paper fails to realize that the motivated application is actually called \\u201ccost-sensitive learning\\u201d and has been studied long time before. The paper also has problems in writing. Finally, there is no comparison with any baseline. Only empirical results of the proposed methods are shown. Due to all these reasons, there is still a long way to go before the paper can be published. I will rate it a clear rejection.\\n\\nMore specially, \\n\\nThe definition of \\u201csuspicious example\\u201d in Sec.3.1 has no relationship with adversary examples. Does the paper focus on adversary examples? If the definition has no relationship, it is classical learning with rejection. \\nIn the last equation of Page 3, there is no definition of \\\\tilde L. Actually, according to Figure 1, x\\u2019s is more close to the decision boundary, it is an example more hard to classify, which could also be \\u201csuspicious\\u201d. \\nIn the definition of \\u201csuspicious example\\u201d at the beginning of Sec.3.1, is both x and x\\u2019 defined as suspicious examples in this way?\\nIn the last equation of page 2, there is a rejection function, so minimizing this loss is a \\u201cseparation-based approach\\u201d. However, at the end of Sec.2 the paper states they \\u201cfollow a confidence-based approach\\u201d. Any comment on the inconsistency?\\n\\nThe motivated problem is not new. It is called cost-sensitive learning in machine learning and can date back to 2001:\\nCharles Elkan. The Foundations of Cost-Sensitive Learning. IJCAI 2001: 973-978.\\nWhere they study the same problem when misclassifying one class of data may cost a lot than misclassifying another class of data. The current paper has not discussed any related work of cost-sensitive learning although they want to study a problem in its field. \\n\\nThe paper should be also improved in writing in the following aspects. \\nThere is a lot of inaccurate statements in the paper. For example, \\u201cIn Sections 3 and 4, we propose and describe our algorithm\\u201d, what is the difference between propose and describe? \\u201can estimator \\\\hat h might return result that differ greatly from h^* in a case with finite samples\\u201d. Actually there are rigorous theoretical results describing how the number of finite samples will impact the estimator \\\\hat h on unseen data. For example, \\nPeter L. Bartlett, Shahar Mendelson. Rademacher and Gaussian Complexities: Risk Bounds and Structural Results. JMLR, 2002.\\nSo inaccurate/unclear statements that will mislead readers should be avoided. \\n\\nIn writing, the paper also lacks the necessary references in many places. For example, \\u201cLearning with rejection is a classification scenario where the learner is given the option to reject an instance instead of predicting its label.\\u201d, \\u201c\\u2026classifies adversary attacks to two types of attacks, white-box attack and black-box attack.\\u201d, \\u201cMethods for protecting against these adversarial examples are also being proposed.\\u201d. Necessary references are needed for these places.\\n\\nThe organization is also problematic. For example, in the second half of Sec.2 introducing two kinds of learning with rejection models, it should be included in a \\u201crelated work\\u201d part. \\n\\n------------------------------------------\\nThank you for the rebuttal. I raised my score a little bit. But I still think this paper has not been ready to be published yet.\"}" ] }

[ "Riley Simmons-Edler", "Ben Eisner", "Daniel Yang", "Anthony Bisulco", "Eric Mitchell", "Sebastian Seung", "Daniel Lee" ]

[ "Deep Reinforcement Learning", "Exploration" ]

{ "note_id": [ "YYopxreh3o", "HyrY1o2oH", "HylANqQFjr", "Byg2jD7FjS", "BJeRhNmYjH", "rJetimXYiH", "rkl5dt-ycH", "H1xbZ_dpYr", "SygTs5emKH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732175, 1573855100517, 1573628470393, 1573627812117, 1573627062158, 1573626784785, 1571916145527, 1571813368975, 1571125924786 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1774/Authors" ], [ "ICLR.cc/2020/Conference/Paper1774/Authors" ], [ "ICLR.cc/2020/Conference/Paper1774/Authors" ], [ "ICLR.cc/2020/Conference/Paper1774/Authors" ], [ "ICLR.cc/2020/Conference/Paper1774/Authors" ], [ "ICLR.cc/2020/Conference/Paper1774/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1774/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1774/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"There is insufficient support to recommend accepting this paper. Although the authors provided detailed responses, none of the reviewers changed their recommendation from reject. One of the main criticisms, even after revision, concerned the quality of the experimental evaluation. The reviewers criticized the lack of important baselines, and remained unsure about adequate hyperparameter tuning in the revision. The technical exposition lacked a sober discussion of limitations. The paper would be greatly strengthened by the addition of a theoretical justification of the proposed approach. In the end, the submitted reviews should be able to help the authors strengthen this paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Additional Paper Revisions\", \"comment\": [\"We thank the reviewers again for their thoughtful comments. We have further revised our manuscript in the last several days to include significant additional experiments requested by reviewers:\", \"We have added two new method comparisons, to DORA and GEP-PG on SparseHalfCheetah, showing that QXplore performs favorably to both.\", \"We have performed more hyperparameter sweeps on our RND implementation, showing that our original parametrization is comparable or better than other configurations tested.\", \"We added some additional comparison to RND for the 0-to-1 reward function as well as QXplore. RND performs slightly better than on the -1-to-0 reward function and is on-par with QXplore for that variant, rather than slightly worse.\", \"We revised and centralized our related work at the request of reviewer 2.\", \"We have clarified the particularities of adapting some related methods, such as DORA and GEP-PG, to our experimental settings.\", \"We hope that these changes and additional experiments have addressed the concerns raised by our reviewers, and that the significance of our contribution to exploration has been more clearly demonstrated.\"]}", "{\"title\": \"Responses to Points Raised\", \"comment\": \"Thank you for your helpful comments on organization and clarity, and for taking the time to collect typos. We have responded to several of the points raised below:\\n\\n- \\u201cinspiration from computational neuroscience models...should be removed from the paper and kept for another paper\\u201d\\n\\nEchoing our response to Reviewer 1, this is a reasonable criticism. We leave this for future work, and have revised our manuscript accordingly.\\n\\n- \\u201csome related work references are dispersed\\u2026[and] the paper is poorly organized\\u201d\\n\\nWe have condensed some of these sections, and reworked some of the confusing organization in the updated manuscript. We will further aggregate related work into a discrete section in a future draft.\\n\\n- \\u201c...whether QXplore can deal with problems taking both some positive and negative rewards\\u201d\\n\\nInformally, we have tested QXplore on tasks with both positive and negative rewards, and found performance to be unaffected, as TD-error is relative to the current predicted Q-value rather than the absolute reward scale.\\n\\n- \\u201cthe authors mention RND and DORA as baselines, but only compare to RND. What about DORA?\\u201d\\n\\nWe decided that DORA would not be a useful comparison, because it has not been demonstrated to be effective on large MDPs, and isn\\u2019t directly applicable to continuous action-space problems. We opted to focus on RND because of its algorithmic relationship to QXplore and because it has been shown to work on large MDPs.\\n\\n- \\u201cthe paper would be much stronger if the comparison was with respect to many other methods...\\u201d\\n\\nThis is an understandable position. However, it is complicated by the fact that in its current form QXplore is formulated to work only with an off-policy RL algorithm, and nearly all other existing exploration methods (including RND, VIME, EMI, ICM, etc.) are implemented on top of policy gradient methods such as PPO and TRPO. We attempted to adapt several of these methods for use in off-policy learning settings, but found we couldn\\u2019t get good performance on a Q-learning baseline using author-provided hyperparameters for any method except for RND. GEP-PG is a reasonable method to compare with, and we will attempt to make a fair comparison by the end of the rebuttal period.\\n\\n- \\u201cis the difference in performance due to using TRPO which is known to be less sample efficient?\\u201d\\n\\nWe recognize that some of our performance gains reported in Table 1 may be attributed to this difference. However, the maximum performance achieved by QXplore is significantly higher than the maximum performance achieved by other methods in any amount of time as reported by the original authors.\\n\\n- \\u201cthe caption of Table 1 should be explicit about which algorithm comes from which paper\\u201d\\n\\nGood suggestion. We\\u2019ve updated the manuscript accordingly.\\n\\n- \\u201c...their algorithm is very sensitive to initialization.\\u201d\\n\\nThe alternate initialization schemes tested were intended to test the performance of QXplore on suboptimal initializations, not to provide equivalent initialization schemes to Kaiming-Uniform. An informal, empirical spot check reveals a relationship between the performance of agents using a given initialization method and the mean values output by the initialized network. Concretely, when 128 experiences are sampled randomly in the SparseHalfCheetah environment, the standard Kaiming-Uniform initialization produces mean Q-values in the [-0.15, 0.15] range, Xavier-Uniform in the [-1.5, 1.5] range, Kaiming-Normal in the [-5, 5] range, Uniform in the [-200, 200] range, and Normal in the [-2000, 2000] range. These increases in initial Q-value range correlate with the decreases in performance when we train QXplore using these initializations (see new Figure 11 in Appendix G). Furthermore, a relationship between training speed/final performance and feature/output magnitudes has been repeatedly shown in the literature (notably in Batch Norm [1]). Our initialization experiments were intended to demonstrate robustness to suboptimal initializations, and show that Qx is still able to find reward given even very bad initializations.\\n\\n- \\u201cSparseHalfCheetah is not that sparse... It would be more informative to use 0 everywhere and 1 when successful\\u201d\\n\\nTo allay any concerns about the change in reward function, we ran QXplore on the 0 to 1 reward function and include results in a revision to our paper. Briefly, without modification QXplore trains slower on this reward function, but adding a hyperparameter to initialize the bias of the output neuron of Q to a nonzero value can correct this slowdown. Among a simple parameter sweep with values of 1, 10, and 100, we found that an initial bias of 10 recovers the full performance we reported with the -1 to 0 reward function. Full details can be found in Appendix E.\\n\\n- \\u201ctypos\\u201d\\n\\nGood catch - thanks. We\\u2019ve fixed them in the updated manuscript.\\n\\n[1] Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167.\"}", "{\"title\": \"Responses to Points Raised\", \"comment\": \"Thank you for your helpful and supportive comments. We are pleased you found our method interesting and intuitive. We have addressed several points below, and hope that we have addressed your concerns sufficiently:\\n\\n- \\u201c[SparseHalfCheetah\\u2019s] reward function is not sparse since it gives -1 at every step except at the goal\\u201d\\n\\nOur use of the term \\u201csparse\\u201d relates to the fact that the reward function is a step function (-1 until the cheetah passes a threshold, 0 afterwards), where the 0-valued region is unlikely to be sampled by a naive exploration policy. Since there are two discontinuous, uniform regions in the reward space, the boundary between which is distant from the starting point, we still consider this to be a \\u201csparse\\u201d reward function. Naive exploration algorithms such as epsilon greedy are not capable of exploring it (see Figure 3).\\n\\n- \\u201cThis [default -1] change to the reward function makes any comparisons to previous work questionable\\u201d\\n\\nTo allay any concerns about the change in reward function, we ran QXplore on the 0 to 1 reward function and include results in a revision to our paper. Briefly, without modification QXplore trains slower on this reward function, but adding a hyperparameter to initialize the bias of the output neuron of Q to a nonzero value can correct this slowdown. In a small hyperparameter sweep we found that an initial bias of 10 recovers the full performance we reported with the -1 to 0 reward function. Full details can be found in Appendix E.\\n\\n- \\u201cIt is unclear to me then which benefits QXplore could have compared to previous algorithms such as DORA.\\u201d\\n\\nQXplore will not outperform previous state novelty methods such as RND or DORA before any reward is found. However, once reward is encountered, Qx will focus exploration on states that lead to unexpected reward (or an unexpected absence of reward), and as such this focused exploration results in faster learning and convergence by Q.\\n\\n- \\u201cSome discussion of possible pitfalls for the algorithm could be added\\u201d\\n\\nYou are correct in your assessment that stochastic rewards create local (but not necessarily global) maxima in the unsigned TD-error landscape, which can cause QXplore to explore insufficiently. We observe that one simple solution is to use signed TD-error instead of unsigned, which avoids this issue at the cost of worse performance on SparseHalfCheetah (based on new results in Appendix E).\\n\\n- \\u201cCould the authors clarify how TD-error-based exploration avoids the exploitation-versus-exploration tradeoff?\\u201d\\n\\nThe intent of this passage was to highlight the fact that driving exploration using TD-error will inherently sample more states that are relevant to exploitation, whereas state-novelty exploration methods are reward agnostic and tend not to sample as many exploitation-relevant states in sparse environments. We have revised the manuscript to make this distinction clearer.\\n\\nAdditionally, we have revised the manuscript to include an experiment that demonstrates how a state-based exploration objective (here, RND) instead of the TD-error-based objective for the Qx policy is insufficient to train Q, and demonstrate that neither policy converges to reward.\\n\\n- \\u201chow are actions chosen for the epsilon-greedy baseline since the action space is continuous (and not discrete)?\\u201d\\n\\nWith probability 0.1 we sampled from a uniform distribution between the minimum and maximum action values (typically -1 and 1).\\n\\n- \\u201cwhat is the difference between the lines Q and Qx are in Fig. 3.\\u201d\\n\\nIn Figure 3, Q is the performance of our exploitation policy implied by Q (which maximizes reward), while Qx is the performance of our exploration policy implied by Qx (which maximizes the TD-error of Q). We have relabled this figure in our revised draft to avoid confusion.\\n\\n- \\u201cWhy was DORA not included as a baseline algorithm?\\u201d\\n\\nDORA was not included as a baseline because DORA was never tested on large MDPs, and thus we were unsure how well it would perform by itself on such cases. We focused on RND instead as it is also related to the behavior of QXplore when no reward has been found, but has been demonstrated on large MDPs previously.\\n\\n- \\u201cAre two separate buffers necessary instead of a single shared buffer (with uniform sampling)?\\u201d\\n\\nWe have revised the paper to clarify that two buffers were used. While not essential, we found that using a skewed sampling ratio of self versus non-self data results in higher performance, possibly by reduced non-self data decreasing off-policy variance/stability issues. See Appendix D for empirical results.\\n\\n- \\u201cwhy was Q(s,a) replaced by V(s)?\\u201d\\n\\nWe opted for V(s) over Q(s,a) for the single-policy variant due to concerns about inaccurate target Q value prediction when training Q fully off-policy, and also to demonstrate that the QXplore paradigm works for value functions as well as Q-functions. We expect this variant to outperform the simpler ablation, and thus represent a better version of 1-policy QXplore.\"}", "{\"title\": \"Responses to Points Raised\", \"comment\": \"Thank you for your helpful comments and advice. We have address several points raised below:\\n\\n- \\u201cIncorrect or untuned RND baseline.\\u201d\\n\\nBecause we had to adapt RND to the off-policy Q-learning paradigm (a change that precludes the direct use of many parameters reported in RND\\u2019s Table 5), it is reasonable to be concerned about the quality of the implementation. However, we feel that our implementation and tuning is representative of a correct adaptation. Where possible, we used the parameters specified by RND in experiments (i.e. fixing intrinsic/extrinsic reward weight at 2), and we performed several baseline experiments across multiple exploration and non-exploration tasks (although not on Atari games) that showed evidence of performance differences in line with the results reported in the original paper. To further assuage concerns about parameter tuning, we will provide parameter sweeps over learning rate and reward weighting in a revision before the end of the rebuttal period.\\n\\n- \\u201c...adapting RND to be more like an ablation of your own approach, such as replacing the Qx TD-error rewards with random network distillation rewards...\\u201d\\n\\nThis is a good idea - thanks! At your suggestion, we\\u2019ve performed a new ablation where we replaced the Qx agent\\u2019s TD-error objective with RND\\u2019s state novelty objective. We found that this variant does not converge to achieve reward in the time allowed, and samples rewarding states too infrequently to train the Q function. We\\u2019ve included this experiment in our revised manuscript.\\n\\n- \\u201cNo comparisons to published RND baselines\\u2026 [including] Montezuma\\u2019s Revenge\\u201d\\n\\nWe considered this comparison, but felt that adapting the QXplore method to the Atari domain (with its image observations and observations) would require too much of a change in the method and require correct implementation of too many non-exploration-relevant details to draw useful inference for the task on continuous control domains. Furthermore, a recent paper [1] performed rigorous comparisons of exploration methods on Atari games and found that performance on Montezuma\\u2019s Revenge is not predictive of performance more broadly within the benchmark, and other exploration methods don\\u2019t significantly outperform epsilon-greedy on Atari tasks. \\n\\n- \\u201cNo comparison to value uncertainty methods...\\u201d\\n\\nThank you for letting us know about this work! We will compare with any value uncertainty methods which can be easily evaluated on SparseHalfCheetah and add such comparisons before the rebuttal period ends.\\n\\n- \\u201c...dropping the \\u2018biologically inspired\\u2019 bits\\u2026\\u201d\\n\\nWhile we believe that the biological relevance of this method should be explored more, we will leave that for future work. We have revised the manuscript and now only mention the connection in passing.\\n\\n- \\u201c...Fetch Results\\u2026 should be contextualized by the results obtained in the HER paper.\\u201d\\n\\nGood suggestion. We\\u2019ve added some description to note that QXplore is not state of the art compared to goal-directed RL methods like HER.\\n\\n- \\u201cNoisy TV problem also isn\\u2019t a problem for RND. Either compare to prediction-error based approaches and show that you're more robust to this issue, or drop it entirely\\u201d\\n\\nOur mention of QXplore\\u2019s robustness to the Noisy TV problem was intended to compare it against naive state-prediction-error approaches, not RND. We\\u2019ve revised the text with more context.\\n\\n[1] On Bonus Based Exploration Methods in the Arcade Learning Environment, Anonymously submitted to ICLR 2019, https://openreview.net/forum?id=BJewlyStDr\"}", "{\"title\": \"Paper Revisions\", \"comment\": \"We thank our reviewers for their helpful comments, and for their positive assessment of the significance of our method, which represents a novel approach to the exploration problem in RL in the realistic case where reward feedback is sparse but not singular, and which performs better than state novelty exploration in cases where novel states are poorly correlated with improving reward.\", \"we_have_made_the_following_major_changes_to_our_manuscript\": \"1. We included results using the original SparseHalfCheetah task with rewards {0, 1}.\\n2. We included an ablation comparing QXplore with a similar 2-policy algorithm where the exploration policy\\u2019s objective is changed to use RND\\u2019s state-novelty objective.\\n3. We included an ablation comparing QXplore with a variant using signed TD-error instead of unsigned TD-error.\\n4. We revised several sections to clarify points of confusion.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"The authors present a novel exploration method wherein an additional Q-function/policy is learned that treats abs(TD-error) of the standard Q-function as its reward function. Both policies are executed in parallel and experience is shared between them for off-policy learning. They demonstrate their method's superiority on a few continuous control tasks relative to RND. Overall, I thought the method was interesting and novel, but several concerns prevent me from endorsing its publication.\\n\\n1) Incorrect or untuned RND baseline.\\nAs the tasks under consideration are quite different from those in the RND paper, RND must be adapted significantly. It is therefor troubling that only a single RND-specific hyper-parameter is reported, when there should be many (e.g. ratio of intrinsic to extrinsic rewards, see Table 5 in their paper for many more).\\n\\nIt could be that you're adapting RND to be more like an ablation of your own approach, such as replacing the Qx TD-error rewards with random network distillation rewards. This is fine (and I'd suggest doing this as an additional baseline if not), but it should be labeled as such.\\n\\n2) No comparisons to published RND baselines. \\nAll of these problems could be avoided if the authors chose to run their method on tasks with published baseline results. Indeed, the lack of a Montezuma's Revenge run is particularly glaring. The author's are right to point out that Atari is perhaps not ideal for exploration since most novel states are rewarding, so I'm not expecting it to outperform RND per se. But the authors also claim that in these situations it won't do worse than RND, and without a Montezuma's Revenge run, this claim isn't well founded empirically.\\n\\n3)No comparison to value uncertainty methods.\\nThis is also made more frustrating when contrasted to the author's claim that \\\"to our knowledge posterior uncertainty methods have thus-far only been demonstrated in small MDP.\\\" Osband, Aslanides, and Cassirer (2018) has results on a continuous control task and Montezuma's Revenge. As QXplore is related to both RND and this class of value uncertainty methods, is feels as though the latter should also be included.\", \"smaller_points\": [\"No mention of prioritized experience replay. This method utilizes TD-error to bias replay rather than act explicitly, but there seems to be an interesting connection (acting to obtain more TD-error vs over-sampling high TD-error transitions) that was disappointingly ignored in this work.\", \"I'd suggest dropping the \\\"biologically inspired\\\" bits from the paper. The evidence you're citing (dopamine represents TD error + dopamine seeking behavior) could be used to justify any approach that optimistic with respect to value function uncertainty (e.g. thompson sampling on value function posterior), and doesn't really add anything.\", \"While the Fetch results are impressive, they should be contextualized by the results obtained in the HER paper. Obviously, HER isn't as general as this method, but it is worth reminding the reader that there is still a large performance gap between QXplore and the SOTA on these tasks.\", \"noisy TV problem also isn't a problem for RND. Either compare to prediction-error based approaches and show that you're more robust to this issue, or drop it entirely.\"], \"rebuttal_edit\": \"I appreciate the authors' responsiveness to my feedback on the text; the framing of biological plausibility and the connection to HER performance are both handled quite well now.\\n\\nI also appreciate the efforts the authors have demonstrated in adding new baseline results. That said, GEP and DORA results seem preliminary still (which is understandable given the timeframe), and would benefit from tuning and immediate results reporting in the case of GEP.\\n\\nI still find the treatment of RND problematic. \\\"Where possible, we used the parameters specified by RND in experiments (i.e. fixing intrinsic/extrinsic reward weight at 2)\\\" -- this is not the right attitude when testing such different domains and reinforcement learning algorithms. All of these hyper-parameters should be swept over in a manner analogous to your own approach before this comparison is valid.\\n\\nI acknowledge the Montezuma's Revenge (MR) isn't representative of exploration problems by itself -- this could be a good place to raise the bar and test against all 6 or so exploration-heavy problems in the Atari Suite. But even just running on MR would be better than introducing new tasks without published results from prominent exploration methods like RND. At the very least adding MR to your set of tasks would allow for comparisons against prior exploration results, which is the primary thing preventing me from raising my score.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes an exploration algorithm for reinforcement learning agents based on learning a separate Q-value function which treats TD-errors as rewards. Experiments on continuous control tasks with a difficult exploration component are used to highlight the effectiveness of the approach.\\n\\nThe proposed algorithm is interesting and has an intuitive appeal, the magnitude of the TD-error as an \\\"auxiliary reward\\\" seems like a natural choice to guide exploration. The experiments are comprehensive, including some ablation studies and robustness tests, with one caveat. Due to some experimental details (particularly concerning the environments), the empirical results may be invalid and it is difficult to assess them. There are also some certain parts that are unclear in the experimental results. Overall, because of these issues, I cannot recommend acceptance although I would be willing to increase my score if my concerns are addressed.\\n\\nConcerning the SparseHalfCheetah task, from Appendix F: \\\"...baseline reward level is -1 while a successful state provides 0 reward, but report reward values on a 0 to 500 scale for direct comparison with previous work\\\". This reward function is not sparse since it gives -1 at every step except at the goal which would incentivize the agent to explore naturally if it were acting greedily (or approximately greedily) with respect to the value function. For a sparse reward task, I would expect the reward to be 0 everywhere except at the goal state(s), where it would have some positive value. This is the case for the original reward function for SparseHalfCheetah described in [1]. This change to the reward function makes any comparisons to previous work questionable (Table 1) and could also affect the qualitative results in section 4.5. \\n\\nSimilarly, the other 3 tasks also do not have a sparse reward structure as the agent receives -1 at every timestep. This again is a confounder for assessing the exploration capabilities of the algorithm. It would be preferable to use other environments such as those in [1], which indeed have sparse rewards. \\n\\nFor these sparse reward problems where most rewards are zero, then before any nonzero reward is observed, it seems like Qx would mostly behave like previous algorithms based on a state novelty term (as discussed in sec 3.4). This could happen since the agent would only be observing rewards of 0 at this point during training. It is unclear to me then which benefits QXplore could have compared to previous algorithms such as DORA.\", \"other_points\": \"1) Some discussion of possible pitfalls for the algorithm could be added. For example, while the noisy TV problem may be avoided, there could be a noisy reward problem now. The absolute TD error could be high simply due to stochasticity in rewards. The agent could also be vulnerable to scenarios in which a stochastic transition brings the agent to either a state eventually leading to very high reward or to a state leading to no reward as this would cause the TD error be high (even if the value function converges).\\n2) The sentence before sec 3.3, \\\"Further TD-error-based exploration with a dedicated exploration policy removes the exploitation-versus-exploration\\ntradeoff that ... to an optimal Q-function.\\\" Could the authors clarify how TD-error-based exploration avoids the exploitation-versus-exploration\\ntradeoff? I do not see the connection here.\\n3) In the experiments, how are actions chosen for the epsilon-greedy baseline since the action space is continuous (and not discrete)? \\n4) I am bit confused as to what the difference between the lines Q and Qx are in Fig. 3.\\n5) Why was DORA not included as a baseline algorithm? It would seem to be the most closely related to QXplore when rewards are mostly constant (or zero).\\n6) Are two separate buffers necessary instead of a single shared buffer (with uniform sampling)? It seems to needlessly introduce complexity and additional hyperparameters (the ratios of samples from each buffer). If this is a key choice, then it should be mentioned in the paper.\\nThe text could also be clarified to indicate there are two buffers. In some places the writing suggests there is only one, e.g. sec3.3 \\\"...policy with a replay buffer shared between Q_\\\\theta andthe Q-function maximizing rx, which we term Qx.\\\"\\n7) In section 4.4, for the \\\"Single-Policy QXplore\\\", why was Q(s,a) replaced by V(s)? If I understand correctly, this experiment tries to test a variant where the TD error is used as a reward bonus. If this is the case, it seems like the best comparison would be to leave the original Q(s,a) which is learned by Q-learning instead of changing it to a state-value function. \\n\\nSuggestions (did not impact score)\\n- I wonder if there is a connection between Qx and the variance of the returns since the latter can be learned by using the squared TD-errror as a reward (see the Direct Variance algorithm in [2]) and the absolute TD-error is a similar quantity. In this way, it could be possible to frame QXplore as following a type of risk-seeking stategy (see [3] for an algorithm that makes use of the variance). \\n- I am not sure 'adversarial' is the right term to describe Q and Qx since the two policies are not in direct competition with each other.\\n\\n[1] \\\"VIME: Variational Information Maximizing Exploration\\\" by Houthooft et al.\\n[2] \\\"Directly Estimating the Variance of the \\u03bb-Return Using Temporal-Difference Methods\\\" by Sherstan et al. \\n[3] \\\"Deep Reinforcement Learning with Risk-Seeking Exploration\\\" by Dilokthanakul and Shanahan\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": [\"The paper proposes an exploration method based on the TD-error as an adversarial reward signal. The authors show that this reward signal has interesting exploration properties. They compare it empirically to RND and epsilon-greedy, showing better performance. Besides, they perform additional ablation studies to better investigate the properties of their approach.\", \"Though the paper proposes an interesting concept, it suffers from many weaknesses, some easy to fix and some which will require more work.\", \"Here are some remarks, in random order:\", \"at the end of the introduction, the authors mention inspiration from computational neuroscience models, but they do not come back to this aspect in their work. To me, these remarks should be removed from the paper and kept for another paper about the biological significance of the model. In the conclusion, the authors come back to the \\\"biological concepts of curiosity boredom, and exploration\\\", I would rather say they are psychological concept, and the authors should have a look at developmental psychology and developmental robotics if they really want to contribute in this respect (but not for this paper).\", \"some related work references are dispersed in the introduction, in Section 2, in the beginning of Section 3.2, in the end of Section 3.3 and in a few other places. The authors should build a proper \\\"related work\\\" section. Globally, the paper is poorly organized, e.g. Section 3.2 refers to Section 3.4 etc.\", \"given that Q_x's reward function is the unsigned TD-error, I would like to see whether QXplore can deal with problems taking both some positive and negative rewards.\", \"the authors mention RND and DORA as baselines, but only compare to RND. What about DORA? Despite the excitement it generated when published, I suspect RND is a rather weak baseline. There are many other exploration frameworks, the paper would be much stronger if the comparison was with respect to many other methods, such as GEP-PG (Colas et al. , ICML 18), Novelty search approaches, etc. To me, the weak comparison is the biggest weakness of this work.\", \"the authors compared themselves to approaches based on TRPO while they were using TD3 (this information is hidden in Appendix A and should be moved in the main paper. But then, is the difference in performance due to using TRPO which is known to be less sample efficient? Again, this makes the comparison very weak, the authors should rely on the same algorithm from both sides.\", \"the caption of Table 1 should be explicit about which algorithm comes from which paper.\", \"I found Section 4.5 very weak, it looks like mere \\\"handwaving\\\" and would deserve a proper quantitative study if the authors want to keep it. In my opinion, the authors should remove it for now and move Appendices E and F to the main paper instead. As is, Appendix E is very poor (by the way, the caption and the figure legend do not match, so we don't know which is which) and I had to look for the number of seeds until I found it hidden in Appendix F.\", \"Appendix C shows that, though the authors try to minimize the importance of this fact, their algorithm is very sensitive to initialization. It is very honest of them to have kept this study in their paper, but I'm afraid it strongly speaks against the algorithm.\", \"As described, SparseHaflCheetah is not that sparse (-1 everywhere and 0 when you succeed, as this reward scheme already favors exploration). It would be more informative to use 0 everywhere and 1 when successful.\"], \"typos\": \"Eq (1) and (3) should finish with a dot as it is the end of a sentence.\", \"p3\": \"MDP's => MDPs\\nSection 3.3, second sentence: avoid starting a sentence with a symbol.\\nSection 4, the second sentence is not a sentence (no main verb)\", \"p14\": \"is also no subject to it => not\"}" ] }

[ "Ilja Manakov", "Markus Rohm", "Volker Tresp" ]

[ "convolutional autoencoder", "bottleneck", "representation learning" ]

{ "note_id": [ "Dx2jD7gwx1", "HyxVI542sr", "HJeMC9mnoS", "HJeaaef2jH", "HygvX-XsiH", "BJlG9CCwsr", "HyeYdRRvsS", "BJezspRDjH", "SJgULT0Dir", "S1li_V7ecS", "Byg17qZy9S", "H1xTS8cptB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732145, 1573829196022, 1573825226001, 1573818565312, 1573757215206, 1573543561878, 1573543537280, 1573543321729, 1573543245980, 1571988594565, 1571916311463, 1571821124895 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1773/Authors" ], [ "ICLR.cc/2020/Conference/Paper1773/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1773/Authors" ], [ "ICLR.cc/2020/Conference/Paper1773/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1773/Authors" ], [ "ICLR.cc/2020/Conference/Paper1773/Authors" ], [ "ICLR.cc/2020/Conference/Paper1773/Authors" ], [ "ICLR.cc/2020/Conference/Paper1773/Authors" ], [ "ICLR.cc/2020/Conference/Paper1773/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1773/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1773/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper investigates the effect of convolutional information bottlenecks to generalization. The paper concludes that the width and height of the bottleneck can greatly influence generalization, whereas the number of channels has smaller effect. The paper also shows evidence against a common belief that CAEs with sufficiently large bottleneck will learn an identity map.\\n\\nDuring the rebuttal period, there was a long discussion mainly about the sufficiency of the experimental setup and the trustworthiness of the claims made in the paper. A paper that empirically investigates an exiting method or belief should include extensive experiments of high quality in to enable general conclusions. I\\u2019m thus recommending rejection, but encourage the authors to improve the experiments and resubmitting.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Results from different seed\", \"comment\": \"Unfortunately, we will not be able to address your criticism of the limited dataset set size in this submission. We thank you for the valuable discussion and will incorporate your feedback into our future work.\", \"regarding_your_other_main_point_of_criticism\": \"As we stated earlier, we have been training more models since the submission deadline. We have managed to temporarily acquire additional hardware resources to speed up training of the remaining models on a different seed. We have included the results in the most recent revision of the paper. The plots for models 12x12x48 and 12x12x192 for STL-10 stop at about 850 epochs, because they are still finishing their training. Nevertheless, you will find that the pattern for this seed remains the same as for the first. We hope that this increases your confidence in our results.\\n\\nWe thank you again for taking so much time to review our paper and discussing it with us.\"}", "{\"title\": \"Another response\", \"comment\": \"Thank you for your response.\\n\\nI understand that you think your findings will likely transfer to full dataset training. I just want to see more concrete evidence. \\n\\nIf you think studying small datasets is more relevant to other domains such as medical imaging, then please show results on medical imaging datasets. It is difficult to imagine how your findings translate to such a setting, as medical images have very different image characteristics as compared to CelebA or STl-10, and often have a fairly high resolution.\"}", "{\"title\": \"Response to response to rebuttal :)\", \"comment\": \"Thank you for replying to our rebuttal and continuing to participate in the discussion. We much appreciate the feedback we get from you!\\n\\nVarying the dataset size is a great idea, and we will definitely investigate it in future work. However, we want to make a few clarifications regarding this submission:\\n\\nHow and when CAEs overfit is one of the questions we try to address in our paper. However, it is only part of our investigation. In this submission, we offer a broad investigation of CAE characteristics: Effect of bottleneck shape on training dynamics, knowledge transfer and representation structure, testing of the copying CAE hypothesis, and overfitting behavior. We find all of these aspects to be underreported in literature. The sources you provide are an excellent example of this. They mostly provide qualitative results, focused on generative elements of the model, while completely lacking a description of the aspects we investigated in our submission. Only the WAE paper mentions that the model was trained for 55 epochs in the appendix, while the others do not disclose information about the training of the model. We want to emphasize that we do not criticize the quality of these papers but rather comment on the current focus of AE literature in general.\\n\\nWe understand you are concerned about the results not replicating when the models are trained on the full dataset. However, we do not believe this outcome to be likely. Our reasoning for this is two-fold:\\ni) Studies, such as Zhu et al. Do We Need More Training Data? (https://arxiv.org/pdf/1503.01508.pdf) have shown that adding more data quickly results in diminishing returns. Therefore, we do not expect generalization to increase to a point where all models achieve equal performance on the test data, even when we increase the number of samples.\\nii) The pattern we are reporting, regarding the effect of bottleneck shape, manifests itself early in training, both in training and test loss. At this stage, the model has clearly not captured all the variance present in the training data. We do not see how increasing the dataset size would drastically change this behavior.\\nAdditionally, many image datasets in other domains, such as medical imaging, do not have 100k+ samples, although they use CAEs frequently, especially for tasks like outlier detection. We, therefore, argue that limiting the dataset size to 10k samples is representative of what can we encounter in many domains.\\n\\nLastly, when I stated that \\u201cFrom our experience prior to conducting this investigation, we knew that CAEs do not seem to overfit in typical training scenarios.\\u201d I should have been more careful with the wording. \\\"Suspected\\\" would have been more appropriate (I am definitely learning about the danger of overstatement from this submission). I was referring to experiments, which we performed on medical imaging data where the validation error did not increase even when trained for over 1000 epochs on datasets with 3-6k samples (full public datasets) (here is a link to the figure https://ibb.co/FbbbLr4). This observation motivated the question as stated in this submission: \\\"How and when do CAEs overfit?\\\". To the best of our knowledge, the literature does not discuss this aspect at all. Therefore, we argue that our observations of the overfitting behavior (or apparent lack thereof for larger feature maps in the bottleneck) already provide novel insights into the workings of CAEs.\"}", "{\"title\": \"Response to rebuttal\", \"comment\": \"Thank you for the extensive rebuttal.\\n\\nFirst, I appreciate that the authors have toned down their claims a little, and I of course understand that resources can be scarce. Please mention the model sizes in the revised vision of your paper.\\n\\nI am still concerned about the lack of results on the influence of dataset size, specifically when looking at this separately from the dataset image complexity.\\nI understand that the question \\u201chow does the bottleneck size and shape influence overfitting\\u201d is what you want to investigate. My concerns/questions remain two-fold: i) Is it enough to only focus on image complexity when studying overfitting and the influence of the bottleneck? As overfitting is strongly influenced by dataset size, this seems like something that can, and in my opinion should, be studied more here. ii) As you state \\u201cFrom our experience prior to conducting this investigation, we knew that CAEs do not seem to overfit in typical training scenarios.\\u201d This then raises the question whether focusing on a scenario of overfitting, which doesn\\u2019t happen so often, is actually impactful enough for a full conference paper as it is unclear how the findings translate to \\u201ctypical training scenarios\\u201d. \\n\\nExperiments where the subset size of the dataset is changed would give more insight into how the findings in this paper can be extended to \\u201ctypical training scenarios\\u201d. I agree that when looking at overfitting, the complexity of the dataset in terms of image complexity is interesting.\\nBut when overfitting is studied, dataset size is also a critical component. As you say, confounding factors should be avoided. Stating that the trends remain the same when switching from the smaller pokemon dataset to a subset of CelebA/STL-10 (which differ both in image complexity as well as the dataset size) does not provide sufficient evidence (in my opinion) that the findings in the paper will transfer to scenarios where for instance the entire celeba/stl-10 datasets are considered. \\n\\nWith my statement \\\"When auto-encoders are studied in other literature, it is uncommon practice to restrict the dataset sizes like this.\\\" I am referring to other autoencoder papers (such as [1, 2, 3]) that also study celeba/stl-10, but don't restrict the size like you do. If you would show results where the size restriction is varied (including the full dataset), it would make the paper more convincing.\\n\\nI want to emphasise that I don't want to discourage the authors from attempting to publish this. I simply think that in its current form it's not impactful or complete enough to be a full conference paper. \\n\\n[1]Wasserstein Auto-Encoders. Tolstikhin et al. https://arxiv.org/abs/1711.01558\\n[2] beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework . Higgins et al. https://openreview.net/forum?id=Sy2fzU9gl\\n[3]Autoencoding beyond pixels using a learned similarity metric. Larsen et al. https://arxiv.org/abs/1512.09300\"}", "{\"title\": \"Response to Reviewer #3 (Sources)\", \"comment\": \"[1] Radhakrishnan, A., Yang, K., Belkin, M., & Uhler, C. (2019). Memorization in Overparameterized Autoencoders. ICML 2019 Deep Phenomena.\\n[2] Berthelot, D., Raffel, C., Roy, A., & Goodfellow, I. (2018). Understanding and Improving Interpolation in Autoencoders via an Adversarial Regularizer.\\n[3] Gregor, K., Papamakarios, G., Besse, F., Buesing, L., & Weber, T. (2018). Temporal Difference Variational Auto-Encoder.\\n[4] Kolouri, S., Pope, P. E., Martin, C. E., & Rohde, G. K. (2018). Sliced Wasserstein auto-encoders.\\n[5] Li, X., Chen, Z., Poon, L. K., & Zhang, N. L. (2018). Learning Latent Superstructures in Variational Autoencoders for Deep Multidimensional Clustering.\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"Dear Reviewer #3,\\nThank you very much for the in-depth review. We are aware that you must have invested a lot of time. Thank you also for the many suggestions for improving our paper. We have re-written portions of our paper to address or clarify the cases of overstatements, which you pointed out. We will revisit those later in our response. However, we wish to start with the two most prominent points of criticism we received from you.\\n\\n\\n\\\"It is unclear to me why the authors have chosen to only take a subset of the CelebA and STL-10 datasets for training and testing.\\\"\\n\\nAn important question we investigated was how and when CAEs overfit and what role the bottleneck shape plays in this process. From our experience prior to conducting this investigation, we knew that CAEs do not seem to overfit in typical training scenarios. Therefore, to be able to observe overfitting, we deliberately limited the dataset size and increased the number of epochs to 1000.\\nTraining on more data certainly affects the generalization of the model. However, we argue that the observed pattern remains unchanged, which is what we observe when going from the small Pokemon dataset to CelebA and STL-10: the generalization gap narrows but the overall trends remain the same\\nWe are not sure which other literature you are referring to when stating, \\\"When auto-encoders are studied in other literature, it is uncommon practice to restrict the dataset sizes like this.\\\" If it serves the research question, we believe that limiting the dataset size is a viable, investigative tool, e.g., as used in [1]. We agree, however, that we were not clear about our reasoning in the paper, and thank you for pointing it out. Thus, we have slightly modified the dataset section in the revision to correct this.\\nWe agree with your statement that \\\"Dataset size seems like a crucial part of dataset complexity.\\\" What we wanted to investigate, however, is how the complexity of the data samples influences the CAE training. We also did not always clearly communicate this and have tried to fix this in the revision.\\n\\n\\n\\\"It seems that the experimental validation is based on one run per CAE model with one single seed.\\\"\\n\\nWe understand your concern about the statistical significance of the results. However, we set up our experiments in a way that minimizes the number of moving parts and, therefore, the risk of confounding. By fixing the seed, we perform an intervention experiment to find the influence of the bottleneck shape on the CAE. Furthermore, we experimented on three different datasets and observed the same outcome every time, which, we believe, gives credibility to our results. Of course, we would have liked to repeat the experiments for other seeds but, unfortunately, our resources are limited. Because we wanted the bottleneck to be the only moving part, we had to preserve the total number of pixels in the input until the bottleneck layer (as explained in the methods section). This constraint resulted in massive models, the smallest of which had ~5x10\\u2077 parameters (the largest had ~9x10\\u2077). These models are on the order of ResNet152, which has ~6x10\\u2077 parameters. The training duration for a single model is on the order of days, and we trained 36 models in total, excluding the classifiers for the knowledge transfer experiments. Since the submission deadline, we have been training more models with a different seed, but so far, only 30 out of the 36 models could be trained. So far, we observe exactly the same pattern. Unfortunately, training will not finish before the revision deadline. Should our paper be accepted, we will, of course, include those results in the camera-ready version. However, we argue that our results show a level of significance and novelty that warrants publication, to inspire other researchers to replicate and expand upon them and hence deepen our understanding of CAEs and perhaps CNNs in general. To this end, we have published all of our code, models, and data, as mentioned in the paper.\\nWe understand that perhaps other authors share our struggle for resources, as many outstanding autoencoder-related papers from last year's ICLR also presented an analysis based on a single initialization. See [2], [3], [4] and [5] for examples.\\n\\n\\nLastly, we wish to thank you for highlighting the points where you believe we overstated our results. Our intention was not to question the works related to the loss landscape in DNNs and curriculum learning. We may have phrased these particular passages poorly, which is why we have changed them in the revised version. Regarding the 8x8x48 model, we wanted to express our surprise, that the model parameters underwent such a drastic change late in training and also hint at the possibility that there is more to discover here. The same goes for the part about curriculum learning.\\n\\nOverall, we thank you again for your consideration and hope we have addressed your concerns. If so, we would be happy if you re-evaluated our paper.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"Dear Reviewer #1,\\n thank you very much for appreciating our work. \\n\\n The novelty in our work lies in the systematic investigation of commonly held beliefs about CAEs. As far as we know (please correct us if we are wrong), these beliefs have never been thoroughly tested, despite their wide-spread adoption. As you have observed, we tried to formulate these beliefs as testable hypotheses and verify them experimentally. In the review, you state that \\\"further experiments to provide initial insights into the further questions raised by the authors would improve the 'novelty' aspect of the paper.\\\" Could you please provide specific experiments that you would like to see added to the paper? We will try our best to include them, if possible, before the revision deadline.\\n \\n The paper you mention in the review is fascinating, although probably not closely related to our work. Belkin et al. investigate how increasing the number of model parameters past the number of training samples increases generalization. In our case, the smallest model that we trained already had ~5x10\\u2077 parameters due to the experiment setup, which is far beyond the \\\"interpolation point\\\" mentioned in Belkin et al. Additionally, we demonstrate an increase in generalization while keeping the number of parameters fixed. For example, the models 3x3x48, 6x6x48, and 12x12x48 have the same number of parameters, yet their generalization is significantly different. Please let us know if we missed something.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"Dear Reviewer #2,\\nthank you very much for investing the time to review our paper and providing suggestions for improvement, which we tried to incorporate into our work.\\n \\n\\n\\\"A. How does (1) above help architecture search when designing CAEs in a new application? \\\":\\n\\n The most considerable areas of application for autoencoders, outside of representation learning in academia, are outlier detection, and image compression and denoising. Our results are readily applicable to both of these domains: In the case of outlier detection, the model should yield a high reconstruction error on out-of-distribution samples. Using smaller bottleneck sizes to limit generalization could prove useful in this scenario. Compression and denoising tasks, on the other hand, seek to preserve image details while reducing file size and discarding unnecessary information, respectively. In this case, a bigger bottleneck size is preferable, as it increases reconstruction quality at the same level of compression.\\nAdditionally, our paper demonstrates that CAE without a bottleneck does not learn simply to copy the input. Although further investigation is needed to determine what is preventing the CAEs from doing so, we believe that identifying the source of this phenomenon could lead to new architectures and training criteria that improve the abstraction capabilities of the CAE. We argue that the work we presented here is a starting point for such an investigation. We recognize that we did not communicate these implications clearly, and thank you for bringing them to our attention. We have re-written a portion of the discussion and conclusion in our paper to incorporate this aspect.\\n\\n\\n\\\"B. How does (2) above help style transfer if ``\\\"CAEs do not learn to copy their input.\\\" \\\"\\n\\nCurrently, we are not aware of CAEs being used in the field of style transfer. To the extent of our knowledge, style transfer mostly relies on a pre-trained VGG, but we have to admit, we are no experts in this field. Could you perhaps provide a specific example for the use of CAEs in style transfer? This way, we could potentially offer insight into how our work is related.\\n\\n\\n\\\"Concerns: If (2) is true, could the authors explain that almost perfect reconstruction results are shown in [*] ?\\\"\\n\\nThank you for pointing us to this paper. It was an interesting read. In the article, the authors use a combination of two representations, one is 64x64, and the other is 32x32, whereas the original input size is 256x256. First of all, copying is, by definition, not possible in this scenario, as the number of neurons in the bottleneck is smaller than the number of pixels in the input. The high-fidelity reconstruction presented in this paper nicely aligns itself with our findings. As we demonstrate in our paper, the reconstruction improves with increased bottleneck size (as in size of the feature maps). In our experiments, we only tested up to a bottleneck size of 16x16 with an original image size of 128x128. Extrapolating our findings, we would expect the reconstruction quality to improve further for bigger bottleneck sizes, which can be seen in Figure 3 of [*] when they include the larger feature maps.\\nFurthermore, they include a sophisticated system for incorporating hierarchical features into the architecture, which presumably also benefits abstraction and in turn, better reconstruction. Additionally, one can formulate the hypothesis, that the bottleneck size in [*] is so big that the task of the decoder is potentially more akin to 4x super-resolution than to decoding, and many papers have demonstrated the power of CNNs in this setting (e.g., Wang 2018). However, this is not investigated in the article, so a separate experiment is necessary to verify this. \\n\\nWang, Xintao, et al. \\\"Esrgan: Enhanced super-resolution generative adversarial networks.\\\" Proceedings of the European Conference on Computer Vision (ECCV). 2018\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper investigates convolutional autoencoder (CAE) bottleneck. The research problem is important, as CAE is widespread adopted. It cam be interesting to shed insight into how the bottleneck works.\\n\\nIn particular, two observations are made: (1) By measuring the performance of the latent codes in downstream transfer learning tasks, the authors show that increased height/width of the bottleneck drastically improves generalization; The number of channels in the bottleneck is secondary in importance. (2) CAEs do not learn to copy their input, even when the bottleneck has the same number of neurons as there are pixels in the input.\\n\\nIt would make this submission more convincing if the authors show the some important applications of their findings. For example, \\n\\nA. How does (1) above help architecture search when designing CAEs in a new application? \\nB. How does (2) above help style transfer if ``\\\"CAEs do not learn to copy their input\\\"\", \"concerns\": \"If (2) is true, Could the authors explain that almost perfect reconstruction results are shown in [*] ?\\n\\n[*] Generating Diverse High-Fidelity Images with VQ-VAE-2\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"The authors evaluate convolutional autoencoders (CAE) by varying the size (width & height) and depth of the bottleneck layer on three datasets and compare test and training performance. They furthermore evaluate the quality of the bottleneck activations for linear classification. The authors also investigate the belief that a bottleneck layer of size equal to the input image will copy the image.\\n\\nI am not an expert in the field of (C)AEs. As such, I cannot approriately judge the relevance of the questions which are answered here. In the following, I will therefore make the assumption that those questions are relevant.\\nIf so, I (weakly) recommend accepting the paper. \\n\\nWhile it does not propose any novel algorithms, it does ask a clear question and provides a compelling experimental answer, which (assuming the question is relevant), should be interesting for the community. \\nOn the other hand, further experiments to provide initial insights into the further questions raised by the authors would improve the 'novelty' aspect of the paper.\", \"minor_comment\": \"I believe the paper \\\"Reconciling modern machine learning practice and the bias-variance trade-off\\\" by Belkin et. al is relevant.\", \"edit\": \"Thank you for your response.\\n\\nI wasn't referring to any specific experiments, but the questions you are raising in the paper, for example in the conclusion section.\", \"regarding_belkin_et_al\": \"Sorry for the misunderstanding, I didn't mean to imply that it's highly relevant, but I thought it might be a good addition to the related work section as it also relates to the generalization of neural networks.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\nThis paper studies some of the properties of fully convolutional autoencoders (CAE) as a function of the shape and total size of the bottleneck. They train and test CAEs with bottlenecks consisting of different ratios of spatial resolution versus number of channels, as well as different total number of neurons. The authors investigate which type of change in the bottleneck is most influential on training behavior, generalization to test set, and linear separability for classification/regression. Their first main finding is that the spatial resolution of the bottleneck is a stronger influencer of generalization to the test set than the number of channels and the total number of neurons in the bottleneck. The second main finding is that even when the total number of neurons in the bottleneck is equal to the data input size, the neural network does not appear to simply learn to copy the input image into the bottleneck.\", \"decision\": \"\", \"weak_reject\": \"It is always refreshing to see papers that address/challenge/investigate common assumptions in deep learning. However, I find the experimental findings and discussion of borderline quality for a full conference paper. It might be more suitable as a good workshop paper.\", \"supporting_arguments_for_decision\": \"It is unclear to me why the authors have chosen to only take a subset of the CelebA and STL-10 datasets for training and testing. It seems like dataset size is also an important factor that increases the complexity for training a model, and it certainly affects how a model can generalize. When auto-encoders are studied in other literature it is uncommon practice to restrict the dataset sizes like this, so this makes me question the applicability of this paper\\u2019s results to the literature. \\n\\nIt seems that the experimental validation is based on one run per CAE model with one single seed. This is on the low side of things, especially when quite extensive claims are made. An example of such a claim is on page 6 when discussing a sudden jump in training and test scores for 8x8x48 model trained on the Pokemon dataset. Because the same behavior appeared when the authors repeated the experiment with the same seed, the authors conclude \\u201cThis outlier suggests, that the loss landscape might not always be as smooth towards the end of training, as some publications (Goodfellow et al., 2014) claim and that \\u2018cliffs\\u201d (i. e., sudden changes in loss) can occur even late in training.\\u201d Making this claim based on something that occurs with one single model for a single seed is not convincing and overstating this finding. \\nAnother example is on page 7 under bullet point 4, where the authors discuss the obtained evidence against copying behaviour when the bottleneck is of the same size as the input. The authors state \\u201cWe believe this finding to have far-reaching consequences as it directly contradicts the popular hypothesis about copying CAEs.\\u201d The paper definitely shows some empirical evidence that supports the claim that copying does not occur, but these findings are all done with a single seed and by considering small subsets of datasets (celebA and stl-10). In my opinion, it is therefore too much to state that the current findings have far reaching consequences. It has potential, but I wouldn\\u2019t go much further than that. \\n\\nOn page 7 in the second to last paragraph the influence of dataset complexity is discussed. The authors state \\u201cthe loss curves and reconstruction samples do not appear to reflect the notion of dataset difficulty we defined in Section 2.3\\u201d and \\u201cThis lack of correspondence implies that the intuitive and neural network definitions of difficulty do not align. Nevertheless, a more detailed study is required to answer this question definitively as curriculum learning research that suggests the opposite (Bengio et al., 2009) also exists.\\u201d It is unclear to me what the authors expected to find here. Moreover, the absence of major differences across the chosen datasets does not immediately make me doubt or question results from curriculum learning. My skepticism is again enhanced by the fact that the authors have taken a subset of the data for the more complex celebA and STL-10 datasets. Dataset size seems like a crucial part of dataset complexity.\\n\\nAn interesting smaller finding is that linear separability of the latent codes for classification is better on the test set for the pokemon dataset, even though the training and test reconstruction losses showed signs of overfitting. The authors hypothesize that overfitting might occur more in the decoder than in the encoder.\\n\\nAdditional feedback to improve the paper (not part of decision assessment)\\n- Section 2.3.1: what was the original resolution of the pokemon dataset?\\n- Section 3.1, c_j is used as the compression level, but in eq 3 and 4 c^i is also used to indicate the number of channels. - For clarity please use n_ci in eq 2 and 3 or choose a different all together for the compression level.\\n- Please increase the font size of the plots in figures 1, 3 and 4.\"}" ] }

[ "Nir Baram", "Shie Mannor" ]

[ "Reinforcement Learning", "Imitation Learning", "Behavior Cloning", "Partial Simulation" ]

{ "note_id": [ "y8Yz3g8YNc", "SygLpncfjS", "Syeg935GoS", "H1eDzhcfiB", "SkeQKHjaYH", "rkgvO5taFB", "r1xJuNzPYH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732113, 1573199037577, 1573198984344, 1573198862842, 1571825019065, 1571818095080, 1571394662611 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1772/Authors" ], [ "ICLR.cc/2020/Conference/Paper1772/Authors" ], [ "ICLR.cc/2020/Conference/Paper1772/Authors" ], [ "ICLR.cc/2020/Conference/Paper1772/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1772/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1772/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper introduces the concept of an Expert Induced MDP (eMDP) to address imitation learning settings where environment dynamics are part known / part unknown. Based on the formulation a model-based imitation learning approach is derived and the authors obtain theoretical guarantees. Empirical validation focuses on comparison to behavior cloning. Reviewers raised concerns about the size of the contribution. For example, it is unclear to what degree the assumptions made here would hold in practical settings.\", \"title\": \"Paper Decision\"}", "{\"title\": \"response to review #3\", \"comment\": \"We thank the reviewer for his detailed response. Below is our response to the main issues raised. Other issues will be fixed or handled in the PDF.\\n\\n\\u2022\\tQ: \\u201c\\u2026This modeling implicitly assumes that the agent cannot influence the unresponsive part of the state space. This is too restrictive\\u2026\\u201d\\n\\u2022\\tA: The eMDP model bridges between BC and contemporary imitation methods by using the BC kernel for features we cant simulate, and the responsive kernel for features we can. While with BC the agent cant influence any part of the state space, our model allows some interaction to occur. Our experiments show that even partial interaction is better than zero interaction.\\n\\n\\u2022\\tQ: I think the practicality of eMDP is too limited.\\n\\u2022\\tA: To the best of our knowledge, this is the first and only method that allows imitation in a multiplayer setup.\\n\\n\\u2022\\tQ: Also, it is unclear what happens when the transition of unresponsive state space is stochastic. In this case, the unresponsive transitions of eMDP would be incorrect since eMDP assumes deterministic transitions as-is from demonstrations. Though I am not sure about this.\\n\\u2022\\tA: eMDP does not assume deterministic transitions.\\n\\n\\u2022\\tQ: The state-value function\\u2019s error bound ignores a policy. The proof utilizes the state distribution P(s\\u2019|s). However, this distribution should depend on a policy. It is unclear to me what is the policy function used in this error bound. Does this error-bound hold for any policy or only for the optimal policy?\\n\\u2022\\tA: The value function error is true for any policy and not just for the optimal one, thus it was omitted. A comment about it should be added to the text.\\n\\n\\u2022\\tQ: Empirical evaluation lacks strong baseline methods. The paper compares the proposed method against behavior cloning, which is known to perform poorly under limited data. The paper requires stronger baseline methods such as those mentioned in Section 3.\\n\\u2022\\tA: Unfortunately, this is not possible. Stronger baselines require full knowledge of F, which does not apply in our setting. For example, we cannot provide rules for the behavior of the opponent human player.\\n\\n\\n\\u2022\\tQ: Does the reward function in Eq. (5) correspond to the Wasserstein distance for any metric d_s? If it is not, then the value function\\u2019s error-bound for the Wasserstein distance is not applicable to this reward function.\\n\\u2022\\tA: The reward is defined over the metric space S while the Wasserstein distance is defined between probability distributions. In Section 5, we relate the metric distance to the distribution distance. Please see footnote 2 at the bottom of page 6.\\n\\n\\u2022\\tQ: What is the reward function r(s) in the value function\\u2019s error-bound that we can control the Lipschitz constant? We do not know the ground-truth reward r(s,a) so it cannot be controlled. The reward r(s,s\\u2019) of eMDP is given by the metric d_s of the state space so again we cannot control it.\\n\\u2022\\tA: Yes. This is the reward from Eq. 5. It is determined by the state space and cannot be controlled.\\n\\n\\u2022\\tQ: What is the metric d_s used in the experiments?\\n\\u2022\\tA: This is the reward from Eq. 5. We assume no knowledge about the state space and use a Dirac delta function at the event s_r=s_{r,e}. In some of the environments (that involve a shooting action) we implemented the Dirac function asymmetrically (agent shoots and experts don\\u2019t is rewarded differently than expert shoots and agents don\\u2019t).\\n\\n\\u2022\\tQ: How do you perform BC without expert actions in the experiments?\\n\\u2022\\tA: The BC baseline \\u201ccheats\\u201d by looking at expert actions, while the eMDP model doesn't access A.\"}", "{\"title\": \"Response to review #2\", \"comment\": \"We thank the reviewer for his positive feedback. Below is our response to the issues raised.\\n\\n\\u2022\\tQ: For the experimental part, BC baselines for multiplayer games just imitates the target player's behavior (player1), completely ignoring the states/actions of player2 (ignoring the states of bricks and balls for breakout-v0)?\\n\\u2022\\tA: The BC baseline receives the original (un-factored) state as an input. In Breakout it is the complete frame including the bricks, ball and expert racket. The baseline is also granted access to expert action which we do not provide to the eMDP model. I.e., the BC baseline \\u201ccheats\\u201d.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"We thank the reviewer for his important comments and apologies for the issue with the link anonymization. Below is our response to the issues raised.\\n\\n\\u2022\\tQ: Isn't it problematic that the imitation agent is trained in an environment where unresponsive state features from the demonstrations are replayed and not affected by the agent's actions?\\n\\u2022\\tA: Yes. Ideally, you would like to use a \\u201cfull\\u201d simulation to eliminate the problem. However, most real-world problems do not have a simulator yet. Our experiments show that even partial interaction of the agent with the environment is better than zero interaction that happens with simulation free methods such as BC. Moreover, our transition kernel becomes less biased as the agent improves.\\n\\n\\u2022\\tQ: It would be nice to expand Section 4 to explain why it makes sense to reflect the unresponsive component in the transition kernel.\\n\\u2022\\tA: Yes, we agree. The reflection operator is the transition kernel of BC where the prediction of the agent on a state does not affect the next state. The eMDP model bridges between BC and contemporary imitation methods by using the BC kernel for features we cant simulate, and the responsive kernel for features we can simulate. As stated before, the reflection operator also possesses the nice property of being unbiased when pi=pi_E and it requires zero calculations, thus dramatically accelerating training time. \\n\\n\\u2022\\tQ: What is the purpose of observing unresponsive features, if the reward function for the imitation agent is only defined in terms of the responsive features?\\n\\u2022\\tA: The reward is defined over a) s_r: the responsive features of the agent which are observed in the state space, and b) s_{r,e}: the responsive features of the expert which are unobserved. How can the agent maximize the reward if it depends on unobserved features? It does so by observing s_u: the unresponsive features, which define (or at least have high mutual information with) s_{r,e}.\\n\\n\\u2022\\tQ: Is it that the imitation policy is conditioned on the unresponsive features?\\n\\u2022\\tA: Yes, of course, the unresponsive features are a part of the state space and the agent and expert share the same state space. Think of Pong for example where the unresponsive features represent the opponent racket and the ball. The agent can't take actions without having this info.\\n\\n\\u2022\\tQ: What distance metric was used to define the reward function?\\n\\u2022\\tA: We report the reward defined in Eq.5 (the optimization problem we wish to solve). In words, the reward is a Dirac delta function on the event s_r=s_{r,e}. It is the most restrictive form of similarity and requires zero prior knowledge. Calculating it requires extracting s_r and s_{r,e} at each step as explained in section 4. In the Atari games, we do the extraction manually from frame pixels. However, with access to the game\\u2019s ROM, this can be done much easier.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposes an imitation learning algorithm that learns from state-only demonstrations and assumes partial knowledge of the transition dynamics. The demonstrated states are decomposed into \\\"responsive\\\" and \\\"unresponsive\\\" features. The imitation policy is trained in an environment where the responsive state features are simulated and controllable by the agent, and the unresponsive state features are replayed from the demonstrations. The agent is rewarded for tracking the responsive state features in the demonstrations.\\n\\nOverall, I was confused by this paper. Isn't it problematic that the imitation agent is trained in an environment where unresponsive state features from the demonstrations are replayed and not affected by the agent's actions? It would be nice to expand Section 4 to explain why it makes sense to reflect the unresponsive component in the transition kernel. It would also be helpful to include some of this information in Section 1.\\n\\nI am also confused by the method. What is the purpose of observing unresponsive features, if the reward function for the imitation agent is only defined in terms of the responsive features? Is it that the imitation policy is conditioned on the unresponsive features?\\n\\nThere are several important experimental details missing from Sections 4 and 6. What distance metric was used to define the reward function? For example, was it Euclidean distance in pixel space for the Atari games? The main experimental results in Figure 2 are difficult to interpret without knowing the how the reward function is defined in the eMDP. Could the authors either provide definitions of the reward functions of the eMDPs in the experiments, or measure performance of the imitation agents using more interpretable or standardized metrics (e.g., game score) for the chosen environments?\\n\\nIt is also concerning that the Google Drive links to code and supplementary materials aren't anonymized.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper provides a theoretical basis called expert-induced MDP (eMDP) for formulating imitation learning through RL when only partial information on the transition kernel is given. Examples include imitating one player in multiplayer games where the learner cannot explicitly have the transition kernel for other players. eMDP just uses the available part of transition kernel (e.g. the target player's state-action transitions) for learning, but substitute the next state from demonstrations 'as is' into the unavailable part. In this situation, using IPM (Wasserstein-like) divergence, it is easily guaranteed that eMDP is consistent with the optimal solution in the limit. However, this does not guarantee anything when the optimum is not reached. The paper's main contributions would be to provides error bounds (Lemma 5.2 in particular) for the convergence rate in terms of Lipschitz constants and the maximum error between eMDP and the demonstration set. This roughly means that when transition functions are continuous and smooth and the sample demonstration set well characterize the underlying MDPs, then eMDP is theoretically guaranteed to work well. If these factors (Lipschitz continuity and the max deviation of state distributions conditioned on the given demonstration set) are expected to be small, eMDP can nicely perform learning with partial knowledge. Some demonstrations confirm eMDP works when simple BC fails in multiplayer games.\\n\\nOverall, the paper is worth accepting. The proposed eMDP would be a technically simple solution, but providing a solid theoretical basis for that solution to these commonly seen situations of imitation learning with partial knowledge would be quite beneficial. The paper is well written with intuitive examples and nice overview summaries for the theoretically complicated parts. Though the implications of the experimental part are not clear, the paper fairly states this point and also provides a video for visual assessment with displaying ghost imitations.\", \"one_minor_question\": \"For the experimental part, BC baselines for multiplayer games just imitates the target player's behavior (player1), completely ignoring the states/actions of player2 (ignoring the states of bricks and balls for breakout-v0)?\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": [\"Paper summary.\", \"The paper considers an IL problem where partial knowledge about the transition probability of the MDP is available. To use of this knowledge, the paper proposes an expert induced MDP (eMDP) model where the unknown part of transition probability is modeled as-is from demonstrations. Based on eMDP, the paper proposes a reward function based on an integral probability metric between a state distribution of expert demonstrations under the target MDP and a state distribution of the agent under the eMDP. Using this reward function, the paper proposes an IL method that maximizes this reward function by RL. The main theoretical result of the paper is that the error between value function of the target MDP and that of the eMDP can be upper-bounded. Empirical comparisons against behavior cloning on discrete control tasks show that the proposed method performs better.\", \"Rating.\", \"The main contribution of the paper is the eMDP model, which enables utilizing prior knowledge about the target MDP for IL. While this idea is interesting, eMDP is too restrictive and its practical usefulness is unclear. Moreover, there are other issues that should be addressed such as clarity and experiments. I vote for weak rejection.\", \"Major comments:\", \"Limited practicality due to an assumption on the unresponsive transition kernel.\", \"In eMDP, the unresponsive transition kernel is modeled from demonstrations by directive using the observed next states (in demonstrations) as a next state of the agent. This modeling implicitly assumes that the agent cannot influence the unresponsive part of the state space. This is too restrictive, since actions of the agent usually influence other parts of the state space such as opponents and objects. For instance, in pong, actions of the agent indeed influence trajectories of the ball. Due to this restrictive assumption, I think the practicality of eMDP is too limited.\", \"Also, it is unclear what happens when the transition of unresponsive state space is stochastic. In this case, the unresponsive transitions of eMDP would be incorrect since eMDP assumes deterministic transitions as-is from demonstrations. Though I am not sure about this.\", \"The equality in Eq. (3) should be an upper-bound. The IPM is defined by the absolute difference between summations (expectations), but the right-hand side of Eq. (3) is summations (expectations) of the absolute difference. These two are not equal, and the right-hand side should be an upper-bound.\", \"The paper is difficult to follow. There are many skipping contents in the paper. Especially, in the description of eMDP and its solution, where the paper describes the optimal solution of eMDP (Section 2) before describing the eMDP model (Section 4). Also, it is unclear from my first reading pass what is the actual IL procedure. Including a pseudo-code would help.\", \"The state-value function\\u2019s error bound ignores a policy. The proof utilizes the state distribution P(s\\u2019|s). However, this distribution should depend on a policy. It is unclear to me what is the policy function used in this error bound. Does this error-bound hold for any policy or only for the optimal policy? In the case that it only holds for the optimal policy, this result still does not provide useful guarantees when the optimal policy is not learned, similarly to the result in Proposition 2.1.\", \"Empirical evaluation lacks strong baseline methods. The paper compares the proposed method against behavior cloning, which is known to perform poorly under limited data. The paper requires stronger baseline methods such as those mentioned in Section 3. Also, it is strange that the vertical axis in Figure 2 represents the reward function of eMDP, which is an artificial quantity in Eq. (5). The results should present the reward function of the target MDP, which is the ground-truth reward of the task.\", \"Besides the major comments, I have few minor comments and questions:\", \"Does the reward function in Eq. (5) correspond to the Wasserstein distance for any metric d_s? If it is not, then the value function\\u2019s error-bound for the Wasserstein distance is not applicable to this reward function.\", \"What is the reward function r(s) in the value function\\u2019s error-bound that we can control the Lipschitz constant? We do not know the ground-truth reward r(s,a) so it cannot be controlled. The reward r(s,s\\u2019) of eMDP is given by the metric d_s of the state space so again we cannot control it.\", \"What is the metric d_s used in the experiments?\", \"How do you perform BC without expert actions in the experiments?\", \"Confusing notations. E.g., the function r is used for r(s,a) (MDP\\u2019s reward), r(s, s\\u2019) (eMDP\\u2019s reward), and r(s) (reward in the error bound\\u2019s proof); these r\\u2019s have different meaning and should be denoted differently.\", \"Typos. E.g., \\u201cLemma 2.1\\u201d in Section 5 should be \\u201cProposition 2.1\\u201d, \\u201cLemma A.1\\u201d should be \\u201cProposition 2.1\\u201d. \\\"tildeF\\\". etc.\", \"*** After authors' response.\", \"I read the other reviews and the response. My major concern was the impracticality of eMDP. While I can see that the eMDP model provides advantages over BC, I still think that it is impractical since it assumes the agent cannot influence the unresponsive part. Moreover, extracting next-unresponsive states as-is from demonstrations is problematic with stochastic transitions, since actual next-unresponsive states can be different from next-unresponsive states in demonstrations even when the sames state and actions are provided. Given that the major concern remains, I still vote for weak rejection.\"]}" ] }

[ "Muhammad Rizki Maulana", "Lee Wee Sun" ]

[ "Deep learning", "few-shot learning" ]

{ "note_id": [ "c8lmtADCE", "SylXjPS3ir", "SJgTwPrnjB", "H1eexvS3sH", "S1l8TWFRFS", "S1esrDQ0YS", "ryx1J6kTFB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732082, 1573832603044, 1573832549266, 1573832423625, 1571881406401, 1571858243508, 1571777751444 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1771/Authors" ], [ "ICLR.cc/2020/Conference/Paper1771/Authors" ], [ "ICLR.cc/2020/Conference/Paper1771/Authors" ], [ "ICLR.cc/2020/Conference/Paper1771/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1771/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1771/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"Main content:\\n\\n[Blind review #3] The authors propose a metric based model for few-shot learning. The goal of the proposed technique is to incorporate a prior that highlight better the dissimilarity between closely related class prototype. Thus, the proposed paper is related to prototypical neural network (use of prototype to represent a class) but differ from it by using inner product scoring as a similarity measure instead of the use of euclidean distance. There is also close similarity between the proposed method and matching network.\\n\\n[Blind review #2] The stated contributions of the paper are: (1) a method for performing few-shot learning and (2) an approach for building harder few-shot learning datasets from existing datasets. The authors describe a model for creating a task-aware embedding for different novel sets (for different image classification settings) using a nonlinear self-attention-like mechanism applied to the centroid of the global embeddings for each class. The resulting embeddings are used per class with an additional attention layer applied on the embeddings from the other classes to identify closely-related classes and consider the part of the embedding orthogonal to the attention-weighted-average of these closely-related classes. They compare the accuracy of their model vs others in the 1-shot and 5-shot setting on various datasets, including a derived dataset from CIFAR which they call Hierarchical-CIFAR.\\n\\n--\", \"discussion\": \"All reviews agree on a weak reject.\\n\\n--\", \"recommendation_and_justification\": \"While the ideas appear to be on a good track, the paper itself is poorly written - as one review put it, more like notes to themselves, rather than a well-written document to the ICLR audience.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Clarification for Official Blind Review #3\", \"comment\": \"Thanks for reviewing our paper.\\n\\n\\nSimilarity with Prototypical Networks and Matching Network\\n---\\nOur proposed method is similar to the prototypical networks (Snell et al., 2017) -- and subsequently Mensink et al. (2013) -- in its use of mean representation of class (or prototypes). The similarity stops there, as the prototypical networks directly perform classification by comparing the distance of the new input to each prototype. They assume that the embedding function that produces the prototypes can sufficiently capture useful and general enough representation that is transferable to the novel set; it only computes global task embedding. As shown in our results, their assumption breaks down when there are changes in class granularity or the label is of fine granularity. In contrast, our method does not directly classify on the prototypes; instead, it transforms the prototypes, producing task-aware embeddings that are locally orthogonal to the shared components belonging to different classes. Thus, classification is performed by computing a Softmax over dot product between the new point and the task-aware prototypes.\\n\\t\\nThe Dissimilarity Network uses context embedding similar to the full context embedding (FCE) extension of the matching network (Vinyals et al., 2016). However, there are some glaring differences in how they operate. The matching network carries the entire support for the prediction. It predicts the label of an unknown point by computing the linear combination of the label of its support set; as the support set grows, the memory increases linearly with it. Their full context embedding conditions the prediction on the entire support set; computing the task-aware embedding is quadratic in the number of elements of the support set. Moreover, as they do not construct an explicit reference for the classes to condition into, it is less clear how reasonable separation of a point belonging to different classes can be maximized. Our model only maintains a set of prototypes and classifies a new point based on how orthogonal its representation to the prototypes. Since we condition the prediction only on the prototypes, it will not grow with the size of the support set. Moreover, computing the task-aware embedding is only quadratic in the number of prototypes (i.e., labels) -- as opposed to the number of support set. We also explicitly computes the representations to be dissimilar; lifting the reliance on learning sufficiently separable inter-class representations only to the loss function.\\n\\nWe have also updated our related works to give more focus on the metric and similarity learning methods.\\n\\n\\nNovelty, Mathematical Inaccuracies, & Clarity issues\\n---\\nWe have updated the paper to better highlight the novelty of our work, as well as incorporating the comments given.\\n\\n\\nComparison Fairness\\n---\\nFour-layer convolution backbone (Conv-4) with an input size of 84x84 is used as a feature extractor for all methods for a fair comparison. Apart from that, our method is parameterized by the self-attention with 2 layers BLSTM for query, key, and value. The rest of the methods follow exactly from their description in their respective papers. For example, RelationNet is parameterized by an additional 2 layers of convolution networks and 2 fully-connected layers.\"}", "{\"title\": \"Clarification for Official Blind Review #1\", \"comment\": \"Thanks for taking the time to review our paper.\\n\\n\\nExplanation about Components\\n---\\nWe have updated the manuscript to include more explanation about the methods that we rely upon.\\n\\n\\nTerms & Similarity Learning\\n---\", \"we_have_updated_the_manuscript_to_use_more_appropriate_terms\": \"inductive bias and similarity learning.\\n\\nOur method adopts the approach of similarity learning. Instead of learning a distance or similarity function, we learn a space (embedding) that works well with a fixed similarity-based classifier in that space. Specifically, our model learns to construct a space that is optimized to separate data that belong to different classes for a classifier that uses dot-product as its similarity function.\\n\\n\\nApplication on General Classification Tasks\\n---\\nIn the usual classification tasks, the labels are fixed during the training and testing. Our approach is advantageous if the labels are changing between task -- as in few-shot classification.\\n\\n\\nDetailed Comments\\n---\\nWe have updated our manuscript based on the above comments. We also updated the name of our harder variant of CIFAR dataset to CIFAR-Hard to avoid confusion -- as CIFAR is already hierarchically labeled. For point (1), it\\u2019s true that our new dataset can be seen through the lens of experiment design. However, it\\u2019s also true that the dataset is new, in a sense that the dataset comprises of a new set of image label pairs, with the label derived from the original CIFAR.\\n\\n\\nResults & Discussions\\n---\\nWe have added more discussion about the results in the manuscript. Please have a look at the updated manuscript.\"}", "{\"title\": \"Clarification for Official Blind Review #2\", \"comment\": \"Thank you very much for your review.\\n\\n\\nSimilarity Learning\\n---\\nWe have updated our manuscript and present the approach as similarity learning. We learn a function that maps to a space that is optimized for dissimilarity. Since our notion of dissimilarity is based on vector orthogonality, it\\u2019s only natural that we use an inner product as a similarity function in such space. \\n\\n\\nPermutation Invariance\\n---\\nWe are aware of the sequential nature of BLSTM, which can be counter-intuitive as we are modeling a set-to-set operation which shouldn\\u2019t have any preference for ordering. However, we found empirically that this setup offers more performance gain compared to the use of traditional linear function as attention embedding function. The BLSTM may learn to ignore the unimportance of set ordering due to the nature of episodic training, which exposes it to many permutations of the possible class-orderings. Moreover, the attention also gives a global context of the member of the set, which could further alleviate the ordering issues (if any).\\n\\n\\nHarder Benchmark\\n---\\nOur approach requires the dataset that we derived from to have at least two different levels of class granularity. For example, CIFAR dataset which has two levels of labels granularity. ImageNet labels also form a hierarchy which -- through this method -- can be derived into several hard few-shot classification datasets. In the case where different labels of granularity are absent, one may be able to construct new labels by exploring the natural hierarchy which may present.\\n\\nWe leave out the detail of the construction of the validation set, not the dataset.\\n\\n\\nComparison Fairness\\n---\\nNo finetuning is performed on CIFAR-H and CIFAR-FS, all the hyperparameters are the one used in the mini-ImageNet dataset. This applies to all methods (including ours).\\n\\n\\nResults & Discussions\\n---\\nWe have added more discussion about the results in the manuscript. Please have a look at the updated manuscript. \\n\\n\\nMathematical Inaccuracies, Clarity, and Grammatical Errors\\n---\\nWe have updated the manuscript to address the issues.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The stated contributions of the paper are: (1) a method for performing few-shot learning and (2) an approach for building harder few-shot learning datasets from existing datasets. The authors describe a model for creating a task-aware embedding for different novel sets (for different image classification settings) using a nonlinear self-attention-like mechanism applied to the centroid of the global embeddings for each class. The resulting embeddings are used per class with an additional attention layer applied on the embeddings from the other classes to identify closely-related classes and consider the part of the embedding orthogonal to the attention-weighted-average of these closely-related classes. They compare the accuracy of their model vs others in the 1-shot and 5-shot setting on various datasets, including a derived dataset from CIFAR which they call Hierarchical-CIFAR.\\n\\nOverall, while we like the concepts/ideas and the problem is definitely important, we were not enthusiastic about the paper. First, we found the write up to be cryptic, involving very long unclear statements. It read as if the authors were writing for themselves and not for ICLR general audience. Beyond the writing style, we found the paper to have:\\n\\n* Inadequate description of the model, including mathematical inaccuracies.\\n* Inadequate description of Hierarchical-CIFAR, motivation, and evaluation.\", \"description_of_model\": \"------------------------------\\nThe manuscripts describe the presented approach as metric learning but make no use of a distance function in the different spaces they map to. Instead, the manuscript defines an inner product over embeddings to compute similarities.\\n\\nThe manuscript describe their \\u201cself-attention operation\\u201d as a dynamic set-to-set operation. While the usual definition of self-attention is permutation invariant, the definition presented here is not and thus cannot be accurately described as a mapping between sets. Specifically, in the usual presentation of self-attention the only sharing of information between different set elements is during the outer product of the key and query vectors. The use of a BLSTM between \\u201cneighboring elements\\u201d of the set of prototype vectors violates this assumption and induces a lack of permutation invariance. This makes the method sensitive to permutations of classes, which does not make sense for predicting unordered classes.\\n\\nIn sections 2.2.3 and 2.2.4 the material is presented twice but slightly differently. For example, the definition of $b_k$ inline before equation 3 differs from equation 6 later in the text.\\n\\nEquation 7 is incorrect and should not exclude the current class from the denominator.\\n\\nThe description of how to classify new points after equation 6 is poorly explained. The description of what happens when $h_V$ is a \\u201cBLTSM\\u201d [sic (should be BLSTM)] is noninformative.\\n\\nThe manuscript describes two dimension sizes $H$ and $M$ but the definition of $attn$ requires that $H = M$.\", \"description_of_new_dataset_and_evaluations\": \"------------------------------------------------------------\\nOne of the stated contributions of the manuscript is a methodology to build harder few-shot learning datasets. Section 4.2 is the only place in the text that appears to address this point, but is unclear where either new finer-grained or coarser-grained labels are coming from (new manual annotation or otherwise). The manuscript \\u201cleave[s] out the detail of its construction for simplicity\\u201d, but it is unclear what is being done here in the first place.\\n\\nThe manuscript does not detail tuning competing methods on the new dataset and so it is unclear whether it is a fair comparison.\\n\\nThe manuscript presents evaluations without any discussion of differences in performance between datasets or the 1-shot/5-shot settings. For example, their method is significantly better on CUB on 1-shot but not so much on 5-shot, on the other hand it is not significantly better on 1-shot for H-CIFAR and CIFAR-HS but then becomes better than the rest with 5-shot.\\n\\nAdditional comments/corrections\\n---------------------------------------------\\nThere were numerous typos and grammatical errors that were present in the manuscript that did not directly impact this evaluation but should be fixed in the future.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The authors propose a new neural network model, called as Dissimilarity Network, to improve the few-shot learning accuracy.\\nOverall the idea is well motivated that by emphasizing the difference among classes, the model can achieve more accurate predictions for classes where only limited data points are available for training.\\nHowever, the paper is not quite well written.\\nFirstly, much of the work is built upon previous work including attention mechanisms, episodic training for few-shot learning. Such components are the core of this work because the attention mechanisms implement the class-awareness, and the episodic training facilitates the LSTM structure. Yet these are not well explained and not much context is provided, thus making the paper hard to follow.\\nSecondly, some terms are fairly overloaded, or not clearly defined. For example, the \\u201cprior\\u201d as mentioned in both the abstract and the introduction doesn\\u2019t refer to the commonly interpreted term as in the Bayesian settings, but rather as a hand-waiving term to indicate the model design. Also, the terms, \\u201cscore\\u201d, \\u201cmetric\\u201d, \\u201cdissimilarity\\u201d are mentioned in the paper but the paper is not really learning the metric, to my understanding. Thus the details of the paper is quite hard to grasp. \\nLastly, the idea of designing the global embedding and the task aware embedding is interesting but shouldn\\u2019t really be restricted to few-shot learning. It would be interesting to test the idea on general classification tasks, for example in a simple cross validation settings.\\nThus I think the paper would be stronger if the above are addressed and it\\u2019s not ready for publishing yet in its current form.\", \"below_are_some_more_detailed_comments\": \"1)\\tIn the abstract, the \\u201cnewly introduced dataset H-CIFAR\\u201d is not precise to me; my understanding is that the paper proposes such an experiment design for testing how well a classifier can predict the labels with hierarchy. The current writing refers to that the authors comprises a completely new dataset with new labels.\\n2)\\tIn the last sentence of the second paragraph in Introduction, the question is asked \\u201cwhat prior\\u201d should be reasonable. Since the authors didn\\u2019t really add any priors in a Bayesian settings but rather designed an architecture, I suggest to reword something like \\u201chow to explicitly encode hierarchies into the model structure\\u201d.\\n3)\\tIn Section 2.1, some more description for \\u201cepisodic training\\u201d would be nice: why should it be used? How is it used and why it makes sense in the few-shot learning context?\\n4)\\tIn Section 2.2, it would be nice to add the mathematical definition of \\u201cprototype\\u201d.\\n5)\\tIn Section 2.2.1, it would be nice to define \\u201cH\\u201d.\\n6)\\tIn Section 2.2.2, is M required to be fixed given it\\u2019s episodic training? Also it would be nice to add more details about the attention mechanism.\\n7)\\tIn the result section, it would be nice to discuss when the proposed method is doing better than other methods, for example RelationNet, as well as when it\\u2019s worse since different datasets show different results.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"In this paper the authors propose a metric based model for few-shot learning. The goal of the proposed technique is to incorporate a prior that highlight better the dissimilarity between closely related class prototype. Thus, the proposed paper is related to prototypical neural network (use of prototype to represent a class) but differ from it by using inner product scoring as a similarity measure instead of the use of euclidean distance. There is also close similarity between the proposed method and matching network\\n\\n\\noverall, the paper does not highlight the novelty of their proposed method especially prototypical network and matching network. Thus, the related work session is so general and does not tackle the close models in details. The experiments do not provide convincing evidence of the correctness of the proposed approach.\\n\\nSeveral parts are unclear/incomprehensible:\\n(1) The Introduction is confusing and does not demonstrate the problem that the paper is trying to solve. Specifically, the described intuition (Mill\\u2019s method of difference) is not convincing \\n(2) the first sentence of the section \\u201cOur work.\\u201d (page 1) is long and unclear \\u2026 \\u201cIn this paper, we propose a model that focuses on the differences in the support set of closely related classes in assigning the class label to a new instance in the novel task.\\u201d\\n(3) the use of the two level of embedding is confusing and not clear. Figure 1 is also confusion and not clear.\", \"the_correctness_of_the_proposed_approach_is_not_proved_by_the_conducted_experiment_and_does_not_provide_convincing_and_fair_comparison_with_soa_techniques\": \"(1) The experiments do not provide the details of the used architecture compared to your baseline. (how many layers are used in both embedding systems)\\n(2) In Table 1 you are using the results reported by Chen et al. (2019) did you use his framework (Resnet or 4 layers CNN)\\n\\n**Minor comments** \\nThe definition of the embedding function f = (f_g o f_f) (in line 1 page 5) is not consistent with the domain of each function f_g is defined on R^H x R^H.\\n\\nK is not defined (last line page 3)\"}" ] }

[ "Zhouxing Shi", "Huan Zhang", "Kai-Wei Chang", "Minlie Huang", "Cho-Jui Hsieh" ]

[ "Robustness", "Verification", "Transformers" ]

{ "note_id": [ "k2F6zynkK6", "rJeXo8rYjH", "BkeqhHSKsr", "SkeQbNBKir", "r1xFcQBYjH", "ryxVD3SG5B", "r1xpvEcpKH", "S1elFelaKS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798732048, 1573635738896, 1573635505940, 1573635067153, 1573634960866, 1572129883535, 1571820644969, 1571778679972 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1770/Authors" ], [ "ICLR.cc/2020/Conference/Paper1770/Authors" ], [ "ICLR.cc/2020/Conference/Paper1770/Authors" ], [ "ICLR.cc/2020/Conference/Paper1770/Authors" ], [ "ICLR.cc/2020/Conference/Paper1770/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1770/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1770/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"A robustness verification method for transformers is presented. While robustness verification has previously been attempted for other types of neural networks, this is the first method for transformers.\\n\\nReviewers are generally happy with the work done, but there were complaints about not comparing with and citing previous work, and only analyzing a simple one-layer version of transformers. The authors convincingly respond to these complaints.\\n\\nI think that the paper can be accepted, given that the reviewers' complaints have been addressed and the paper seems to be sufficiently novel and have practical importance for understanding transformers.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Discussions on comparisons to SOTA methods and our novelty\", \"comment\": \"We thank the reviewer for the constructive feedback. We address point 1, 2, and 4 in \\u201cweakness bullets\\u201d below (we have addressed point 3 in another reply).\\n\\n(1) Regarding comparison with previous SOTA methods, and time cost comparison:\\n\\nBecause we propose the first algorithm for verifying Transformers, no previous work is directly applicable for comparison. Previous methods were specifically designed for simpler neural networks such as feed-forward networks or RNNs. They cannot be extended to Transformer in a straightforward manner, except for naive IBP (Gowal et al., 2018). We have adapted IBP to verifying Transformers as a baseline in the paper. MILP can hardly be extended to Transformers (you may refer to our response to your question regarding MILP below).\\n\\nWe have already included time costs for fully forward, fully backward and our \\u201cbackward & forward\\u201d methods (Table 2 in the initial draft and Table 3 in the revised one). We think it is unnecessary to compare time costs with IBP, because IBP cannot produce non-vacuous bounds for Transformers at all (lower bounds computed by IBP are extremely small; see Table 1).\\n\\n(2) Regarding citations and MILP:\\n\\nWe thank the reviewer for raising the concern of listing previous verification works. In our revised paper, we have improved the introduction and related work sections and included many related papers that we missed, and also fixed some inaccurate citations.\\n\\nWe cannot compare to MILP based methods (Tjeng et al., 2019; Dutta et al., 2018), as they mainly focus on ReLU non-linearity. It is non-trivial to extend MILP based method to more complicated functions like self-attention, as the MILP formulation crucially depends on the property of ReLU (or other piecewise linear) neurons: ReLU neurons have an \\u201cactive\\u201d state that passes the input, and an \\u201cinactive\\u201d state that outputs a constant 0. This can be formulated as a binary variable in MILP (see Eq. (6) in Tjeng et al., 2019), but cannot be easily extended to self-attention layers with complex cross-nonlinearity and non-piecewise linear functions (dot product, weighted sum, softmax, tanh, etc).\\n\\nAdditionally, since MILP searches the optimal solution for the non-convex verification problem (Eq. 1), they are significantly slower. Weng et al., 2018 showed that similar methods can be over 10,000 times slower than backward bounding; Salman et al., 2019 had to use a cluster with 1000 CPU nodes to conduct MILP verification for simple CNN networks. So we believe MILP cannot scale to the transformer verification problem at all. Our method efficiently finds a lower bound instead, similar to CROWN (Zhang et al., 2018) and other efficient verifiers.\\n\\n(3) Regarding our novelty:\\n\\nOur work is the first work on verifying Transformers. Previous methods were specifically designed for simple architectures such as feed-forward networks and RNNs, and they cannot be directly applied to Transformers with a self-attention architecture.\\n\\nImportantly, verifying Transformers is not just about dealing with different nonlinearities by directly using methods from previous work. It is significantly more challenging in the following aspects:\\n\\n1. Cross-nonlinearity: cross-nonlinearity appears in nonlinear functions involving two variables under perturbation. Deriving the linear relaxations for such functions is different and more challenging than that for univariate functions. The gradient descent based approach for RNN in (Ko et al., 2019) is inefficient (hundreds of iterations needed) and poses a computational challenge for the much more complex transformer verification problem. In contrast, we derive a novel closed-form linear relaxation in $O(1)$ complexity.\\n\\n2. Cross-position dependency: this is special to self-attention, and there is no such dependency in feed-forward networks or RNNs verified by previous works. Resolving the cross-position dependency is critical for Transformers, otherwise the algorithm would be too slow. As the ablation study (Table 3 in revised paper) shows, if we directly extend the fully backward bounding framework (CROWN) to Transformers, it is much slower than our \\u201cbackward & forward\\u201d algorithm which resolves the cross-position dependency. Our novel \\u201cbackward & forward\\u201d algorithm reduces the time complexity by a factor of $O(n)$ ($n$ is the input length) compared to the fully backward algorithm, and still produces comparably tight bounds (much tighter than those by the fully forward algorithm).\\n\\n3. Additionally, we have demonstrated an application beyond robustness verification using our techniques -- identifying important words (Table 4 in our revised paper). Thus, we believe that our proposed algorithm has enough novelty in both methodology and empirical study.\\n\\n\\nWe thank the reviewer again for the helpful comments, and we have addressed all the \\u201cweakness bullets\\u201d raised by the reviewer (response for point 3 in another reply below). We hope the reviewer can re-evaluate our paper based on our response.\"}", "{\"title\": \"We have results for both 1-layer and 2-layer Transformers and we have further added new results on 3-layer Transformers.\", \"comment\": \"We thank the reviewer for the constructive feedback. We address the concern regarding multi-layer Transformers below:\\n\\nOur initial submission included results for multi-layer transformers. As you can see in Table 1, a column called \\u201cN\\u201d indicates the number of layers. We had experiments for both 1-layer and 2-layer Transformers. We have also further added experiments for 3-layer Transformers in the revised paper.\", \"each_transformer_layer_actually_contains_many_nonlinear_operations\": \"dot product operations, softmax, weighted summation, and also element-wise activation functions in the feed-forward layers after self-attention. Especially, bivariate nonlinearity in dot product and weighted summation are much more complicated than nonlinearities in a MLP/CNN/RNN layer (e.g. a MLP layer is just a linear operation plus an element-wise activation function). Therefore, we believe that even a *single* Transformer layer already has an amount of nonlinear operations comparable to those in many MLP/CNN/RNN layers. We believe that the provided up to 3-layer Transformer results in Table 1 should be sufficient to demonstrate the ability of our algorithm in the multi-layer setting.\"}", "{\"title\": \"We have extended our method to multiple perturbed positions and provided additional experiments. We also fixed the table on identifying important words and added a diagram for the forward and backward processes.\", \"comment\": \"We thank the reviewer for correctly identifying the main contributions of our paper and providing the very constructive feedback. We have revised our paper accordingly, and we provide answers below:\\n\\n(1) Regarding identified important words:\\n\\nWe believe the presentation of the table caused a misunderstanding. In our initial submission, listed words were most important (or least important) words identified by different methods, and the red/bold words were *improperly* selected words by each method (less red words are better, and our method has the least red words). We have updated the table (Table 4 in our new revision) and captions for a better illustration, as detailed below.\\n\\nIn our new table, boldfaced words are considered to have strong sentiment polarities. For the rows of \\u201cmost important\\u201d words, baselines improperly select more words with no sentiment polarity (words that are not boldfaced in the revised table). E.g., for the gradient method, \\u201cdinner\\u201d and \\u201cfood\\u201d are improperly selected, while our method chooses only one word (\\u201clevel\\u201d) improperly. We also have quantitative results evaluated on more examples (column \\u201cScore (SST)\\u201d in the table) showing that the most important words identified by our method are indeed more important than those by baselines, and the least important words identified by our method are indeed less important.\\n\\nWe hope it is now clear that our proposed method works better on identifying important words. You may also refer to our response to reviewer #1, who had the same confusion.\\n\\n(2) Regarding perturbing multiple positions:\\n\\nOur algorithm can be easily generalized to perturbations on multiple positions, and we have revised our methodology section for this generalization. We have also added experiments about perturbing two positions in Table 2 of the revised version, and they are consistent with previous conclusions. To support multi-position perturbation, we mainly need to modify the bound propagation to the input layer. We change from representing the bounds as linear functions of the one perturbed position ${\\\\rm x}^{(r)}$ to linear functions of multiple perturbed positions ${\\\\rm x}^{(r_1)}, {\\\\rm x}^{(r_2)}, \\\\cdots, {\\\\rm x}^{(r_t)}$ ($t$ is the number of perturbed positions), and also accordingly modify the part of taking global bounds (Eq. (3), (4)) and reusing previous linear bounds (paragraph \\u201cbackward process to self-attention layers\\u201d). Please refer to our revised paper for more details. However, the number of perturbed positions should still be small, since imperceptible adversarial examples may not have many perturbed positions (Gao et al., 2018; Ko et al., 2019).\\n\\n(3) Regarding the scope of tasks:\\n\\nAs this is the first paper for verifying Transformers, we focus on developing an algorithmic framework for verification. Existing works on robustness verification also commonly focus on classification (Weng et al., 2018; Zhang et al., 2018; Ko et al., 2019), since the robustness of classification tasks has been well-defined. As long as the specification of robustness can be defined as a special case of Eq. (1), our proposed method can be applied to any model consisting of Transformer encoder/decoder in other complex NLP tasks, such as machine translation, question answering, parsing, or even tasks beyond languages. For the sake of simplicity and due to space limitation, we leave empirical analysis of these complex applications as future work, since in many cases there is even no well-accepted definition of robustness, unlike in classification tasks.\\n\\n(4) Regarding a diagram for the forward and backward processes:\\n\\nThis is a great suggestion. We have provided a diagram in Appendix A to help readers understand the difference between forward, backward, and our novel backward&forward bound propagation processes.\\n\\nLastly, thank you for pointing out the typos and we have fixed them. We improved writing of our paper in our revision, and we will continue working on polishing it. \\n\\n\\n\\nWe greatly appreciate the helpful comments by the reviewer, and we hope the reviewer has a better understanding of our paper now. Feel free to let us know if there is anything still unclear, and we hope the reviewer can re-evaluate our paper based on our response.\"}", "{\"title\": \"We have improved the table on important words in our paper; sorry for the misleading presentation\", \"comment\": \"We thank the reviewer for the encouraging feedback! We believe the presentation of this table caused a misunderstanding. We have improved Table 3 (which becomes Table 4 in our revised paper) for a better presentation of our results. Our proposed method outperforms other baselines.\\n\\nPreviously, we presented two sets of results as indicated by the \\u201ctype\\u201d in the first column. The first part (\\u201cmost\\u201d) shows words that are the MOST salient while the second part (\\u201cleast\\u201d) shows words that are the LEAST salient. Therefore, words such as \\u201ca\\u201d, \\u201cthe\\u201d, \\u201cour\\u201d are indeed the LEAST important words identified by our algorithm as expected. We have updated the table and the captions to make them more clear: we have separated the list of \\u201cmost important\\u201d words and \\u201cleast important\\u201d words in different rows, and used boldfaced words to indicate words with strong sentiment polarities; boldfaced words should appear as \\u201cmost important\\u201d words rather than \\u201cleast important\\u201d words. Our proposed method clearly performs the best.\\n\\nWe hope our response has resolved your question. Please let us know if you have any further questions regarding our paper, and we will appreciate it if you can consider increasing the rating.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Contributions:\\nThis paper develops an algorithm for verifying the robustness of transformers with self-attention layers when the inputs for one input word embedding are perturbed. Unlike previous work the present work can deal with cross nonlinearity and cross position dependency and the lower bounds derived in the paper are much tighter than the Interval Boundary Propagation (IBP) method which uses backward propagation. The core contribution is expounded by developing bounds for multiplication (xy) and division (x/y) and using this to compute tight bounds on self-attention layer computations. Further by introducing a forward process then combining it with a backward process, they can substantially reduce computation time as compared to the IBP method (in experiments demonstrated to be over an order of magnitude).\", \"experimental_results\": \"Very reasonable and well thought out experiments demonstrate that a) the lower bounds produced by the method are very tight (much better than IBP for instance), and the backward&forward approach is an order of magnitude faster than the fully backward approach though they are both equally tight (the fully forward method is much looser by contrast)\\n\\nThe experiments on using the bound to detect the word whose perturbation will cause biggest impact (ie the most important word for explaining the decision) is much less convincing. In particular in Table 3, the second example shows that the method identifiews \\\"a\\\", \\\"the\\\" \\\"our\\\" etc, which at least do not appear to be most salient to this reviewer. [Question to authors: perhaps I misunderstood something here -- if so please explain and help me understand it better]. \\n\\nEvaluation / Suggestions. \\nOverall I liked the paper quite a bit. I may not have fully understood something about table 3 and would welcome some explanation about why the results shown there indicate that their method works better int he second example.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\nThis paper builds upon the CROWN framework (Zhang et al 2018) to provide robustness verification for transformers. The CROWN framework is based upon the idea of propagating linear bounds and has been applied to architectures like MLP, CNNs and RNNs. However, in Transformers, the presence of cross-nonlinearities and cross-position dependencies makes the backward propagation of bounds in CROWN computationally intensive. A major contribution of this paper is to use forward propagation of bounds in self attention layers along with the usual back-propagation of bounds in all other layers. The proposed method provides overall reduction in computational complexity by a factor of O(n). Although the fully forward propagation leads to loose bounds, the mixed approach (forward-backward) presented in this work provides bounds which are as tight as fully backward method.\", \"strength\": \"Use of forward propagation to reduce computational complexity is non-trivial\", \"strong_results_on_two_text_classification_datasets\": \"Lower bounds obtained are significantly tighter than IBP\\nThe proposed method is an order of magnitude faster than fully backward propagation, while still maintaining the bounds tight.\", \"weakness\": \"Experiments only cover the task of text classification. Experiments on other tasks utilizing transformers would have made the results stronger.\\nThe paper makes the simplifying assumption that only a single position of an input sentence will be perturbed. They claim that generalizing to multiple positions is easy in their setup but that is not supported. The paper needs to declare this assumption early on in the paper (abstract and intro). As far as I could tell, even during the experiments they perturb a single word at a time.\\n\\nThe paper is more technical than insightful. I am not at all convinced from a practitioners viewpoint that such bounds are useful. However, given the hotness of this topic, someone or the other got to work out the details. If the math is correct, then this paper can be it. \\n\\nThe presentation requires improvement. Some parts, example, the Discussion section cannot be understood.\", \"questions\": \"The set-up in the paper assumes only one position in the input sequence is perturbed for simplicity. Does the analysis remain the same when multiple positions are perturbed?\", \"suggestions\": \"A diagram to describe the forward and backward process would significantly improve the understanding of the reader.\\n\\nIn Table 3, I am surprised that none of the sentiment bearing words were selected as the top-word by any of the methods. Among the \\u2018best\\u2019 words, the words chosen by their method does not seem better than those selected by the grad method.\", \"several_typos_in_the_paper\": \"spelling mistakes in \\u201cobtain a safty guarantee\\u201d, poor sentence construction in \\u201cand independent on embedding distributions\\u201d, subject verb disagreement in \\u201cUpper bounds are discrete and rely on the distribution of words in the embedding space and thus it cannot well verify\\u201d.\\n\\nI have not verified the math to see if they indeed compute a lower bound.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Overview:\\n\\nThis paper is dedicated to developing robustness verification techniques. They claim their methods can deal with verification problems for Transformers which includes cross-nonlinearity and cross-position dependency. The paper solves these key challenges which are not traceable for previous methods. Moreover, the author demonstrates their certified robustness bounds are significantly higher than those by naive Interval Bound Propagation. They also point out the practice meaning through sentiment analysis.\", \"strength_bullets\": \"1. It is an interesting design that combines the backward process with a forward process. The author also conducts an ablation experiment to show its advantages both in the bound estimation and computation time cost.\\n2. The derivation is very solid and detailed. The author not only gives the certified bounds but also further analyzes whether the bounds are reasonable.\", \"weakness_bullets\": \"1. For comparison experiments, especially in Table 2, the author doesn't compare with other previous state-of-the-art methods. It just an ablation among fully forward, fully backward and combine two. To be more convincing, the author needs to post not only bounds but also time costs cross several methods on one or two datasets.\\n2. The experiment only uses a single-layer transformer, I don't think it contains much more nonlinear operations then MLP, CNN or RNN in the previous work. The author needs to add experiments of multi-layer transformers. \\n3. The author doesn't list all famous previous verification framework, like MILP (EVALUATING ROBUSTNESS OF NEURAL NETWORKS WITH MIXED INTEGER PROGRAMMING), even doesn't give a cite. MILP is a powerful verification technique that can deal with cross-position dependency, which can be potentially applied to transformers. The author needs to compare these methods to prove the advantages of the paper's approach.\\n4. The novelty of the main contribution of the paper is arguable. Although it is the first one who does robustness verification on transformers, the linear relaxation is similar to previous work just deal with different nonlinearity. The author may provide more detail and explanation about the creative modification in relaxation or other parts of the proposed verification framework.\", \"recommendation\": \"For lack of necessary experiments and limit novelty, even if I like part of the approach design, this is a weak reject.\"}" ] }

[ "Yiding Jiang*", "Behnam Neyshabur*", "Hossein Mobahi", "Dilip Krishnan", "Samy Bengio" ]

[ "Generalization", "correlation", "experiments" ]

{ "note_id": [ "8EatWJ_2_U", "HkeGTU-KoH", "ryeab8ZKjS", "ByxTYS-KsB", "S1xtJSZtjH", "HJxvboKicr", "ByxTlM6e5H", "SJewlC1jFr", "H1eyz2M2DH", "rJxvnPrqwS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment", "comment" ], "note_created": [ 1576798732016, 1573619385636, 1573619205489, 1573619076897, 1573618913023, 1572735743191, 1572028916670, 1571646958684, 1569627143182, 1569507247180 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1767/Authors" ], [ "ICLR.cc/2020/Conference/Paper1767/Authors" ], [ "ICLR.cc/2020/Conference/Paper1767/Authors" ], [ "ICLR.cc/2020/Conference/Paper1767/Authors" ], [ "ICLR.cc/2020/Conference/Paper1767/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1767/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1767/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1767/Authors" ], [ "~Ilya_Tolstikhin1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper provides a valuable survey, summary, and empirical comparison of many generalization quantities from throughout the literature. It is comprehensive, thorough, and will be useful to a variety of researchers (both theoretical and applied).\", \"title\": \"Paper Decision\"}", "{\"title\": \"Author Response to Review #3\", \"comment\": \"Thank you for your thoughtful feedback. We have added many more experiments and addressed all of your concerns. We hope that after looking at the revision and reading our response, you would consider increasing your score.\", \"we_will_next_address_your_concerns_in_details_below\": \"1) We have significantly improved the strength of our empirical results in the revision. More specifically, we repeated all experiments for Street View House Numbers (SVHN) dataset (table 6). Furthermore, we ran all experiments on CIFAR-10 5 times and calculated the standard deviation over different runs. Both these results are indeed consistent with the observations we made on CIFAR-10. Note that this means training over 10K networks! Please see the revision for details. Even though we believe current experimental results are more than enough to verify the conclusions in the paper, we are currently running additional experiments on CIFAR100 and STL10 datasets to make the experiments even more comprehensive. Given the scale of the experiments, these results would be available for the final version.\\n\\n 2) When studying generalization, we follow the common practice in the field which mainly focuses on over-parameterized regime where the model capacity is high enough to fit the training data. In this regime, the test error and generalization gap become the same. We believe this regime is more interesting as in the under-parametrized regime the test error can be usually improved by using better optimization algorithms and hence generalization is less mysterious. We agree with you that observations in over-parameterized regime might not generalize to the under-parametrized regime. However, in order to address your concern about test error vs generalization, we included the results of CIFAR-10 at cross-entropy=0.1 (table 7) in the revision where the training errors are much higher and the observations are largely consistent with cross-entropy=0.01, indicating that we are indeed capturing the generalization gap rather than test accuracy.\\n\\n 3) We believe there might be a confusion here. What we stated is that only changing a single hyperparameter and only looking at correlation when changing one single hyperparameter can be misleading. Instead, we do controlled experiments on seven hyperparameters, meaning that we change them one by one (and each time, we change one, the other six hyperparameters are kept fixed). The expected performance in the controlled experiments are much better indicator compared to only changing one hyperparameters or changing all of them together. You also asked why we didn\\u2019t sample all hyperparameters together. We actually do report that as well in the table column called \\u201coverall $\\\\tau$\\u201d. Please see section 2.3 for discussions on why the controlled experiments are preferred to overall $\\\\tau$. Finally, we added even further analysis using conditional mutual information in the appendix B that is consistent with the rank correlation results reported in the main paper. \\n\\n 4) Not all these measures can be converted to a bound. Furthermore, a tighter bound does not necessarily correlate better with generalization. That is mainly because all current bounds are very loose so their tightness relative to each other is not informative. What we try to study is whether these measures can predict the observed generalization rather than bounding it. We chose convolutional neural networks because they are heavily used in both research and in industry. We agree with you that these results might not generalize to other types of networks such as recurrent neural networks and we have made this clear on the conclusion of the original submission. However, we believe the protocols we establish in this work will be useful for future research on other types of networks.\\n\\n 5) Canonical order is not a single measure but rather a number specific to the changes made to a single hyperparameter, which means that there are 7 canonical measures (please see Section 4 for definition). Hence, each canonical measure is only predictive when we change its corresponding hyperparameter and is uncorrelated with generalization when we change any other hyperparameter. Therefore, each of these measures perform poorly overall.\"}", "{\"title\": \"Author Response to Review #1\", \"comment\": \"Thank you for your encouraging review! We next address your questions:\\n\\n 1) The stochasticity due to initialization and optimization: We have added this discussion to the revision. We ran all experiments 5 times and calculated the standard deviation (table 8&9). The resulting standard deviation shows that our result is statistically significant. Please see the paper for more discussion.\\n\\n 2) All results are reported on CIFAR10: Have now repeated all experiments on Street View House Numbers (SVHN) dataset and added the result to the revision (table 6). This result is consistent with that of CIFAR10 which shows that our conclusions are not limited to CIFAR10. In order to improve the strength of the paper even further, we are currently running similar experiments on STL10 and CIFAR100 which will be added to the final version.\\n\\n 3) Making make P & Q have the same variance equation 43 in App. C: Similar to Neyshabur et al, we mainly do this since it makes the derivation simpler and the final bound more interpretable. That is however not the optimal choice. \\n\\n 4) Relationship between sharpness and grad noise: You are right that these quantities seem very related. We indeed believe that this relationship is very interesting and can be a subject of future study! \\n\\nFinally, thank you for the minor suggestions. We have incorporated all of them in the revision.\"}", "{\"title\": \"Author Response to Review #2\", \"comment\": \"Thank you for your encouraging and thoughtful comments! We added more details to section 6 and address your other concerns below:\\n\\n 1) Margin normalization: We agree that these measures should be normalized. However, we want to point out that there is an implicit normalization effect by the fact that we stop at a certain cross-entropy which is closely related to the margin when training error is zero or very small. Therefore, the margin values of all these networks are similar (but not the same). That being said, we have normalized all quantities that usually require normalization. If you look at the normalized versions of the measures, they do not behave significantly different than the unnormalized ones. The reason we do not Euclidean distance measure is that they usually shows up in bounds together with other norm-based terms and the right way to normalize it is not clear (again since the cross-entropy and margin values are almost the same, we don\\u2019t think this makes a big difference).\\n\\n 2) Including sum-of-frob/margin: Thanks for pointing this out. We agree the quantity normalized by margin is important and have included it in table 2, and added corresponding discussion in the first paragraph of page 8.\\n\\n 3) Negative correlation of spectral bound: We agree that it is not surprising that spectral bound would be extremely large and it is not surprising if it doesn\\u2019t correlate with generalization. However, we believe that the negative correlation is very surprising. Note that the negative correlation is constantly present in the controlled experiments when two model different only differ in learning rate, or batch size, or dropout, or optimization algorithm. Therefore, this negative correlation cannot be relevant to the model size since in all these cases, two models have the exact same architecture.\\n\\n 4) Sum-of-fro: We completely agree with you that the main term is fro/spec and we are surprised that it performed worse than sum-of-fro. This perhaps can be a subject of future study.\"}", "{\"title\": \"Revision Summary\", \"comment\": \"We thank all reviewers for their valuable feedback. We have updated the paper to include new experiments and revisions suggested.\\n 1. We repeated all the experiments in the paper on SVHN and show that the result is consistent with what we already observed on CIFAR-10 (Table 6).\\n 2. We repeated all the experiments on CIFAR-10 5 times and show that the observation is robust to stochasticity of initialization and optimization (mean in Table 8 and std in Table 9). \\n 3. We compute the same statistics for CIFAR-10 at cross-entropy=0.1 and the performance of the majority of the measures (most notably spectral and sharpness) are consistent with cross-entropy=0.01. This shows that we are indeed capturing the generalization gap rather than test error (Table 7).\\n 4. We have added more analysis in the appendix B using conditional mutual information. The results are consistent with the ones reported in the main paper.\\n 5. Improved writing including related work, notation etc and incorporated all the suggestions made by reviewer 1&2.\\nMerged related work and subsection into subsections of section 1\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper addresses the well-known generalization problem for deep networks: why do these models successfully generalize even though they contain more parameters than the size of the training set? Recent works on this problem have proposed various measures of model complexity (that is, different from the number of parameters) with the claim that they (at least partially) capture the \\\"true\\\" capacity of deep networks in order to answer this question. The authors report the results of a large-scale, systematic study of the empirical effect of these complexity measures on the observed generalization behavior of deep networks.\\n\\nThe empirical study of generalization behavior is inherently tricky, and the authors make a non-trivial contribution just by designing an experiment that can convincingly address this question. The authors give careful attention to various potential obstacles related to experimental design and statistical validity. \\n\\nThe results of the authors' experiments are highly valuable to the research community. I recommend to accept.\\n\\n*********************************************************************************\", \"technical_comments_for_the_authors\": \"-Please rewrite or elaborate on the first two paragraphs of section 6 for clarity.\\n\\n-It seems to make little sense to consider norm-based complexity measures that are not margin-normalized (or at least normalized in some other way). Please include a margin-normalized version of each such measure in Table 5; certain key quantities (e.g. \\\"frob-distance\\\") have no corresponding margin-normalized entry.\\n\\n-The quantity \\\"sum-of-frob/margin\\\" is certainly a quantity of interest and it should be included in table 2 and possibly merits some discussion.\\n\\n-I disagree somewhat with the discussion of the spectral bound in section 7. In particular, I would not agree that the poor performance of \\\"spec-orig\\\" is surprising, because it contains an astronomically-sized \\\"proof artifact\\\" (as you call it). Namely, in this bound and in similar norm-based bounds, the term which is the product of the matrices in the network is (heuristically) unnecessary; in particular it (heuristically) ought to be able to be replaced with the network's \\\"average-case Lipschitz constant\\\" (this is one of the issues which Arora et al. 2018 attempted to address; unfortunately formalizing an adequate notion of \\\"average-case Lipschitz constant\\\" can be quite cumbersome). This superfluous product of matrix norms is of course highly correlated with model size, so it is not surprising that the quantity would have a sizable negative correlation with generalization. The quantity I would consider to be the \\\"main term\\\" is the quantity \\\"fro/spec\\\". Furthermore, the fact that this quantity (which appears in both Bartlett et al. and Arora et al.) performs somewhat worse than simply \\\"sum-of-fro\\\" is fairly interesting and possibly merits some discussion.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper aims at providing a better understanding of generalization for Deep Learning models. The idea is really interesting for the ML community as, despite their broad use, the astounding property of deep neural networks to generalize that well is still not well understood.\\nThe idea is not to show new theoretical bounds for generalization gaps but stress the results of an empirical study comparing the already existing measures. The authors choose 7 common hyperparameters related to optimization and analyze the correlation between the generalization gaps effectively observed and the ones predicted by different measures (VC dim, cross-entropy, canonical ordering \\u2026).\\nThe writing of the paper is clear and easily understandable. Besides, I believe that the study is relevant to ICLR conference.\\n\\nHowever, I believe the level of the paper is marginally below the threshold of acceptance and therefore would recommend to reject it.\\n\\nThe paper is solely empirical but I believe that the empirical section is a bit weak, or at least some important points remain unclear. If I appreciate the extent efforts made in trying to evaluate different measures of generalization gaps, I do not believe that the findings are conclusive enough.\\n\\n1) First, all this empirical result are based on one-dataset (CIFAR-10) only thus limiting the impact of the study. Indeed, a given measure might very correlate with generalization gap on this specific dataset but not on others.\\n\\n2) Specifically, we see that on this specific dataset, all training accuracies are already quite good (cf : Figure 1, distribution of the training losses). Consequently, authors are more correlating the chosen measures with the test error rather than with the generalization gaps. On other more complicated datasets where the training loss is higher, the VC dimension might consequently have way better results.\\nSimilarly, in Section 6, the authors say that the results \\u00abconfirm the widely known empirical observation that over-parametrization improves generalization in deep learning. \\u00bb In this specific case, no reference was given to support the claim. I would agree with the claim \\u00ab over-paramatrization improves test accuracy (reduces test error) \\u00bb but the link between over-parametrization and generalization is less clear. \\n\\n3) In Section 4, the authors say \\u00ab drawing conclusion from changing one or two hyper-parameters \\u00bb can be a pitfall as \\u00ab the hyper-parameter could be the true cause of both change in the measure and change in the generalization \\u00bb. I totally agree with the authors here. Consequently, I do not understand why the correlations were measured by only changing one hyper-parameter at a time instead of sampling randomly in Theta.\\n\\n4) It is still not clear to me how the authors explain why some measures are more correlated with generalization gaps than others. Are some bounds tighter than others ? This empirical study was only applied to convolutional neural networks and consequently one may wonder that for example the VC dim bounds computed in the specific case of neural networks are too loose. However, this measure could be efficient for type of models.\", \"i_would_like_the_authors_to_clear_the_following_points\": [\"How do you ensure that the empirical study clearly correlated measures predictions with generalization gaps and not simply with test errors (or accuracies) ? (point 2)\", \"Could you please also answer Point 4 ?\", \"Finally, how would you explain the fact that the canonical order performs so well compare to many other measures and that it is a really tough-to-beat baseline ?\"]}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"This work all starts with the observation, made by many, that deeper (>=2 layers) networks generalise differently to the simple models we used to use, and that the good old bias-variance trade-off doesn't really apply.\\n\\nThis paper is a natural successor to a long line of papers, most recently Zhang at el (best paper at ICLR 2017, though some like Tom Dietterich weren't impressed), Neyshabur et al (NIPS 2017) and Jiang et al (ICLR 2019). Whereas Jiang etal and Neyshabur etal had a few metrics, a few data sets and a few models, this paper instead has loads and loads of metrics and just one data set and one model class. You have, however, been clearly strongly influenced by the prior work and attempted to fix what you saw as their empirical shortcomings.\\n\\nI agree with the choice of stopping according to cross entropy. One thing not done in Section 5 is a discussion of\", \"variation\": \"choice of initialisation and the impact of stochasticity in the optimisation.\\n\\nAppendix C lists your copious metrics. I am not deeply embedded in the theory, though I understand the major concepts. I was impressed with the coverage and variations done. I expect Appendix C may raise copius matching arguments amongst the theory community, with some valid and some invalid complaints. For me, I wonder why you make P & Q have the same variance (see before equation 43 in App. C). You credit Neyshabur, but I'm left wondering.\\n\\nSo, the short version of my review is that I believe this paper is imcomplete, in that not enough different data sets were investigated. With this, I fear you could well be uncovering perculiarities pertinant to CIFAR-10 and the ease of getting near perfect training error on it. However, I am impressed with what you have done, and think you made a great starting point. So I say, publish and let the real battle begin. Let the theoreticians argue (about metrics) and the practicioners implement (with more data), and let's see what happens. You have laid a great ground work for folks, bulding on the earlier work. Perhaps Google can be encouraged to support this in that Jiang's paper is a precursor.\\n\\nInteresting observations in Appendix A.1, as are some of the discussions in Sections 7, 8, 9.\\n\\nGiven the best metric seems to be sharpness, which exceeds PAC-Bayesian, shouldn't you relate this to grad noise (metrics 61 and 62) which measure the flatness. Are these related? Note, also, a Bayesian always wants broad peaks for their parameter surfaces.\", \"minor_issues\": \"(1) some repeated words: \\\"an an\\\", \\\"the the\\\" (2) some strange grammar \\\"each canonical ordering that only be predictive for its own category\\\", \\\"VC-dimension bounds or and parameter counting\\\", \\\"possible to find to calculate\\\", \\\"in the Equation equation C.3 (3)\\\" Generalisation gap is poorly introduced. Its officially defined in footnote 4 but used way earlier! Put in the main text: its an important clarification. (4) Table 1 legend should mention the \\\"numerical results\\\" are for equations 2, 4 and 5. (5) Somewhere you need to mention that numbers in red in tables refer to equations in appendix C.\"}", "{\"comment\": \"Thank you so much for reading our paper and posting your feedback (on the second day after submission!) We are glad that you liked our paper!\\n\\nWe agree with your comment that in many applications we are interested in the test error rather than he generalization gap. As you pointed out yourself, we chose a setting where training error is almost zero so that these two coincide. This choice was indeed intentional to avoid choosing between test error and the generalization gap. A very interesting observation that we had during our experiments was that we realized if we do not control for the training accuracy, then by simply looking at the training loss one could easily predict which model will generalize better: while this is not always true, in our case we observed that at convergence models with higher loss generalize better. That made us choose training performance as a stopping criterion and look at the generalization gap. Another potential choice was to stop at a higher training error say 0.1 in which case the test error and generalization error only differ by 0.1 for all models, instead of 0.01. One issue with that choice is that there is not enough variation in the generalization of different models at that training accuracy as the models are not close to convergence. Another issue is that since the training error improves with a fast rate at each update around 0.1, it is almost impossible to stop the optimization at that exact stopping criterion. Also, one could argue that at training error 0.1, the main bottleneck to improve the test error is to fit training data better and even though this is important, it is not the focus of this study. All these difficulties led us to believe that the current setting is the best choice. We will add more discussions around this issue in next update. \\n\\nOnce again, thank you very much for reading the paper and catching the typos, and for the encouraging comment!\", \"title\": \"Thank you for your comment\"}", "{\"title\": \"Whether generalization gap is what we should be looking at\", \"comment\": \"Thank you for the paper. I enjoyed reading it. I wanted to know your opinion on the following point.\\n\\nYou define the generalization gap as the difference between the train and test accuracy and this gap plays the central role in the whole empirical study (i.e. in conclusions regarding which hyper-parameter values are beneficial and which are not). In short, the paper assumes that the smaller this gap, the better. I would like to argue that what is more important (and what the paper actually looks at) is the test accuracy, not the generalization gap. Notice that by design all the models in the study achieved very small training loss (high training accuracy). According to Figure 2 in appendix, majority of the models achieved training misclassification error of less than 1%. This means, generalization gap takes the form of $acc_train - acc_test$, where $acc_train$ takes values in [0.99, 1]. In that context saying \\\"small generalization gap\\\" is equivalent to saying \\\"large test accuracy\\\".\\n\\nI am pointing this out because in many other contexts these two statements (\\\"small generalization gap\\\" / \\\"large test accuracy\\\") may actually have two different meanings. Indeed, it seems that in most of the interesting applications DNNs manage to achieve the perfect training accuracy and thus the argument above holds (i.e. the two statements are equivalent). But I feel, what we are really looking for are the cases when we achieve a good test accuracy. For DNNs it just happens so that in those cases the generalization gap is very small. But small generalization gap obviously does not imply that the training succeeded: think of cases where the hyper-parameters are completely off and you end up with accuracy at random level both at training and test: the generalization gap is zero, but the training failed and there is no generalization. \\n\\nThat being said, your experimental design (looking only at models with high training accuracy) resolves this problem and indeed makes you focus on the good test performance. Nevertheless, I feel that making this point explicit (or even switching from \\\"generalization gap\\\" to \\\"test loss\\\" in your discussions) could be very useful in helping future follow up works (which I am sure will arrive sooner or later) to focus on important questions / numbers.\", \"small_typos\": \"\", \"page_5\": \"\\\"this this\\\", \\\"to choose ensure\\\"\", \"page_8\": \"sentence starting with \\\"We next look at...\\\".\\nPage 8 \\\"to find to calculate\\\"\"}" ] }

[ "Jiadong Lin", "Chuanbiao Song", "Kun He", "Liwei Wang", "John E. Hopcroft" ]

[ "adversarial examples", "adversarial attack", "transferability", "Nesterov accelerated gradient", "scale invariance" ]

{ "note_id": [ "G4WBcHLfr7", "HkepTE9csB", "Byl6Fbc9jB", "S1gH-At9oB", "rklpKcYcsS", "BygheOf79r", "SyxQje86tr", "HJl0R8y5KB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731987, 1573721285497, 1573720452582, 1573719549429, 1573718661511, 1572182003998, 1571803291413, 1571579605773 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1766/Authors" ], [ "ICLR.cc/2020/Conference/Paper1766/Authors" ], [ "ICLR.cc/2020/Conference/Paper1766/Authors" ], [ "ICLR.cc/2020/Conference/Paper1766/Authors" ], [ "ICLR.cc/2020/Conference/Paper1766/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1766/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1766/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"Under the optimization formulation of adversarial attack, this paper proposes two methods to improve the transferability of adversarial examples, namely Nesterov Iterative Fast Gradient Sign Method (NI-FGSM) and Scale-Invariant attack Method (SIM). NI-FGSM adapts Nesterov accelerated gradient into the iterative attacks to effectively look ahead and avoid the \\u201cmissing\\u201d of the global maximum, and SIM optimizes the adversarial perturbations over the scale copies of the input images so as to avoid \\u201coverfitting\\u201d on the white-box model being attacked and generate more transferable adversarial examples. Empirical results demonstrate the effectiveness of the proposed methods. The ideas are sensible, and the empirical studies were strengthened during rebuttal.\", \"title\": \"Paper Decision\"}", "{\"title\": \"RE: Review #3\", \"comment\": \"Thank you for your insightful comments. We have performed the corresponding revision based on your constructive suggestions.\\n\\nA1. Thank you for the valuable suggestion. We agree that the comparison of Nesterov Accelerated Gradient (NAG) with other momentum methods is important. Indeed, we have compared our NI-FGSM with MI-FGSM (Momentum Iterative Fast Gradient Sign Method), and in the revision, we make a more thorough comparison with various number of iterations (Figure 2) and provide more analysis in Sec. 4.6. \\nFor other momentum methods (e.g. Adam, momentum-SGD, etc), we also think it is a good direction to try for adversarial attacks, and the key is how to migrate the optimization method to the gradient-based iterative attack. We add the description in conclusion as our future work. Thank you so much.\\n\\nA2. Our scale operation is conducted after the input normalization. For example for inception_v3, the input normalization will scale the input image $x$ to $[-1, 1]$, and we will then apply scale operation so the range of input after SIM is $[-1/s, 1/s]$. The pixel value is in float type, so there is no need to downsample.\\n\\nA3. Thank you for the insightful comments. We perform additional analysis for the difference between NI-FGSM with MI-FGSM (Figure 2 in the revision). The adversarial examples are crafted on Inc-v3 with various number of iterations ranging from 4 to 16, and then transfer to attack Inc-v4 and IncRes-v2. As shown in Figure 2, NI-FGSM yields higher attack success rates than MI-FGSM with the same number of iterations. In another view, NI-FGSM needs fewer number of iterations to gain the same attack success rate of MI-FGSM. The results not only indicate that NI-FGSM has a better transferability, but also demonstrate that with the property of looking ahead, NI-FGSM can accelerate the generation of adversarial examples.\"}", "{\"title\": \"RE: Review #2\", \"comment\": \"Thank you for your insightful comments. We have performed the corresponding revisions based on your constructive suggestions.\\n\\nA1. Thanks for reminding us. Following your suggestion, we have updated in Sec. 3.1 and cited Dong et al. (2018) in the revision.\\n\\nA2. Thanks for your insightful and valuable suggestion. Yes, batch normalization may mitigate the impact of the scale change. If we consider a layer $t$ with a ReLu activation function $g(\\\\cdot)$, followed by a batch normalization layer $BN(\\\\cdot): z = BN(RL(Wx^t+b))$, where $x^t$ is the current layer\\u2019s input, the weight matrix $W$ (Convolution operator can also be regarded as a sparse linear operation $W$) and bias vector $b$ are the current layer\\u2019s parameters. \\nWhen at layer 0, we scale an input $x$ with factor $s$, we can get: \\n$z = BN(g(W(s \\\\cdot x)+b)) = BN(s \\\\cdot g[(Wx+b/s)] ).$\\nTherefore, for various scales on the input $x$, we have similar outputs with slightly different truncations by ReLu, and BN will shrink their difference and mitigate the impact of the scale change. \\nSimilarly, at the next layer after ReLu and BN, the differences will be shrunk again. After several Conv+ReLu+BN layers, the output will have some scale invariance for a range of the scales on the input. \\n\\nIn Figure 1, Inc-v3 and Inc-v4, Inc-v4 is deeper than Inc-v3, and shows a better scale invariance, which is consistent with the above explanation. For Res-101 and IncRes-v2, their scale ranges are smaller due to the residual connection among the connections, which is also consistent with the above explanation. Due to the time limit, we could not try more experiments to demonstrate this observation, but we are running the experiments, and will add in the final version. \\n\\nA3. In this paper, we consider to demonstrate that our method SI-NI is easy to combine with other orthogonal state-of-the-art methods to boost the transferability of adversarial examples. Following this spirit, we directly compare SI-NI-DIM with DIM (also TIM vs. SI-NI-TIM; TI-DIM vs. SI-NI-TI-DIM). Under such motivation, it is fair to conduct the experiments with the same number of iterations to show the improvement of the performance under the same parameters.\\n\\nFollowing your suggestion, we also did a comparison on the performance with the same number of gradient calculations. We increase the number of iterations to 50 for TIM, DIM and TI-DIM, denote as TIM (50), DIM (50) and TI-DIM (50), respectively. The results are as follows:\\n\\n1) For TIM, TIM (50) gets the worse performance than TIM (10).\\n2) For DIM, DIM (50) improves the transferability for normally trained models but reduces the transferability for adversarially trained models.\\n3) The case of TI-DIM is the same as that of DIM.\\n\\nOverall, as comparing the performance with the same number of gradient calculations, our method still achieves a considerable improvement on the transferability for normally trained models and adversarially trained models. \\n\\n---------------------------------------------------------------------------------------------------------------------------\\n| Attack | Inc-v3* | Inc-v4 | IncRes-v2 | Res-101 | Inc-v3$_{ens3}$ | Inc-v3$_{ens4}$ | IncRes-v2$_{ens}$ |\\n---------------------------------------------------------------------------------------------------------------------------\\n| TIM (10) | 100 | 47.8 | 42.8 | 39.5 | 24.0 | 21.4 | 12.9 |\\n| TIM (50) | 100 | 46.9 | 41.4 | 36.8 | 15.3 | 14.7 | 8.2 |\\n| SI-NI-TIM (10) | 100 | 77.2 | 75.8 | 66.5 | 51.8 | 45.9 | 33.5 |\\n| DIM (10) | 98.7 | 67.7 | 62.9 | 54.0 | 20.5 | 18.4 | 9.7 |\\n| DIM (50) | 99.1 | 77.2 | 71.9 | 61.7 | 13.1 | 13.8 | 6.1 |\\n| SI-NI-DIM (10) | 99.6 | 84.7 | 81.7 | 75.4 | 36.9 | 34.6 | 20.2 |\\n| TI-DIM (10) | 98.7 | 66.1 | 63.0 | 56.1 | 38.6 | 34.9 | 22.5 |\\n| TI-DIM (50) | 100 | 78.2 | 71.7 | 63.3 | 27.6 | 25.4 | 15.6 |\\n| SI-NI-TI-DIM (10) | 99.6 | 85.5 | 80.9 | 75.7 | 61.5 | 56.9 | 40.7 |\\n------------------------------------------------------------------------------------------\\n(Note: the adversarial examples are crafted for Inc-v3, which means it is white-box attack for Inc-v3 and black-box attack for other models)\\n\\nA4. Thank you for your suggestion. But we are afraid there is no efficient way to speed up the gradient calculation for scale-invariant attacks. \\n1)\\tFor the work of Dong et al. (2019), they make an assumption that the translation-invariant property is nearly held with very small translations. Based on this assumption, it is easy to get the gradient of the translated image \\n $T_{ij}(x)$ by translating the gradient of the original image $x$.\\n2)\\tBut in our scale-invariant method, as for one layer of conv+relu+batch-norm, \\n $z = BN(g(W(s \\\\cdot x)+b)) = BN(s \\\\cdot g[(Wx+b/s)] ).$\\nwhere $W$ is a sparse matrix representing the conv operator. The gradient of $s \\\\cdot x$ could not be calculated by scaling the gradient of $x$ with $s$.\"}", "{\"title\": \"RE: Review #1\", \"comment\": \"Thank you for your insightful comments. We have performed the corresponding revision based on your constructive suggestions.\\n\\nA1. We acknowledge that a solid theoretical analysis on why our attack methods can improve the transferability is very important. However, though many efficient adversarial attack methods have been proposed in the literature, so far there is little theoretical results on the transferability, and researchers usually provide some intuitive explanation or just provide some empirical evidence. We will try our best to answer your question as follows: \\n\\n(1) For NI-FGSM: Typical gradient-based iterative attacks (e.g., I-FGSM) greedily perturb the images in the direction of the sign of the gradient at each iteration, which usually falls into poor local maxima, and shows weak transferability than single-step attacks (e.g., FGSM). Dong et al. [1] show that adopting momentum into attacks can stabilize the update directions, which helps to escape from poor local maxima and improve the transferability. Compared to momentum, beyond stabilize the update directions, the anticipatory update of NAG gives previous accumulated gradient a correction that helps to effectively look ahead. Such looking ahead property of NAG can help us escape from poor local maxima easier and faster, resulting in the improvement on transferability.\\n\\nTo provide empirical evidence, we conduct additional experiments in Figure 2 in the revision. It shows that, using the same number of iterations, NI-FGSM yields higher attack success rates than MI-FGSM (Momentum Iterative Fast Gradient Sign Method), not only for the white-box setting (a), but even better for the black-box settings (b and c), demonstrating that NI-FGSM has a better transferability. \\n\\n(2) For SIM: Similarly to the generalization of models can be improved by feeding more training data, the transferability of adversarial examples can also be improved by attacking more models simultaneously [1, 2]. Essentially, SIM derives an ensemble of models to be attacked from the original model via the loss-preserving scale transformation. Such model argumentation will help improve the transferability of adversarial examples. \\n\\nIn the revision, we also add more explanation in Sec. 3.2.\\n\\nA2. Following your constructive suggestion, we have made more comprehensive comparisons with some recent classic methods in adversarial attacks, including FGSM, I-FGSM, PGD and C&W. The results are reported in Table 4 in the revision (also list in the following). Under the white-box setting, our methods achieve 100% attack success rate, which is as good as C&W, and is better than FGSM, I-FGSM, and PGD. Under the black-box setting, our methods significantly outperform all the four baseline methods.\\n\\n---------------------------------------------------------------------------------------------------------------------------------------------\\nAttack |Inc-v3 |Inc-v4| IncRes-v2| Res-101 | Inc-v3$_{ens3}$ | Inc-v3$_{ens4}$ | IncRes-v2$_{ens}$ | AVG|\\n--------------------------------------------------------------------------------------------------------------------------------------------\\nFGSM | 67.1 | 26.7 | 25.0 | 24.4 | 10.5 | 10.0 | 4.5 | 24.0 \\nI-FGSM | 99.9 | 20.7 | 18.5 | 15.3 | 3.6 | 5.8 | 2.9 | 23.8 \\nPGD | 99.5 | 17.3 | 15.1 | 13.1 | 6.1 | 5.6 | 3.1 | 20.9 \\nC&W | 100.0 | 18.4 | 16.2 | 14.3 | 3.8 | 4.7 | 2.7 | 22.9 \\nNI-FGSM | 100.0 | 52.6 | 51.4 | 41.0 | 12.9 | 12.8 | 6.4 | 39.6 \\nSI-NI-FGSM | 100.0 | 76.0 | 73.3 | 67.6 | 31.6 | 30.0 | 17.4 | 56.6\\n---------------------------------------------------------------------------------------------------------------------------------------------\\n\\nA3. In the revision, we evaluate the attack performance in Table 3 with three more advanced defense methods, which have shown to be robust on the ImageNet dataset, namely FD (Feature Distillation) [3], ComDefend [4], and RS (Randomized Smoothing) [5]. Our method SI-NI-TI-DIM achieves an average attack success rate of 90.3% on the six advanced defense methods, surpassing the state-of-the-art method TI-DIM [6] by a large margin of 14.7%. Thus we can claim the effectiveness of our methods. Thank you for the valuable suggestion.\\n\\n[1] Boosting adversarial attacks with momentum. CVPR 2018\\n[2] Delving into Transferable Adversarial Examples and Black-box Attacks. ICLR 2017\\n[3] Feature Distillation: DNN-Oriented JPEG Compression Against Adversarial Examples. CVPR2019\\n[4] ComDefend: An Efficient Image Compression Model to Defend Adversarial Examples. CVPR2019\\n[5] Certified Adversarial Robustness via Randomized Smoothing. ICML 2019\\n[6] Evading defenses to transferable adversarial examples by translation-invariant attacks. CVPR 2019\"}", "{\"title\": \"General Response\", \"comment\": \"We deeply appreciate all reviewers for the thorough comments and valuable suggestions, which definitely help the improvement of our paper.\\n\\nWe would like to briefly summarize our modification in the updated paper and provide specific response in the individual comment.\", \"our_main_modifications_are_as_follows\": [\"We have polished the overall writing.\", \"In Section 2.3, we introduce and discuss more advanced attack and defense methods.\", \"In Section 3.2, we add more explanation on why our NI-FGSM method can improve the transferability.\", \"In Section 4.3, we rewrite this subsection and add an experiment to further show the advantage of SI-NI-FGSM.\", \"In Section 4.5, we evaluate our methods on six most recent defense techniques to further demonstrate the effectiveness of our methods.\", \"In Section 4.6, we made further analysis to compare NI-FGSM and MI-FGSM, and also do comparison with several popular adversarial attacks (FGSM, I-FGSM, PGD and C&W).\", \"In Section 5, the conclusion section, we add one paragraph for our future work on NI-FGSM and SIM.\", \"By regarding the process of generating adversarial examples as an optimization problem, we propose two novel attack methods to improve the transferability of adversarial examples. We provide intuitive explanations as well as strong and extensive empirical support for the effectiveness of the proposed methods.We believe our attack methods can serve as strong baselines to assess the robustness of deep learning algorithm.\", \"We hope all our effort can make our paper more comprehensive and address most of your concerns. Thank you very much!\", \"Best regards,\", \"Authors\"]}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"In this paper, the authors proposed two methods of Nesterov Iterative Fast Gradient Sign Method (NI-FGSM) and Scale-Invariant attack Method (SIM) to improve the transferability of adversarial examples. Empirical results on ImageNet dataset demonstrate its effectiveness. In general, the paper is clearly written and easy to follow but I still have several concerns:\\n1.\\tAlthough the method is easy to understand, the authors are expected to clarify why the methods can improve the transferability. The authors are expected to make more theoretical analysis.\\n2.\\tThe authors are expected to make more comprehensive comparisons with the recent methods in adversarial attacks, e.g, PGD, and C&W even if some methods are designed for white-box attack. \\n3.\\tThe authors are expected to make more evaluations on the models with defense mechanism, and numerous important methods are missing. Without this, the authors cannot claim its effectiveness since only experiments on NIPS2017 is not enough.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"In this paper, the authors apply the Nesterov Accelerated Gradient method to the adversarial attack task and achieve better transferability of the adversarial examples. Furthermore, the authors introduce a scale transformation method to provide the augmentation on the model, which also boosts the transferability of the attack method. Experiments are carried out to verify the scale-invariant property and the Nesterov Accelerated Gradient method on both single and ensemble of models. All experiments turn out to be a positive support to the authors' claim.\\n\\nHowever, one small drawback of this paper is that the author does not claim any comparison between the Nesterov Accelerated Gradient Method and other momentum methods (e.g. Adam, momentum-SGD, etc). This experiment is somehow important since it shows the better transformability is obtained from 1) Nesterov Accelerated Gradient Method only, or 2) all momentum method, which is significant for further research.\\n\\nAlso, in the setting of the Scale-Invariant Transformation, the authors forget to address that what if the attacked network has an input normalization. Does it mean to downsample the value of each pixel in the input image? If so, is the equation $S_i(x) = x / 2^i$ better to be $S_i(x) = [x / 2^i]$ where $[]$ means casting to the nearest integer?\", \"one_more_question_of_this_work_is\": \"The Nesterov Accelerated Gradient method is known for its proveable fast descent property comparing to the traditional Gradient method. Do you observe any speed-up during your training?\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper studies how to generate transferable adversarial examples for black-box attacks. Two methods have been proposed, namely Nesterov Iterative Fast Gradient Sign Method (NI-FGSM) and Scale-Invariant attack Method (SIM). The first method adopts Nesterov optimizer instead of momentum optimizer to generate adversarial examples. And the second is a model-augmentation method to avoid \\\"overfitting\\\" of the adversarial examples. Experiments on ImageNet can prove the effectiveness of the proposed methods.\\n\\nOverall, this paper is well-written. The motivation of the proposed methods are generally clear although I have some questions. The experiments can generally prove the effectiveness.\", \"my_detailed_questions_about_this_paper_are\": \"1. The motivation in Section 3.1, which regards generating adversarial examples as training models, and transferability as generalizability, is first introduced in Dong et al. (2018). The authors should acknowledge and refer to the previous work to present the motivation.\\n2. It's not clear why deep neural networks have the scale-invariant property. Is it due to that a batch normalization layer is usually applied after the first conv layer to mitigate the effect of scale change?\\n3. It's not fair to directly compare DIM with SI-NI-DIM (also TIM vs. SI-NI-TIM; TI-DIM vs. SI-NI-TI-DIM), since SI-NI needs to calculate the gradient over 5 ensembles. It's better to compare the performance of two methods with the same number of gradient calculations.\\n4. Is there an efficient way of calculating the gradient for scale-invariant attacks like translation-invariant attacks in Dong et al. (2019)?\"}" ] }

[ "Hyojin Bahng", "Sanghyuk Chun", "Sangdoo Yun", "Jaegul Choo", "Seong Joon Oh" ]

[ "Generalization", "Bias", "Dataset bias" ]

{ "note_id": [ "XvS-dkKZyVx", "xZyM4u9x8d", "H1gotx12iH", "Bke_IxkhiS", "rJlz_k13or", "HklOfkk3or", "HJl8kkyhjH", "BJeTKL-siS", "B1gpPXAK5B", "ByewI_D15B", "S1g5ZEg6KB" ], "note_type": [ "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1593665606550, 1576798731959, 1573806210566, 1573806160304, 1573805929995, 1573805839995, 1573805789798, 1573750404661, 1572623205448, 1571940430714, 1571779585879 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Paper1765/Authors" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1765/Authors" ], [ "ICLR.cc/2020/Conference/Paper1765/Authors" ], [ "ICLR.cc/2020/Conference/Paper1765/Authors" ], [ "ICLR.cc/2020/Conference/Paper1765/Authors" ], [ "ICLR.cc/2020/Conference/Paper1765/Authors" ], [ "ICLR.cc/2020/Conference/Paper1765/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1765/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1765/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1765/AnonReviewer3" ] ], "structured_content_str": [ "{\"title\": \"To be presented at ICML 2020\", \"comment\": \"Dear Readers, Official Reviewers, and the Area Chair,\\n\\nThank you for your interest in this paper. Thanks to the constructive feedback from ICLR reviewers and the AC, we have further advanced the methodology and expanded the set of experiments over the past months. The new paper is accepted at ICML 2020 and will be presented in two weeks!\\n\\nPaper (ICML version): https://arxiv.org/abs/1910.02806\", \"code\": \"https://github.com/clovaai/rebias\\n\\nHighlights of changes (ICLR->ICML):\\n- Simplification of ReBias: (1) remove unnecessary normalization (CKA) and (2) match the inner and outer objective for the minimax optimization.\\n- Experiments on action recognition (Kinetics with 3D CNNs).\\n- Comparison against further baselines that appeared after the ICLR 2020 deadline (RUBi and LearnedMixin).\\n- Code is released.\\n\\nThanks again for letting us improve our research, and hope you enjoy the new paper & code.\\n\\nThanks,\\nReBias authors\"}", "{\"decision\": \"Reject\", \"comment\": \"This paper provides and analyzes an interesting approach to \\\"de-biasing\\\" a predictor from its training set. The work is valuable, however unfortunately just below the borderline for this year. I urge the authors to continue their investigations, for instance further addressing the reviewer comments below (some of which are marked as coming after the end of the feedback period).\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer3 (2/2)\", \"comment\": \"#4 How well do Bagnets alone perform on the benchmarks in Table 3? \\n \\n> Please see the table below.\\n\\nModels \\tBiased \\tUnbiased IN-A\\nResNet18 \\t93.3 \\t85.8 \\t30.5\\nBagNet18 \\t72.4 \\t58.6 \\t19.5\\nResNet18-BagNet18 \\t93.7 \\t88.4 \\t31.7\\n \\nModels \\tBiased \\tUnbiased IN-A\\nResNet50 \\t91.7 \\t78.3 \\t29.5\\nBagNet50 \\t73.0 \\t60.9 \\t21.4\\nResNet50-BagNet50 \\t88.7 \\t89.2 \\t31.3\\n \\nBagNets alone perform worse than the vanilla ResNets themselves as they are biased towards texture by design (i.e., small receptive field size). We have updated the results in the paper (Table 3).\\n \\n#5 How well do other methods do in these domains (Unbiased and IN-A)?\\n \\n> As requested by R3, we compare REBI\\u2019S against the data augmentation strategy using Stylized ImageNet [9].\\n\\nModels \\tBiased \\tUnbiased IN-A\\nResNet18 \\t 93.3 \\t85.8 \\t30.5\\nStylizedImageNet ResNet18 \\t92.5 \\t87.6 \\t29.7\\nREBI\\u2019S (ResNet18-BagNet18)\\t \\t93.7 \\t88.4 \\t31.7\\n \\nStylized ImageNet augmentation shows improvements in reducing texture bias (\\u201cunbiased\\u201d accuracy from 85.8 to 87.6), but it does not increase the generalizability to the challenging natural adversarial examples (ImageNet-A accuracy from 30.5 to 29.7). REBI\\u2019S improves upon the StylizedImageNet-trained model both in terms of the unbiased and ImageNet-A accuracies (1.2 pp and 2.0 pp, respectively). We have updated the results in Table 3.\\n\\n<References>\\n[1] https://arxiv.org/abs/1610.02413\\n[2] http://openaccess.thecvf.com/content_ECCV_2018/papers/Yingwei_Li_RESOUND_Towards_Action_ECCV_2018_paper.pdf\\n[3] https://arxiv.org/abs/1904.07911\\n[4] https://arxiv.org/abs/1712.00377\\n[5] https://arxiv.org/abs/1902.01007\\n[6] https://arxiv.org/abs/1907.07355\\n[7] https://arxiv.org/abs/1908.10763\\n[8] https://arxiv.org/abs/1909.03683 \\n[9] https://openreview.net/pdf?id=Bygh9j09KX \\n[10] https://arxiv.org/abs/1811.11155 \\n[11] https://arxiv.org/abs/1903.06946\\n[12] https://arxiv.org/abs/1812.02725\"}", "{\"title\": \"Response to Reviewer3 (1/2)\", \"comment\": \"We thank all the reviewers for their recognition of the task \\u201cinteresting and important\\u201d (R1), and finding the proposed method \\u201csound and interesting\\u201d (R2, R3). In particular, R3 has commented that \\u201cthe paper does a really good job of defining the problem setting\\u201d. Yet, we find some lack of detail and clarity in our paper could potentially have led to misunderstanding our work -- we have clarified them in reviewer-specific responses and have updated the paper accordingly.\", \"we_summarize_our_revisions_to_the_paper_as_follows\": \"\", \"section_2\": [\"Based on the helpful comments of R1 and R3, we have clarified that\", \"G may capture important signals as well as biases (Section 2.3, paragraph 3).\", \"G can be easily defined in a broad class of problems in machine learning applications (Section 2.3, paragraph 4).\", \"We have updated additional related works provided by R2 (Section 2.3, paragraph 4).\"], \"section_3\": [\"We have clarified that REBI\\u2019S does not suppress the signal captured by G due to its formulation as conditional independence (Section 3.1, Independence versus separation).\"], \"section_4\": [\"As requested by R3, we have added additional comparison on vanilla BagNets and Stylized ImageNet augmentation [1] on ImageNet (Table 3).\", \"There was a minor error in our evaluation code on ImageNet-A, and we have updated the correct values for every method (Table 3).\", \"We have fixed typos throughout the paper.\"], \"our_response_to_reviewer_3_is_as_follows\": \"#1 I'm not convinced that the proposed biased models (G) only use B for making predictions. \\n \\n> We agree with R3 that it is difficult to define G that only use B for making predictions (perfect precision, see Section 2.3). Nevertheless, G does not have to be perfect; please also see our response to R1. Our framework does not remove signals captured by G because REBI\\u2019S is built upon the \\u201cconditional independence\\u201d criterion (Section 3.1, \\u201cIndependence versus separation\\u201d paragraph) that f(x) is encouraged to be independent of g(x), **given** the target task Y. Unlike independence that requires f(x) to ignore g(x) altogether, conditional independence allows f(x) to still retain cues captured by g(x) if they highly correlate with Y. In short, our goal is to **discourage over-reliance** towards cues captured by g(x) and push the model to learn other features that are useful for prediction. In the revision, we have improved the description of the precision and recall conditions for G in Section 2.3: those conditions are not meant as hard requirements but as desiderata. Even if they are not precisely met, conditional independence will do the rest of the job.\\n\\n#2 The solution just pushes the problem of 'learning a model that only uses the true signal to make predictions' to 'learning a set of models that only use the noise to make predictions.' It's not clear why the latter is easier than the former. \\n \\n> As R3 suggested, there can be problems where directly using the true signal is easier than only using bias for prediction. If that is the case, then we would suggest to use the signals directly. Our work is concerned with the cases where i) the signal is highly entangled with the bias and there is no easy method to disentangle them yet, and ii) we have some evidence for the type of bias and the corresponding set of models G capturing it. We will give examples of such scenarios in #3.\\n\\n#3 Identify a broader class of problems for which G is given or can easily be defined.\\n\\n> As addressed in #1 (and in response to R1), G does not have to be perfect and we are provided certain evidence for bias types and corresponding set of representations G in many application scenarios [2-9]. Action recognition models have been reported to rely heavily on static cues without learning temporal cues [2,3]; though the actual bias may not precisely be _any static cue_, we can still regularize the 3D convolutional networks towards better generalization across static cue biases by defining G to be the set of 2D convolutional architectures. It has been argued that visual question answering (VQA) models, too, rely overly on language biases rather than the visual cues (e.g. without looking at the image, one knows the answer to \\u201cwhat color is the banana\\u201d is \\u201cyellow\\u201d) [4]. We can define G as the set of models looking at the language modality only. Entailment models are biased towards the presence of certain words (e.g. when there are many \\u201cnot\\u201ds, the sentence is \\u201ccontradictory\\u201d), rather than really understanding the underlying meaning of sentences [5,6]. We can design G to be the set of bag-of-words classifiers [7,8]. In the revised paper, we have supplemented the above examples in Section 2.3.\"}", "{\"title\": \"Response to Reviewer2\", \"comment\": \"We thank all the reviewers for their recognition of the task \\u201cinteresting and important\\u201d (R1), and finding the proposed method \\u201csound and interesting\\u201d (R2, R3). In particular, R3 has commented that \\u201cthe paper does a really good job of defining the problem setting\\u201d. Yet, we find some lack of detail and clarity in our paper could potentially have led to misunderstanding our work -- we have clarified them in reviewer-specific responses and have updated the paper accordingly.\", \"we_summarize_our_revisions_to_the_paper_as_follows\": \"\", \"section_2\": [\"Based on the helpful comments of R1 and R3, we have clarified that\", \"G may capture important signals as well as biases (Section 2.3, paragraph 3).\", \"G can be easily defined in a broad class of problems in machine learning applications (Section 2.3, paragraph 4).\", \"We have updated additional related works provided by R2 (Section 2.3, paragraph 4).\"], \"section_3\": [\"We have clarified that REBI\\u2019S does not suppress the signal captured by G due to its formulation as conditional independence (Section 3.1, Independence versus separation).\"], \"section_4\": [\"As requested by R3, we have added additional comparison on vanilla BagNets and Stylized ImageNet augmentation [1] on ImageNet (Table 3).\", \"There was a minor error in our evaluation code on ImageNet-A, and we have updated the correct values for every method (Table 3).\", \"We have fixed typos throughout the paper.\"], \"our_response_to_reviewer_2_is_as_follows\": \"#1 Most of the biases studied are synthetic and trivial.\\n \\n> Studying biases in models is difficult because it requires a dataset where bias is well-controlled (Section 4.1). We approach the difficulty by presenting two sets of experiments: Biased MNIST (synthetic but fully controlled biases) and ImageNet (realistic but less well-controlled biases). While the former may seem trivial, it allows us to prove that the novel algorithm REBI\\u2019S does work as intended on the simplest type of biases. With full control over the degree of bias, we can measure its performance on perfectly unbiased test data, which is infeasible to achieve in real-world datasets (i.e., all datasets are biased in their own manner [1]). The synthetic bias allows us to explicitly evaluate the bias-specific (color and texture) performances (Table 1). On ImageNet, we evaluate our method against realistic texture biases [2]. Experiments show that REBI\\u2019S does remedy biases towards texture and improves the generalization to the unbiased set and the ImageNet-A dataset (Table 3).\\n \\n#2 A few very related recent works were missing.\\n \\n> Thank you for mentioning the recent related works. We have included them in the revised paper. We believe extending our framework to the entailment task is an interesting future work. \\n \\n<References>\\n[1] http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.208.2314&rep=rep1&type=pdf\\n[2] https://openreview.net/pdf?id=Bygh9j09KX\"}", "{\"title\": \"Response to Reviewer1 (2/2)\", \"comment\": \"#2 In practice, their claim is texture and color are biases, yet they are still important cues for object recognition.\\n \\n> We do not claim texture and color to be biases in _all scenarios_. In Section 2.1, we have defined bias as \\u201ccues not essential for the recognition but correlated with the target Y\\u201d and the key property for a bias is that \\u201cintervening on [the bias] should not change [the target Y]\\u201d. In the biased MNIST experiments, we refer to color and texture as biases (B) because changing them do not change the semantics of the digits (Y). In the ImageNet experiments, we use the prior knowledge that image classifiers tend to be biased towards \\u201clocal patterns\\u201d such as color and texture [9], though they may still be important cues for recognition. Essentially, our goal is to **discourage over-reliance** towards those cues rather than removing them altogether. As a result, we show improved performances in the realistic image classification task (Table 3).\\n \\n#3 Compare with previous texture-shape disentanglement works.\\n \\n> There are several works that attempt to disentangle texture and shape for the image generation task [10-12]. However, as R1 would agree, they are irrelevant because texture is not always a bias. REBI'S is conceptually better because its conditional independence does not remove all texture cues. As a similar line of work, Stylized ImageNet [9] attempts to achieve the same effect by augmenting texturized versions of the images during training. The aim is to make the model rely more on shape cues than texture. We have empirically compared against Stylized ImageNet (also requested by R3).\\n\\nModels \\tBiased \\tUnbiased IN-A\\nResNet18 \\t 93.3 \\t85.8 \\t30.5\\nStylizedImageNet ResNet18 \\t92.5 \\t87.6 \\t29.7\\nREBI\\u2019S (ResNet18-BagNet18)\\t \\t93.7 \\t88.4 \\t31.7\\n \\nStylizedImageNet-trained ResNet18 shows improvements in reducing texture bias (improved \\u201cunbiased\\u201d accuracy from 85.8 to 87.6), but it does not increase the generalizability to the challenging natural adversarial examples (ImageNet-A accuracy from 30.5 to 29.7). REBI\\u2019S improves upon the StylizedImageNet-trained model both in terms of the unbiased and ImageNet-A accuracies (1.2 pp and 2.0 pp, respectively). We have updated the results in Table 3.\\n\\n<References>\\n[1] https://arxiv.org/abs/1610.02413 \\n[2] http://openaccess.thecvf.com/content_ECCV_2018/papers/Yingwei_Li_RESOUND_Towards_Action_ECCV_2018_paper.pdf \\n[3] https://arxiv.org/abs/1904.07911 \\n[4] https://arxiv.org/abs/1712.00377 \\n[5] https://arxiv.org/abs/1902.01007 \\n[6] https://arxiv.org/abs/1907.07355 \\n[7] https://arxiv.org/abs/1908.10763 \\n[8] https://arxiv.org/abs/1909.03683 \\n[9] https://openreview.net/pdf?id=Bygh9j09KX \\n[10] https://arxiv.org/abs/1811.11155\\n[11] https://arxiv.org/abs/1903.06946\\n[12] https://arxiv.org/abs/1812.02725\"}", "{\"title\": \"Response to Reviewer1 (1/2)\", \"comment\": \"We thank all the reviewers for their recognition of the task \\u201cinteresting and important\\u201d (R1), and finding the proposed method \\u201csound and interesting\\u201d (R2, R3). In particular, R3 has commented that \\u201cthe paper does a really good job of defining the problem setting\\u201d. Yet, we find some lack of detail and clarity in our paper could potentially have led to misunderstanding our work -- we have clarified them in reviewer-specific responses and have updated the paper accordingly.\", \"we_summarize_our_revisions_to_the_paper_as_follows\": \"\", \"section_2\": [\"Based on the helpful comments of R1 and R3, we have clarified that\", \"G may capture important signals as well as biases (Section 2.3, paragraph 3).\", \"G can be easily defined in a broad class of problems in machine learning applications (Section 2.3, paragraph 4).\", \"We have updated additional related works provided by R2 (Section 2.3, paragraph 4).\"], \"section_3\": [\"We have clarified that REBI\\u2019S does not suppress the signal captured by G due to its formulation as conditional independence (Section 3.1, Independence versus separation).\"], \"section_4\": [\"As requested by R3, we have added additional comparison on vanilla BagNets and Stylized ImageNet augmentation [1] on ImageNet (Table 3).\", \"There was a minor error in our evaluation code on ImageNet-A, and we have updated the correct values for every method (Table 3).\", \"We have fixed typos throughout the paper.\"], \"our_response_to_reviewer_1_is_as_follows\": \"#1 The authors assume that \\u201cthere is a family of feature extractors (G), such that features learned by any of these extractors would correspond to pure bias.\\u201d In practice, it is very hard to define the family of biased feature extractors.\\n \\n> In Section 2.3, we have defined G as the set of \\u201cbias extractors except for ones that can also recognize [essential cues]\\u201d. In practice, we agree with R1 that defining G that precisely captures bias is difficult, and it is likely in many applications that any definition of G still encodes \\u201cimportant cues\\u201d along with biases (imperfect precision). \\n\\nNonetheless, we argue that G does not need to precisely capture the bias. Our framework does not remove cues captured by G if they are essential for the target task. Note that REBI\\u2019S is built upon the \\u201cconditional independence\\u201d criterion (Section 3.1, \\u201cIndependence versus separation\\u201d paragraph) that f(x) is encouraged to be independent of g(x), **given** the target task Y. Unlike independence that requires f(x) to ignore g(x) altogether, conditional independence allows f(x) to still retain cues captured by g(x) if they highly correlate with Y. In short, our goal is to **discourage over-reliance** towards cues captured by g(x) and push the model to learn other features that are useful for prediction. In the revision, we have improved the description of the precision and recall conditions for G in Section 2.3: those conditions are not meant as hard requirements but as desiderata. Even if they are not precisely met, conditional independence will do the rest of the job. Please also see our response to R3 because many of our points are relatable.\\n\\nSince G does not have to be perfect, it is easy to define G\\u2019s in many interesting application scenarios using prior knowledge on the relevant biases. We supply many examples below (also included in the response to R3). Action recognition models have been reported to rely heavily on static cues without learning temporal cues [2,3]; though the actual bias may not precisely be _any static cue_, we can still regularize the 3D convolutional networks towards better generalization across static cue biases by defining G to be the set of 2D convolutional architectures. It has been argued that visual question answering (VQA) models, too, rely overly on language biases rather than the visual cues (e.g. without looking at the image, one knows the answer to \\u201cwhat color is the banana\\u201d is \\u201cyellow\\u201d) [4]. We can define G as the set of models looking at the language modality only. Entailment models are biased towards the presence of certain words (e.g. when there are many \\u201cnot\\u201ds, the sentence is \\u201ccontradictory\\u201d), rather than really understanding the underlying meaning of sentences [5,6]. In this case, we can design G to be the set of bag-of-words classifiers [7,8]. In the revised paper, we have supplemented the above examples in Section 2.3.\"}", "{\"title\": \"Important concerns\", \"comment\": \"It seems that reviewer 1 and I have very similar concerns.\\n\\nI would highly encourage the authors to address these while they can. I think it's crucial to answer these, and without a response, it's unlikely that I'll keep recommending an accept for the paper.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This manuscript discusses the problem of bias shortcut employed by many machine learning algorithms (due to dataset problems or underlying effects of any algorithmic bias within an application). The authors argue that models tend to underutilize their capacities to extract non-bias signals when bias shortcuts provide enough cues for recognition. This is an interesting and important aspect of machine learning models neglected by many recent developments.\\n\\nThe only problem is that the paper seems to be a bit immature as the exemplar application is too naive for illustrating the idea. The authors\\u2019 idea is to assume that\\n\\u2018there is a family of feature extractors, such that features learned by any of these extractors would correspond to pure bias. Then in order to learn unbiased features, the goal is to construct a feature extractor that is ''as different as this family\\\" as possible\\u2019\\nin practice, their claim is texture and color are biases; one should learn shape instead of texture and color. So the family of feature extractors are the ones with small receptive field that can only capture texture and color. Therefore, what they eventually achieved is the unbiased feature extractor only learns the shape of object and avoids learning any texture and color. \\n\\nSo, the problem is, in practice, it is very hard to define the family of biased feature extractors. It really depends on the dataset and the goal. Texture and color, in general, are still important cues for object recognition, removing this information is NOT equivalent to removing bias. Just as a suggestion, the background scene might be a better definition of bias. However, with the proposal in this paper, it would be unclear how to define the family of feature extractors for describing background. Therefore, the solution given for this important problem seems to be too ad-hoc and not generalizable. \\n\\nThe second example (that does not have an experiments on) is action recognition; the family of biased feature extractors is 2D-frame-wise CNNs (object recognition). The authors claim that objects are biases for action recognition systems, but again a large part of action recognition is indeed object recognition. Many actions are defined based interaction of humans with objects (e.g., opening bottle or pouring water from bottle). Some objects may be instroducing bias in the task, but not all. Again, the proposed solution in this paper cannot disentangle this. \\n\\nThe authors need to survey previous texture-shape disentanglement works and then compare with those methods.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper describes a methodology for reducing model dependance on bias by specifying a model family of biases (i.e. conv nets with only 1x1 convs to model color biases), and then forcing independence between feature representations of the bias model and the a full model (i.e. conv nets with 3x3 convs to also model edges).\\n\\nOverall the method is very interesting. A few very related recent works were missing: https://arxiv.org/pdf/1908.10763.pdf and https://arxiv.org/abs/1909.03683. These works provide different formulations for factoring out know \\\"bias oriented\\\" models (and focus more on NLP, although the second is applied to VQA). Although, I do appreciate that the bias models used in this work are slightly more general in terms of family than those studied before, they do encode significant intuition about the target bias to be removed and perhaps in this context, the methods cited above could also be compared. That being said, I don't feel the paper needs to provide this comparison given how recent those works are, but it would improve the quality of the paper. \\n\\nOne aspect that worries me about this paper is that most of the biases studied are synthetic, so specification of the bias family is trivial, in contrast to the works I mentioned above (where the bias model is potentially somewhat misspecified and needed to be discovered by different researchers). But I do really like the experiments on Imagenet-a. \\n\\nOverall the paper presents an interesting contribution that would be useful for future study in reducing bias dependance in ml.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"### Summary\\n\\nThe paper proposes a method for regularizing neural networks to mitigate certain known biases from the representations learned by CNNs. The authors look at the setting in which the distribution of biases in the train and test set remains the same, but the distribution of targets given biases changes.\\n\\nTo remove the bias, they propose hand-designing a set of models that rely exclusively on these biases to make predictions. They then train a new unbiased model by making sure that this new model uses sufficiently different information than the biased models to make predictions (By adding a regularization term)\\n\\n### Decision and reasons\\n\\nI vote for a weak accept.\", \"strengths\": \"\", \"1\": \"The paper does a really good job of defining the problem-setting. Contrasting cross-bias with cross-domain and in-distribution makes the goal of the paper very clear. The notation used to formalize the problem setting in Section 2.1 is also clear and concise. Moreover, the experiments on the toy dataset help clarify the proposed solution whereas experiments on Biased MNIST and Imagenet show that it successfully mitigates the bias. Finally, the authors show the importance of each component of the proposed solution by factor analysis.\", \"weaknesses\": \"\", \"2\": \"In the most general case, it is not obvious why it is easier to define and learn a set of models that only use noise to make predictions (which is the required first step for their proposed solution) as opposed to learning a model that only uses signal (which is the goal of the problem). These two problems seem equally hard. The paper builds on the premise that in some cases former is easier (i.e. in some cases, it is easier to learn a set of models that use only noise as opposed to learning a debiased model directly). The authors give two such examples (They only explore the first experimentally.)\\n1. Learning a model that relies on local texture. They achieve this by limiting the receptive field of the features. \\n2. Learning a model that relies on static images to make predictions about actions in videos. \\nI feel that the two given examples are very narrow. It would be nice if the authors could identify a broader class of problems for which G is given or can easily be defined. Moreover, even in these two examples, I'm not convinced that the proposed biased models only use B for making predictions. For example, for some classification tasks, the local texture could be part of the signal and not just the bias. Similarly, static images from a video do contain important information for making the prediction.\", \"3\": \"The authors only compare their method to a baseline that does nothing to debias the representations. Even though this is an important comparison (as it shows that the proposed method can debias the representations), it does not tell the reader how effective the proposed method is compared to other possible solutions. The results would be more meaningful if the authors could include at-least a simple baseline that tries to remove the bias in other ways (For example, they could use the style-transfer baseline used by Geirhos et al., 2019).\\n\\nI vote for accepting the paper as a poster. It introduces an interesting approach for debiasing representations. However, due to its narrow scope and missing baselines, I would not recommend the paper for an oral presentation. \\n\\n\\n### Questions\\n\\n1. What are some broader class of problems for which defining G is easier than directly regularizing for debiased representations?\\n\\n2. How well do Bagnets alone perform on the benchmarks in Table 3? I would expect to see that Bagnets alone do worse than vanilla Resnets on Unbiased and IN-A. Is that so? \\n\\n3. How well do other methods do in these domains? (Such as methods that directly debias the training data against texture by applying style-transfer). \\n\\n\\n### Update after Author's response\\n\\nThe author's response has clarified the motivation behind the proposed approach to an extent. They have also added a comparison with a method that directly promotes learning the shape as opposed to the texture ( by training on stylized Imagenet) \\n\\nI agree with R1 on all accounts (i.e. it is very hard to define the family of biased feature extractors, the proposed approach is ad-hoc, the authors need to compare to texture-shape disentanglement methods, etc), however at the same time, I can see that proposed approach can act as a useful heuristic for regularizing neural networks to pay attention to certain kind of information. \\n\\nAn interesting use-case of the proposed method (which the authors indirectly mentioned in their response to my review) is in a multi-modal setting. It's not trivial to enforce deep learning systems to utilize all data modalities in a multi-modal setting. By defining G to be models trained on individual modalities, it would be possible to nudge our models to pay attention to the information in all modalities. \\n\\nSome important results in deep learning have been ad-hoc (For example skip connections in deep networks, ReLUs) and have nonetheless progressed the field. This work is not as widely applicable as skip connections or ReLUs, but it is, nonetheless, providing a heuristic for solving an important problem.\"}" ] }

[ "Rui Shu", "Yining Chen", "Abhishek Kumar", "Stefano Ermon", "Ben Poole" ]

[ "disentanglement", "theory of disentanglement", "representation learning", "generative models" ]

{ "note_id": [ "3ooTOTTjsE", "rkll00W3jS", "HylboUWhsB", "Bke7wjpsjr", "BylIHsasjB", "HylxlVijor", "B1gJb49sor", "SyereqHjiS", "S1ehjYSojS", "SJgpa_BijS", "HJeDluCQjB", "SJgQ9TMl5r", "rkez3PlAtr", "HyeZnJF6KH", "r1lvXDuHYr", "ryli6_oGKr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment", "comment" ], "note_created": [ 1576798731931, 1573818055751, 1573815961024, 1573800794963, 1573800766454, 1573790695726, 1573786615261, 1573767660799, 1573767588504, 1573767364842, 1573279726736, 1571986826832, 1571846057895, 1571815336677, 1571288862736, 1571104962681 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1764/Authors" ], [ "ICLR.cc/2020/Conference/Paper1764/Authors" ], [ "ICLR.cc/2020/Conference/Paper1764/Authors" ], [ "ICLR.cc/2020/Conference/Paper1764/Authors" ], [ "ICLR.cc/2020/Conference/Paper1764/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1764/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1764/Authors" ], [ "ICLR.cc/2020/Conference/Paper1764/Authors" ], [ "ICLR.cc/2020/Conference/Paper1764/Authors" ], [ "ICLR.cc/2020/Conference/Paper1764/Authors" ], [ "ICLR.cc/2020/Conference/Paper1764/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1764/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1764/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1764/Authors" ], [ "~Weili_Nie1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper first discusses some concepts related to disentanglement. The authors propose to decompose disentanglement into two distinct concepts: consistency and restrictiveness. Then, a calculus of disentanglement is introduced to reveal the relationship between restrictiveness and consistency. The proposed concepts are applied to analyze weak supervision methods.\\n\\nThe reviewers ultimately decided this paper is well-written and has content which is of general interest to the ICLR community.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response\", \"comment\": \"Dear Reviewer,\\n\\nThank you for your kind reconsideration of our paper. We will adjust our paper accordingly to emphasize that consistency and restrictiveness are primarily intended as tools for theoretical analysis.\\n\\nRegarding your hypothesis that using consistency and restrictiveness metrics for model selection will select for less informative models, our finding on synthetic datasets is that this is not the case (Appendix G provides visualizations of the models selected by consistency. The latent interpolations indicate that the representations are informative). As such, we believe that uninformativeness is unlikely to be a confounder for Figure 4. To verify this, we will investigate this further and check, for example, that the (high-consistency + high-restrictiveness) models in Figure 4 are not attributable to uninformativeness. We apologize for not conducting this analysis before the rebuttal deadline, but we are confident about our hypothesis and will commit to releasing this analysis when it is ready.\\n\\nFor non-synthetic datasets where $g(Z)$ and $g^*(S)$ are likely to deviate considerably from each other, we agree that using solely consistency/restrictiveness for model selection may be biased toward uninformativeness models. One potential way to defend against uninformative models would be to estimate the mutual information between $X$ and $Z$ (which can be done in an unsupervised manner) and restrict model selection to models above a certain information threshold. Ultimately, however, we believe that model selection is best done with a proper disentanglement scoring rule, and that the development of such a scoring rule is still an open and important question.\"}", "{\"title\": \"Response\", \"comment\": \"Dear Reviewer,\\n\\nThank you for your kind reconsideration of our paper. Regarding your concerns about the practicality of the proposed methods, we wish to provide an adapted version of our response to Reviewer #3.\\n\\n----------------------\\n\\nRegarding the practicality of the proposed methods, we wish to note that the specific weak supervision methods we analyze (such as restricted labeling) are already employed on real-world datasets, for example, in style-content disentanglement. Our contribution is thus to demonstrate the right way these methods can be used or combined to guarantee disentanglement. \\n\\nWe acknowledge that the degree of supervision needed for methods to actually guarantee disentanglement might not necessarily be scalable in practice. We believe that determining when methods with disentanglement guarantees are scalable ultimately need to be addressed on a case-by-case basis in practice. Going forward, we think the following questions will be of increasing importance to the disentanglement community:\\n\\n1. What is the actual cost of weak supervision methods with guarantees for a particular setting? \\n\\n2. Are there settings that admit cheaper types of weak supervision methods with guarantees, compared to the specific methods analyzed in this paper?\\n\\n3. In what settings are weak supervision methods (that do not have guarantees) practically reliable? And when they are reliable, why are they reliable?\\n\\nWe believe our paper naturally inspires these important questions and hope that future researchers will study these questions through the lens of the perspective and tools developed in our paper.\"}", "{\"title\": \"Response (cont'd)\", \"comment\": \"References\\n---------------\\n\\nLocatello et al. (2019) Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations\\nBurgess et al. (2018) Understanding disentangling in \\u03b2-VAE\\nKim et al. (2018) Disentangling by Factorising\\nRidgeway et al. (2018) Learning Deep Disentangled Embeddings With the F-Statistic Loss\\nDenton et al. (2017) Unsupervised Learning of Disentangled Representations from Video\\nMathieu et al. (2016) Disentangling factors of variation in deep representations using adversarial training\\nWatters et al. (2019) Spatial Broadcast Decoder: A Simple Architecture for Learning Disentangled Representations in VAEs\"}", "{\"title\": \"Response\", \"comment\": \"Dear Reviewer,\\n\\nThank you for acknowledging the significance of our theoretical framework. The goal of this paper is to clarify some confusion regarding the concept of disentanglement in the current literature, and to build up a theoretical framework that clarifies the guarantees actually conferred by several popular weak supervision methods. And we hope that you will re-assess the merits of our paper with this perspective in mind.\\n\\n\\nOn the Significance of our Theoretical Analysis\\n-------------------------------------------------------------\\nFirst, we would like to emphasize the inherent value of studying the theoretical guarantees of various weak supervision methods. Theorem 2 shows that existing approaches such as restricted-labeling in style-content disentanglement literature do not actually provide disentanglement guarantees. In contrast, employing our calculus and Theorem 1 allows us to design weak supervision methods that guarantee disentanglement. We believe our demonstration is of inherent value, as it exemplifies the theoretical analysis of guarantees that have so far eluded the disentanglement community. In light of the findings in Locatello et al. (2019) regarding the existing literature's reliance on model inductive bias, we hope the reviewer will agree that providing a proper theoretical framework for understanding weak supervision is both timely and valuable.\\n\\n\\nOn the Significance of the Synthetic Experiments\\n-----------------------------------------------------------------\", \"the_experiment_section_can_broadly_be_divided_into_two_categories\": \"1. Experiments to test our theoretical guarantees (Sections 6.2.1 and 6.2.3)\\n\\nSections 6.2.1 and 6.2.3 primarily serve as demonstrations that the guarantees in Theorem 1 can be achieved in practice with machine learning models on synthetic datasets. Since ground truth factors are available on synthetic datasets, we used this as an opportunity to take extensive measurements to validate the various facets of our theoretical statements (Figures 3-5 and 7-16). Furthermore, the synthetic datasets that we chose are common benchmarks used in the existing disentanglement literature (Locatello et al. 2019; Burgess et al. 2018; Kim et al. 2018; Ridgeway et al. 2018; Denton et al. 2017; Mathieu et al. 2016; Watters et al. 2019). The performances of our models on existing disentanglement metrics (Figure 13) can also be directly compared with the scores reported in Locatello et al. (2019).\\n\\nWe agree that the use of non-synthetic data would further enhance the validation of these theoretical guarantees, but we hope that this does not detract from the legitimacy and extensive nature of our existing experiments.\\n\\n2. Experiments that assess the gap between theory and practice (6.2.2)\\n\\nFigure 4 of Section 6.2.2 addresses the following question: if Theorem 2 states that disentanglement is not guaranteed in, for example, style-content disentanglement, why have we as a community not established this as common knowledge yet? \\n\\nFigure 4 shows that the correlation between consistency and restrictiveness is quite strong across models tested across a broad range of hyperparameters. This finding leads us to make the following two conclusions.\\n\\n1. It is easy to mistake the practical reliability of style-content disentanglement as a theoretical guarantee. \\n\\n2. The practical reliability of style-content disentanglement is poorly understood and we wish to call attention to the need to provide a proper theoretical characterization of this reliability.\\n\\nWe believe both conclusions are important in contextualizing the current literature on weakly supervised disentanglement, and hope that more researchers will be aware of these findings. \\n\\n\\nOn the Practicality of Weak Supervision Methods with Guarantees\\n----------------------------------------------------------------------------------------\\nSince our paper shows examples of weak supervision methods that guarantee disentanglement, the reviewer is rightfully concerned about whether these methods can scale in practice.\\n\\nWe believe that the scalability of these methods ultimately need to be addressed on a case-by-case basis in practice. While not within the scope of our paper, we acknowledge that the following questions are both important and challenging:\\n\\n1. What is the actual cost of weak supervision methods with guarantees for a particular setting? \\n\\n2. Are there settings that admit cheaper types of weak supervision methods with guarantees, compared to the specific methods analyzed in this paper?\\n\\n3. In what settings are weak supervision methods (that do not have guarantees) practically reliable? And when they are reliable, why are they reliable?\\n\\nWe believe our paper naturally inspires these important questions. If the reviewer believes it is appropriate, we would like to include these questions as concluding thoughts in the paper to facilitate further discussion on weakly supervised disentanglement.\"}", "{\"title\": \"Thank you for your response.\", \"comment\": \"Thank you for your response. Now I understand that the 3 distribution matching cases in Theorem 1 imply $G(Z) \\\\overset{d}{=} g^*(S)$, which is a prerequisite for consistency and restrictiveness.\\n\\nI suggest the authors explicitly state that the consistency and restrictiveness defined in Equations (6) and (10) are more for theoretical analysis, rather than a practical metrics for disentanglement performance; because these metrics do not measure the informativeness. Note that $G(Z) \\\\overset{d}{=} g^*(S)$ is not likely to be true in practice. Therefore, if one uses consistency and restrictiveness metrics for model selection, it is very likely that a less informative model will be selected. In fact, I suspect the high correlation in Figure 4 $c(i)$ and $r(i)$ are due to the informativeness of the model, rather than the true relationships between consistency and restrictiveness. Because it has been observed that some disentanglement models, such as $\\\\beta$-VAE, give worse reconstructions when better disentanglement is achieved. This might suggest that in practice, a model that achieves a better disentanglement is likely to be less informative.\\n\\nI have updated my review rating because my major concerns are addressed.\"}", "{\"title\": \"updated score\", \"comment\": \"I still have some concern about the practicality of the proposed method but I think it is a valuable contribution and I have updated my score accordingly.\"}", "{\"title\": \"Response (final)\", \"comment\": \"Regarding Figure 4\\n--------------------------\", \"figure_4_indicates_two_phenomena\": \"1. When trained with single-factor weak supervision, examples of models that are \\u201cconsistent but not restrictive\\u201d and \\u201crestrictive but not consistent\\u201d do occur. This is consistent with our theory, since restrictiveness does not imply consistency and vice versa. \\n\\n2. The occurrence of models that are \\u201cconsistent but not restrictive\\u201d and \\u201crestrictive but not consistent\\u201d are rare in practice. \\n\\nOf the weak supervision choices shown in Figure 4, we note that restricted labeling is a common approach used to achieve disentanglement in, for example, the style-content disentanglement literature. The existing literature may mislead readers into thinking that restricted labeling can guarantee disentanglement. Our theory and phenomenon (1) indicates that restricted labeling does not have this guarantee. However, phenomenon (2) indicates that examples that violate disentanglement happen infrequently in practice. Taken together, our findings are interesting because\\n\\n1. We highlight that it is easy to mistake the practical reliably of style-content disentanglement as a theoretical guarantee.\\n\\n2. The practical reliability of style-content is poorly understood and we wish to call attention to the need to provide a proper theoretical characterization of this reliability.\\n\\n\\nRegarding Minor Feedback\\n------------------------------------\\nThank you for pointing out the theorem labeling mis-match in the appendix. We have updated the paper to fix this issue.\"}", "{\"title\": \"Response (cont'd)\", \"comment\": \"Regarding Figure 3\\n--------------------------\\nAs stated in the caption, Figure 3 is trained on Shapes3D. Figure 3 is a condensed version of a subset of the results shown in Figures 7, 8, 9, and 10. The key goal of Figure 3 is to verify our hypothesis that: restricted / share / rank pairing on a particular factor guarantees consistency on that factor, and that change pairing on a particular factor guarantees restrictiveness on that factor. \\n\\nAs stated in the captions, this hypothesis predicts that the diagonal components of the heatmap should have the highest scores. And indeed, this is generally the case. Even for the third heatmap, the diagonal components of the heatmap achieve the highest scores by a statistically significant margin (Figure 8 column 1 provides additional quantile information not displayed in the heatmap). \\n\\nWe believe the reviewer\\u2019s concern is moreso regarding why the margin of improvement (of diagonal v. non-diagonal scores) appears to be smaller for restrictiveness than for consistency. It is worth noting that single-factor restrictiveness measures the expected L2-discrepancy over a single factor, whereas single-factor consistency measures the expected L2-discrepancy over (n - 1) factors. This, we believe, is the main contributor to why restrictiveness scores tend to be higher than consistency scores on the off-diagonal elements. While we do not have a complete explanation for why Factors 3/4/5 for Shapes3D tended to have even higher off-diagonal restrictiveness scores than Factors 0/1/2, we note that Factors 0/1/2 control the color attributes (wall color, floor color, shape color) whereas Factors 3/4/5 control non-color attributes (object size, object shape, and azimuth).\", \"regarding_the_right_most_figure\": \"each column represents a specific intersection experiment. For example, column 0 is trained on a weak supervision of S_{0, 1, 2} and S_{0, 3, 4, 5}. Thus, by the calculus of intersection, only consistency of factor 0 (i.e. S_0) is guaranteed.\\n\\n\\nRegarding the Use of Consistency and Restrictiveness as Metrics\\n---------------------------------------------------------------------------------------\\nWe wish to emphasize that consistency and restrictiveness are intended primarily as concepts that establish the theoretical guarantees possible with weak supervision, and not as metrics that should be used in practice for evaluating disentanglement. In our paper, we converted these concepts into metrics (i.e. normalized consistency and restrictiveness scores) solely as a means of probing the behavior the models and for checking whether the behavior is consistent with our theoretical guarantees. \\n\\nAs probes, normalized consistency and restrictiveness also allow us to identify models that are consistent but not restrictiveness for a particular factor (and vice versa). As evident from Figure 4, such models do exist. We have added appendix B to illustrate two prototypical examples of models from disentanglement_lib that are consistent but not restrictive (and vice versa) that we found using our normalized consistency and restrictiveness probes across the population of 12800 pretrained models.\\n\\nWhile we do have some experiments demonstrating that summary statistics using consistency/restrictiveness can serve as reasonable surrogate metrics for the selection of disentangled models (see Figure 16 and Appendix F), we believe the development of a proper disentanglement scoring rule remains an open problem. We note that the primary goal of this paper is not to develop such a metric. Our goal is to clarify some confusion regarding the concept of disentanglement in the current literature, and to build up a theoretical framework that clarifies the guarantees (or lack thereof) that are actually conferred by various weak supervision methods.\"}", "{\"title\": \"Response\", \"comment\": \"Dear Reviewer,\\n\\nThank you for your feedback! We wish to address your concerns as follows.\\n\\nRegarding the message of the paper\\n-------------------------------------------------\\nThe goal of this paper is to clarify some confusion regarding the concept of disentanglement in the current literature, and to build up a theoretical framework that clarifies the guarantees actually conferred by several popular weak supervision methods\\n\\n\\nRegarding the seeming verbosity of the definitions\\n------------------------------------------------------------------\\nAs evident from Figure 2, consistency and restrictiveness are distinct concepts that play critical roles in theory. We agree that there is a noticeable contrast in the simplicity of the intuition of consistency/restrictiveness versus the lengthy mathematical description. Although those concepts are intuitive, the mathematical formalism we give is carefully designed with several important subtleties in mind:\\n\\n1. The specific sampling procedure in Eq (1, 2) and Eq (4, 5) are important for ensuring that the definition of disentanglement allows for correlated but interpretable features (similar to Suter et al 2018). We note that many existing definitions for disentanglement break down when the ground truth factors are correlated, and so we believe it is important to break away from the trend of making the definition of disentanglement implicitly reliant on the assumption of statistically independent ground truth factors.\\n\\n2. Since consistency and restrictiveness are asymmetric concepts, generator-side consistency and restrictiveness and encoder-side consistency and restrictiveness are highly related, but not completely identical concepts. And this is the reason why Theorem 1 explicitly provides guarantee for both generator and encoder-side consistency.\\nWe aim to further simplify our exposition in future iterations of the paper.\\n\\n\\nRegarding the Sufficiency for Disentanglement Formalism\\n-----------------------------------------------------------------------------\\nThe primary significance of this section is in formalizing what is meant by particular supervision method being insufficient. Our definition of sufficiency aims to distinguish the guarantees that arise solely from the choice of supervision versus the inductive bias of architecture or the objective function. Locatello et al.\\u2019s impossibility result is restricted to a specific kind of learning algorithm (matching p_data(x) with p_theta(x)). In contrast, our definition for sufficiency is agnostic to the choice of learning algorithm and inductive bias.\\n\\n\\nRegarding normalized consistency/restrictiveness scores\\n----------------------------------------------------------------------------\\nThe proof for why the score is bounded between [0, 1] directly follows from Lemma 1 in the Appendix I.2.1. We have updated the paper to make this connection explicit. Because the scores in the figure are estimated via monte carlo sampling, the monte carlo estimator is not guaranteed to be within the interval [0, 1]. However, we hope it is evident from the magnitude of the negative numbers (and from Figures 8, 9, 10) that the small negative values are noisy estimates of a score of zero.\"}", "{\"title\": \"Our paper addresses latent code informativeness (i.e. relevance)\", \"comment\": \"Dear Reviewer,\\n\\nThank you for acknowledging the significance of the problem being tackled in this paper. We would like to address your concern regarding the issue of degenerate generators that always output the same image for all choices of $Z$. In our paper, we referred to such a latent space $Z$ as being \\u201cuninformative\\u201d.\\n\\nThis scenario is explicitly prevented by our requirement, stated on page 3, that our analysis is subject to the condition that \\u201c$g(Z)$ =d= $g^*(S)$ are equal in distribution\\u201d. We would like to emphasize that the decomposition of disentanglement into restrictiveness and consistency is appropriate when this condition ($g(Z)$ =d= $g^*(S)$) is satisfied, otherwise we would indeed be admitting degenerate models that have uninformative latent spaces. We addressed this point once again on page 6, stating that\\n\\n\\u201cThe distribution matching requirement $g(Z)$ =d= $g^*(S)$ ensures latent code informativeness, i.e., preventing trivial solutions where the latent code is uninformative (see Theorem 7 for formal statement)\\u201d\\n\\nIn the appendix, Theorem 7 formalizes what it means for the latent space to be informative and states that informativeness of the latent space is guaranteed when $g(Z)$ =d= $g^*(S)$. The formal statement and proof are available on Page 33.\\n\\nIn other words, our proposed concepts of consistency and restrictiveness are best thought of as complementing latent code informativeness, and not as replacements for checking the informativeness of the latent code. We also wish to note that our theoretical guarantees in Theorem 1 explicitly require $g(Z)$ =d= $g^*(S)$. This is because the distribution matching setup in Theorem 1 implies distribution matching of $g(Z)$ with $g^*(S)$. \\n\\nAs such, with respect to our theoretical analysis, we hope that you will agree that our submission takes careful measures to explicitly combat the type of degeneracy that you pointed out. \\n\\nIn practice, when distribution matching is not guaranteed, we fully support that practitioners should check for both latent code informativeness in addition to restrictiveness and consistency. Ultimately, we note that generator restrictiveness and consistency are concepts that complement latent code informativeness, and not replacements for it. We will strive to make this point even clearer when we update the paper.\\n\\n---\", \"regarding_minor_comments\": \"In Figure 2, the way to interpret the vertical axis is that each row corresponds to a randomly sampled choice of $(z_2, z_3)$. This is stated in the caption as, \\u201ceach row denotes a fixed choice of $(z_2, z_3)$\\u201d. \\n\\nIn the calculus, you are correct that $n$ denotes the number of dimensions. We shall make this point clear when updating our paper.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper tries to bring some theoretical foundation to the weakly supervised disentanglement. Overall it is a good contribution, but the message of the paper is not clear. The authors propose two notions: consistency and restrictiveness, which they don't imply each other. However, the experiment on real data shows that they are highly correlated. Up until the experiment section, the paper is well written (although a bit verbose). It seems that it is great but unfinished work.\\n\\nThe paper is well written, but in my opinion, there is too much verbosity on page 4-5 on rather trivial definitions consistency and restrictiveness and a big box in the calculus of disentanglement that steals space from the main results. In my opinion, those sections can be reduced so that other theorem can be covered. In my opinion, the theorem nine should be part of the main text. \\n\\nI understand the definition of \\\"Sufficiency for Disentanglement \\\" but it is not clear why it is important. Sure, it is a strong definition that says for any $\\\\mathcal{H}$ (and not a subset) the algorithm ($\\\\mathcal{A}$) should be able to match the distribution of the observation but why is it a big deal according to the next paragraph?\\n\\nI don't see any proof that Eq.11 should be between [0,1]. Yes, g is optimal, and if you enter suboptimal values to it, one expects the nominator to be less than dominator. However, g a function that is optimal in expectation, which does not mean for every s value it nominator is less than the denominator. In fact, some of the values in fig 3 are small negatives.\", \"fig_3_is_not_explained_well\": \"you are showing normalized consistency and restiveness. First of all, what is the dataset you tried this on? Second, why some values are negative?! These are supposed to be between [0,1]. Third, what is the take-home-message of this figure? the first two matrices from left show that the factors are consistent b/c they are almost diagonal. The third one from left shows that the algorithm you used is not restrictive? Then are you suggesting this as a metric of evaluation? I am not sure I understand the first figure from the right.\\nOverall, the authors perform a significant amount of experiments, but they did a poor job in summarizing the results.\\nFinally, the authors claim \\n \\\"...We believe this correlation between consistency and restrictiveness to have been a general source of confusion in the disentanglement literature, causing many to either observe or believe that restricted labeling or share pairing on $S_i$ (which only guarantees consistency) is sufficient for disentangling Si ...\\\" \\n\\nEach of those methods should be analyzed separately to ensure that their algorithms do not induce restiveness. I just don't see the natural connection between your figure 4 and this conclusion that you made.\", \"minor\": \"Where is the proof for Theorem 1? In the Supp, it starts with Theorem 8, I guess you meant Lemma 8? You need to clean up the Supp so that one can find the proof easily. I suggest restructuring the Supp to less and finally proof of Thorem 1.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper first discusses some concepts related to disentanglement. The authors propose to decompose disentanglement into two distinct concepts: consistency and restrictiveness. Then, a calculus of disentanglement is introduced to reveal the relationship between restrictiveness and consistency. The proposed concepts are applied to analyze weak supervision methods.\\n\\nThis paper is well structured. The presentation is easy to follow. The problem discussed is important to the machine learning community. The concepts discussed are supported by a large number of experiments. \\n\\nThe assumption that disentanglement can be decomposed into consistency and restrictiveness might be flawed. Let us consider a generator $g(Z)$ that always generates the same image for all $Z \\\\sim p(Z)$. Note that $g(Z)$ gives perfect consistency and perfect restrictiveness as defined in Equation (3) and (6). However, we consider $g(Z)$ is a bad generator, and we do not think the corresponding latent representation $Z$ achieves perfect disentanglement. Note that such $Z$, in general, gives low values in the existing disentanglement metrics.\\n\\nThis implies that we might need to introduce a third component to disentanglement, which I call it relevance. We should additionally assume that different $z_{ \\\\setminus I}$ leads to different generated images. It might be challenging to measure relevance quantitatively under the probabilistic framework, but I believe this is necessary. \\n\\nIn summary, I think the idea presented is interesting and useful. I believe this paper is promising and impactful after proper revision. However, I do not recommend acceptance because it looks technically flawed.\", \"minor\": \"In Figure 5, the illustration is clear to me, but I am not sure how the vertical axis simultaneously represents two variables $z_2, z_3$.\\nIn the table on page 5, $n$ represents the number of dimensions, right?\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary\\n\\nThe paper tries to construct a theoretical framework to rigorously analyze the disentanglement guarantees of weak supervision algorithms. In particular, it focuses on two concepts, consistency and restrictiveness which provides a formalism that precisely distinguishes when disentanglement arises from supervision versus model inductive bias. \\n\\nStrengths\\n\\nThe framework uses two simple concepts, consistency and restrictiveness for both generator and decoder. It also gives rise to a calculus. It is very useful to demonstrate the conditions under which various supervision strategies guarantee disentanglement. \\n\\nThe paper also did a good job clarifying how consistency and restrictiveness differ from other disentanglement concepts used in the literature. \\n\\nWeaknesses\\n\\nThe paper does not propose effective methods for disentanglement in the weak supervision setting. \\n\\nThe experimental section uses very toy datasets. It is not clear how the weak supervision signal can come from in realistic applications.\"}", "{\"comment\": \"Thank you for the astute inquiry about the experimental results. Indeed, we expected Full-Label GAN to outperform the weakly-supervised GANs and were also surprised to find in Figure 12 that Change and Share pairing noticeably outperformed Full-Label on a number of metrics on the Cars3D dataset. While this could be attributable to a number of possible confounders (optimization issue, architectural differences, etc), our main hypothesis is that this is because one of the factors of variation in Cars3D is car-type (with labels {0, 1, \\u2026, 182}), for which an ordinal representation in a single dimensional continuous latent variable may be inappropriate.\\n\\nWe had chosen for simplicity to use only continuous latent variables in all of our experiments (this experimental design choice mirrors the Locatello 2019 study. We shall make this point explicit when we update the paper appendix). This means that Full-Labeling (and Rank Pairing), which imposes a specific ordering for how the car types are embedded along the latent dimension z3 (if one-indexing), may have trouble with using z3 to encode the car-type in the order prescribed by the actual labels. In contrast, Change and Share pairing do not have to adhere to the ordering prescribed by the car-type label. We believe this additional flexibility allowed for the Change/Share models to choose a more favorable ordering when casting car-type into a single-dimensional continuous representation, thus enabling better encoding of car-type into the latent variable z3. \\n\\nThis hypothesis is consistent with our interpretation of Figure 13, which shows the normalized consistency score for each label. Notice in (Figure 13, Cars column, Row 3) that Change/Share models significantly outperformed Rank/Full-Label on normalized consistency of the car-type factor, indicating that the Rank/Full-Label models had difficulty learning a representation of car-type that is consistent with the car-type labels.\\n\\nAs for SmallNORB, Figure 12 shows that Full-Label GAN is the top performer in 2 / 6 metrics, and a top-2 performer 5 / 6 metrics. As such, we do not think we can definitively claim that the weakly supervised GANs outperform Full-Label GAN on SmallNORB.\\n\\n=====\", \"we_have_also_uploaded_an_anonymized_version_of_our_code_to\": \"https://drive.google.com/open?id=1VjMNOD3uFrCx4nZSgLlKrVLVbluPnGPa\\n\\nPlease let us know if you have any trouble running it.\", \"title\": \"Response and link to code\"}", "{\"comment\": \"Thanks for the interesting work, but I have a question about the experiments in Section 6.2.3: Why is the fully-supervised GAN worse than weakly-supervised GAN, in particular on SmallNORB and Cars3D? Because, if I understand correctly, we normally we expect the fully-supervised method performs better. Also, May I ask if the authors could provide the code for the implementation of weakly-supervised GANs?\", \"title\": \"One question about the experiments\"}" ] }

[ "Zezhen Zeng", "Jonathon Hare", "Adam Prügel-Bennett" ]

[ "VAE", "GAN" ]

{ "note_id": [ "qgQCJN2MA", "SJeDpXb3jH", "ByeO-pZ7jr", "BJlR2h-Qjr", "HylIK2ZQjr", "B1eCVhZXjS", "Syg8gRa6tr", "Bkxzxw96Kr", "Skg_RY4tdB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731904, 1573815230723, 1573227775910, 1573227702499, 1573227645809, 1573227573665, 1571835373710, 1571821290360, 1570486736314 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1763/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1763/Authors" ], [ "ICLR.cc/2020/Conference/Paper1763/Authors" ], [ "ICLR.cc/2020/Conference/Paper1763/Authors" ], [ "ICLR.cc/2020/Conference/Paper1763/Authors" ], [ "ICLR.cc/2020/Conference/Paper1763/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1763/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1763/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper proposes a new method for improving generative properties of VAE model. The reviewers unanimously agree that this paper is not ready to be published, particularly being concerned about the unclear objective and potentially misleading claims of the paper. Multiple reviewers pointed out about incorrect claims and statements without theoretical or empirical justification. The reviewers also mention that the paper does not provide new insights about VAE model as MDL interpretation of VAE it is not new.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to rebuttal\", \"comment\": \"Thank you for the rebuttal.\\n\\nIf the goal is not to propose a new model, but to provide new insight into the workings of a VAE, I fail to see what the actual new insights are that this paper provides.\", \"on_the_reconstruction_error_term\": [\"Your reconstruction loss ($N \\\\log(E_{MSE})/2$) is unbounded from below, this cannot be a good loss. Stating you\\u2019ve never observed numerical instabilities due to this does not mean it\\u2019s a sound thing to do.\", \"If you model $p(x|z) = N(x|\\\\mu=\\\\hat x(z), \\\\sigma^2(z))$, the exact choice of the dependence of $\\\\sigma^2(z)$ on $z$ is a modeling choice for your decoding distribution. Choosing $\\\\sigma^2=const$ is one of the valid choices. Note that even though for this choice no gradients will flow through $\\\\log(2\\\\pi\\\\sigma^2)$, it does influence the ratio between the reconstruction term and the KL. And yes, when restricting $\\\\sigma^2$ to a constant, picking a different constant will lead to a different ratio of the gradient sizes coming from the reconstruction term and the KL divergence, but this is a consequence of choosing a different parameterization for the Gaussian decoding distribution. One could also choose $\\\\sigma^2$ to be a function of $z$, and independently vary the beta in a $\\\\beta$-vae.\"]}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"We are glad that the paper was a pleasure to read. We feel it was taken\\nin the spirit it was written as a thought provoking reflection on VAEs.\\nWe are sad that in the end the reviewer did not see the point. Clearly,\\nthis was a failure of our paper, but we feel that one of the problems\\nwas that the main novelty was taken to be LSR-GAN, which for us was just\\na visualisation tool. The main novelty was the interpretation of the\\nVAE and drawing conclusions from that interpretation.\\n\\nTo clarify why LSR-GAN is useful. Although VAEs give reasonable\\nreconstructions on training and test data, for many data sets they have\\npoor performance in generating images from randomly sampled parts of the\\nlatent space. Clearly this is because the VAE decoder has not been\\ntrained in these regions. However, these are regions of great interest\\n(the latent representation of VAEs are there real strength and we would\\nlike to these to be meaningful away from training images). LSR-GAN give\\nus much better images when we sample from the latent space (as evidenced\\nby FID scores for example). Due to page limits we had to relegate\\nexamples of these to Appendix E, but these are remarkably more realistic\\n(to be fair surreal) than we would get from using the VAE decoder.\\n\\n*** 1.\\n\\nWe don't quite see what you are disagreeing about. We called this a\\n\\\"perceived problem\\\" because, as we strongly argued, VAEs are not\\ndesigned for reconstruction. Clearly a lot of work on VAEs is focused\\non the latent representation they learn. Nevertheless, some of the\\nliterature has focused on improving performance. We have no desire to\\nbelittle work, for example, on preventing latent variable collapse, but\\nwe also think it is worth stating that this could also be viewed as\\nautomatic dimensionality selection mechanism, which is a desirable\\nproperty rather than a problem that needs to be fixed.\\n\\n*** 2.\\n\\nWe are trying to present a new perspective and in doing so we wanted to\\nmake a strong (provocative?) statement. We accept that the decoder is a\\ngenerative model in that it generates something. Our point is that it\\ndoes not generate a realistic image by design (in contrast to a GAN\\nwhich is designed to generate a realistic image). Undoubtedly different\\ninformation will be stored in the latent space depending on the decoder,\\nbut \\\"incompressible\\\" information (e.g. information that appears only in\\none training example) won't be stored for any decoder.\\n\\n*** 3.\\n\\nWe tried to give a concrete example (inserting the word probably because\\nwe are speculating). Our point was to explain why you would expect some\\nblurriness. A precise description of an image would seem incompressible\\n(otherwise VAEs should achieve zero reconstruction error which in our\\nexperience is quite far from what we find).\\n\\n*** 4.\\n\\nWe totally agree! We concede that it would be useful to make this\\nexplicit. Our whole paper is to point out that a VAE shouldn't be\\nviewed as a pauper's GAN. By design the decoder is not meant to produce\\ngood reconstructions.\\n\\nWe have clearly failed to communicate the purpose of this paper. We\\nbelieve the MDL interpretation of VAEs provide many novel insights. We\\ncannot claim the MDL interpretation is new, but the insights are not\\nwell known. They include the observation that the decoded images should\\nbe poor quality, that latent variable collapse is desirable, that latent\\nspaces need to imagined because they won't capture all the information\\nin an image and even that the correct form of the reconstruction term is\\n$N \\\\log(\\\\sigma) + const$ (i.e. the entropy of $N$ independent variables from a\\ndistribution $N(0,\\\\sigma^2)$). These are not revolutionary, but at the same\\ntime they are not well known and we feel they are important.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We strongly defend our claim about the form of the reconstruction error.\\nAlthough this was made more as a side-remark we believe that it is very\\nimportant to calculate the reconstruction error as we stated.\\n\\nThe reconstruction error is actually an expected negative log\\nprobability, which by definition is the entropy of the error. Under the\\nstandard assumption that the errors are zero mean normal variables this\\nterm should be the entropy of a Gaussian/normal distribution which is\\nindeed equal to $\\\\log(\\\\sigma^2)/2+const$ as we assert. The exponential\\nof the negative entropy for a zero mean Gaussian is not a Gaussian\\ndistribution (or a distribution at all) so we don't accept your\\nstatement that this has to be wrong---it follow algebraically by putting\\nin the definition of $\\\\sigma^2$ into equation (2). If you want to put\\nin an unbiased statistical estimator for the variance it just changes\\nthe constant term by a very small amount. Given we don't know $\\\\sigma$\\na-priori using an empirical estimate seems the only alternative (given\\nthe number of errors we also expect that the empirical estimator (biased\\nor unbiased) will be a very good approximation.\\n\\nWe accept that our term is not common (we have not seen it used\\nanywhere), but we are entirely confident that it is correct!\\n\\nThe assertion that you can use any value of sigma you choose, we find\\ndifficult to understand. If instead of using $\\\\sigma^2=1/2$ we used\\n$\\\\sigma^2=1/20$ this would be equivalent to changing the relative\\nproportion of the KL term by a factor of 10. In the \\\"true\\\" VAE this\\nshould be the a true probability of making an error. Given the errors\\nhave a variance, we should use that variance. When you do this you will\\nget the expression we derived.\\n\\nThe reconstruction error will only overflow if the mean squared error\\nfor all pixels, colour channels and the whole mini-batch is zero. We\\nhave never experienced this! The modification is very easy to\\nimplement. All you need to do is replace MSE in the loss function by\\nN*log(MSE) where N is the number of colour channels times the number of\\npixels times the size of the minibatch. In our experience this is\\nequally easy to learn and will give you better reconstructions (although\\nthis could depend on the scaling you use in your implementation).\\n\\nAlthough in the MDL interpretation it is clear that the negative\\nlog-probability term is just an entropy which you wish to minimise, it\\nis also the case in the original paper that you must minimise the\\nlog-probability of the errors. Arbitrarily choosing $\\\\sigma^2=1/2$ no\\nlonger makes this a probability of anything meaningful.\\n\\n*** Comments on experiments and related work\\n\\nThere are no experimental comparisons because that was not the point of\\nthe paper. The paper was about understanding VAEs. LSR-GAN was\\nintroduced as a tool to visualise the information stored in the latent\\nspace of a traditional VAE. Other VAE-GAN hybrids are not intended to\\ndo this so a direct comparison makes no sense.\\n\\n*** Other Sections\\n\\nThe whole point of the paper is to explain what at VAE does using the\\nMDL framework. This, in our opinion, provides some new insights.\\nWithout Section 3 there is really no content.\\n\\n*** Additional Feedback\\n\\nOur comment \\\"VAEs are often taken as a pauper's GAN\\\" was to succinctly\\ncapture two aspect of VAE (a) they are easier to train than GANs and\\n(b) they produce blurrier outputs than GANs. However, the whole paper\\nis written to point out that VAEs produce blurry reconstructions by\\ndesign. It is not the real point of VAEs. VAEs produce rich latent\\nembeddings of images. Our intention was very far from insulting VAEs\\n(clearly we would not have written the paper if we felt VAEs were\\nuninteresting). We can, of course, rephrase this, we had deliberately\\nchosen the phrase to capture one view of VAEs, but the point of the\\npaper is to show that this view is superficial and misses the point.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"*** Concern 1)\\n\\nWe believe the statements we have made are justified both theoretically\\nand empirically. We take the statements highlighted\\n\\n\\\"the whole point of VAEs is to capture only compressible information and\\ndiscard information specific to any particular image\\\"\\n\\nThis is a crucial argument we believe is justified by the minimum\\ndescription length (MDL) interpretation of the VAE loss function. The\\nloss function consists of an expected negative log-probability or the\\nentropy of the errors. By Shannon's theorem this is also the minimum\\ncode length needed to communicate the error. The KL term is a relative\\nentropy that has an information theoretic interpretation of the code\\nlength needed to communicate a random variable with distribution $q(z|x)$\\nusing a coding with an underlying distribution of all code words of\\n$p(z)$. Therefore we can see the VAE loss function as the cost of\\ncommunicating the original image by sending a code $z$ that the decoder\\ndecodes as $\\\\hat{x}$ and an additional cost of repairing the image by\\ncommunicating the errors.\\n\\nAs a consequence, if a single image contains a butterfly in a corner, it\\nmakes no sense for this to be encoded in the latent space as this would\\nbe at least as costly as communicating the error. If many images\\ncontained butterflies then you can reduce the overall code length by\\ncoding this into the latent space. In contrast, an autoencoder would\\nattempt to encode all information in the latent space. This is what we\\nmean by \\\"only compressible information\\\" and \\\"information specific to a\\nparticular image\\\". We believe these are theoretically justifiable\\nstatements by the MDL interpretation.\\n\\n\\\"strength of a VAE is that it builds a model of the dataset that does\\nnot over-fit\\\"\\\"\\n\\nThis follows theoretically from the above argument, but it leads to an\\ninteresting prediction. A normal autoencoder (or a beta-VAE with\\nbeta<1) will provide better reconstructions of the training images (we\\ndidn't explicitly show this, but this is quite well known). However, it\\nwon't produce better reconstructions on unseen testing images because the\\nextra-information encoded by the autoencoder is image specific\\n(over-fitting the training set). The graphs in Figure 5 confirm this\\nprediction empirically showing for a beta-VAE the reconstruction error\\non images from the CIFAR-10 test set are the same for $\\\\beta\\\\leq1$ and\\nonly become worse when $\\\\beta>1$.\\n\\n\\\"VAEs are not broken...\\\"\\n\\nOur point is that VAEs are designed to balance the two terms in their\\nloss function which will lead to poor reconstructions. Fixing this\\nwould mean that image-specific information was encoded in the latent\\nspace which is precisely what VAEs are designed to avoid.\\n\\n*** Concern 2)\\n\\nAs we explained this paper is about different view of VAEs. LSR-GAN was\\nnot designed to produce better images than other GAN-VAE hybrids its\\npurpose is to show (imagine) what is encoded in the latent space of a\\nVAE. Perhaps it was a mistake to introduce LSR-GAN as it has clearly\\nbeen a distraction, but its purpose was to make a point that the latent\\nspace of a VAE has to be imagined as it does not encode all the\\ninformation in an image.\\n\\n*** Concern 3)\\n\\nClearly we have to accept that the paper has failed to communicate what\\nwe set out to communicate. However, as we pointed out this is different\\nto most papers. Our new method is a visualisation tool to understand\\nVAEs rather than yet another attempt to generate high quality images.\\n\\n*** General Comment\\n\\nWe feel this is a bit harsh. A lot of the argumentation seems to have\\nbeen missed leading to an impression that our statements were rhetorical\\nrather than based on a sound theoretical basis. We chose a tone to\\nhighlight that we are challenging common held beliefs, but we can clearly\\nchange this if has caused confusion.\"}", "{\"title\": \"To All Reviewers\", \"comment\": \"We would like to thank the reviewers for the time spent on reviewing the\\npaper. The purpose of the paper is to take a step back and ask the\\nbigger question \\\"What is the purpose of a VAE?\\\" (why add the KL term).\\nWe believe understanding this is important when trying to improve their\\nperformance. It is not our intention to propose a new network that\\nbeats the current state-of-the-art on some metrics. We propose LSR-GAN\\nas a tool for visualising what is encoded in a standard VAE (as we argue\\nthis involves imaging details that won't be encoded). As other VAE-GAN\\nhybrids were not designed for this purpose a direct comparison seems\\nredundant.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposes a new method for improving generative properties of VAE model. The idea is to train VAE in two stages: at first, train the vanilla VAE, then at the second stage freeze the encoder part and train the decoder part as a GAN generator with an additional regularizer which encourages cycle consistency in the latent space. Also the authors claim that other VAE-GAN hybrids which try to improve VAE model are \\u201cmisguided\\u201d and poor samples and reconstructions of VAE are the consequence of minimum description length problem.\", \"concerns\": \"1) The main concern about this paper is the inaccuracy and very general statements without theoretical or empirical justification. For example, the authors claim that \\u201cthe whole point of VAEs is to capture only compressible information and discard information specific to any particular image\\u201d. What is the definition of \\u201conly compressible information\\u201d or \\u201cinformation specific to any particular image\\u201d? Is there an experiment which can support this statement? Other examples of such general statements: \\u201cstrength of a VAE is that it builds a model of the dataset that does not over-fit\\u201d, \\u201cthe latent code does not contain enough information to do the reconstruction\\u201d, \\u201cVAEs are not broken and \\u201cfixing\\u201d them is actually likely to break them\\u201d. \\n2) Poor experiment comparisons with other baselines. The authors compare their method only with the vanilla VAE which is clearly insufficient. The authors claim that other VAE-GAN hybrids break the \\u201cstrength of the VAE\\u201d. Could you please provide examples of problems and provide experiments where VAE-GAN baselines will be worse than VAE or the proposed method?\\n3) The paper structure is very confusing. The main part and experiments part are mixed. Therefore, it is hard to follow the text. The experiment setup is not clearly stated. For example, it is unclear for which dataset luminance-normalized Laplacian was computed. \\n\\nOverall, the paper proposes a new method and gives an alternative view on the VAE model. However, statements from the paper are very pretentious and are not rigorously proven. Also there are not empirical comparisons with other VAE-GAN hybrids. Therefore, I would suggest rejecting the current version.\\n\\n--------------------------------------------------------\\nUpdate after author rebuttal\\n\\nThank you for your thoughtful response. However, I still think that the paper does not give new insights about VAE model and has poor experiment justifications of their statements. Considering MDL interpretation of VAE it is not new (see [1]). Therefore, the contribution of this paper is very limited. \\n\\nAfter reviewing the other reviews and the author rebuttal, I do not change my original score.\\n\\n[1] Xi Chen et al., Variational Lossy Autoencoder, 2016\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary:\\nThis paper proposes a hybrid VAE-GAN model, called the latent space renderer-GAN (LSR-GAN), with the goal to \\u201cimagine\\u201d the latent space of a VAE, and to improve the decoding and sampling quality of a VAE. First, a VAE-like model is trained, after which the encoder weights are frozen, and the decoder is trained as the generator of a GAN (together with an auxiliary discriminator). The generator loss also contains a reconstruction-like term in the latent space, described by the negative log density of the encoding distribution of the original latent conditioned on the output of the generator: -log q(z|g(z)).\", \"decision\": \"reject\\nThis paper contains incorrect claims. Leaving those mistakes aside, the experiments don\\u2019t lead to new insights and no comparison against other VAE-GAN hybrids is made.\", \"supporting_arguments_for_decision\": \"In the introduction the authors state \\u201cThis is a consequence of the well known evidence lower bound or ELBO objective function consisting of a negative log-probability of generating the original image from the latent representation (this is often implemented as a mean squared error between the image and the reconstruction, although as we argue in Appendix A this term should be proportional to the logarithm of the mean squared error)...\\u201d. This statement is surprising and I don\\u2019t see how it can be correct. If -log p(x|z) is something like the log of the mean squared error, then the density p(x|z) should be a squared error function, which is not even a valid distribution.\\n\\nTo be more precise, according to the authors, if one takes a Gaussian p(x|z), the reconstruction error should be modeled with \\n\\n-log p(x|z) = log 1/N sum (x_i - \\\\hat x_i)^2 + const \\t\\t\\t\\t\\t(1)\\n\\nwhere \\\\hat x_i is a function of z. Note the logarithm on the right-hand side here. In appendix A, the authors observe that most other implementations instead optimize \\n\\n-log p(x|z) = sum_{i=1}^N (x_i -\\\\hat x_i)^2/(2\\\\sigma^2) + N/2 log(2 pi sigma^2) \\t\\t(2)\\n\\nWhere often sigma is set to \\u00bd (so that the last term on the lhs of (2) drops out when gradients are taken). The authors claim the latter is incorrect because sigma should be equal to the (biased) empirical variance. They then insert sigma=empirical variance into (2), take derivatives while ignoring the dependence of sigma on \\\\hat x_i, and then arrive at something like eq. 1. They also argue that those implementations that use sigma=\\u00bd are actually optimizing a beta-VAE because of this \\u201cincorrect\\u201d prefactor of the reconstruction error. \\nI\\u2019m confident that the above claims are not correct. \\nAs an example, using a Gaussian decoder distribution requires parameterizing the mean and variance of that Gaussian distribution as a function of z. Here, setting sigma = \\u00bd (as is commonly done in other literature) is valid, contrary to the above claim. One is always allowed to just set the variance to a constant and ignore its dependence on z. This just leaves you with a less flexible distribution to model p(x|z).\\n\\nThe authors appear to use (1) as the reconstruction error term in the ELBO during VAE optimization, and then play with different prefactors of the KL term similar to beta-VAEs. The combination of the reconstruction error (1) and the prefactor no longer makes this a VAE, but more like a regularized auto-encoder with an unconventional reconstruction error that should not be interpreted as coming from the negative of a log density. Due to the logarithm in front of the mean squared error, the loss function is less sensitive to large errors in reconstruction. It is surprising that the reconstruction error does not cause overflow as log (0) --> - infinity.\", \"comments_on_experiments_and_related_work\": [\"Experiments only show images of reconstructions of a VAE and LSR-GAN, a mean squared error plot of reconstruction errors and an accuracy plot of a classifier evaluated on the reconstructions produced by both models. In terms of MSE the LSR-GAN actually performs worse than their VAE baseline. The authors also train a classifier on cifar-10 and measure its accuracy using the ground truth labels and the reconstructed images of a VAE and LSR-GAN. Here the LSR-GAN performs marginally better than the VAE, but the overall accuracy is very poor.\", \"No experimental comparison is made against other VAE-GAN hybrids. Related work on hybrid VAE-GANS is discussed in the appendix, not in the main paper.\"], \"other_sections\": \"- Section 3 contains a very long interpretation of the minimum description length (MDL). It is unclear what the goal of this section is, as it does not lead to any results, nor does it help bring the point across as to why the proposed model is good at imagining the latent space of a VAE.\\n\\n\\nAdditional feedback to improve the paper (not part of decision assessment):\\nThe additional loss term of the generator is very similar to what is used in reweighted wake sleep, although this is not mentioned. It could be worth making a connection here. \\nIn the conclusion the authors state \\u201cVAEs are often taken to be a pauper\\u2019s GAN\\u2019. This is not a very scientific statement and can be perceived as insulting for various reasons. Please rephrase this.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes to augment the VAE objective with two additional terms: (1) a GAN-like objective that ensures that the generated samples are not distinguishable from real samples, and (2) an additional regularizer to make sure that the latent variable that generated the image can be reconstructed from the VAE encoder.\\n\\nOverall the paper is written very well and was a pleasure to read. In particular I appreciated the differences between the present work and the many VAE/GAN hybrids in section B of the Appendix. However, after reading the paper a couple of times, it was still not clear to me what the \\\"main point\\\" of the paper is. To be more specific:\\n\\n- From the perspective of obtaining good unconditional samples from a VAE-like model, the authors do not compare their approach against like methods like IntroVAE (https://arxiv.org/pdf/1807.06358.pdf) which also adds an adversarial objective to the VAE. Qualitatively at least, it seems apparent that the generated samples here are worse in quality that the IntroVAE work.\\n\\n- From the perspective of learning more faithful reconstructions, it seems clear from Figure 5(a) that a regular beta-VAE does better. And to me it is not clear why we would want faithful reconstructions?\\n\\n- The main novelty of the proposed method is in adding an extra term to the generator loss controlled by lambda (equations 1 and 4). In the appendix the authors experiment with varying the lambda parameter and find that generally having lambda > 0 produces \\\"better\\\" metrics, though the choice of these metrics are somewhat questionable. It would be great to see the generated samples as \\\\lambda is varied (to me this is more interesting than seeing the generations as \\\\beta is varied).\\n\\nFurther, I take several issues with the authors' point of view regarding the current state-of-affairs in VAEs:\\n\\n1. \\\"However, one of their perceived problems is their reconstruction performance.\\\" \\n\\nI somewhat disagree with this characterization. Sure, there has been much work on modifying the VAE objective such that the latent variable is not ignored (i.e. posterior collapse), but the point of these works is not to get \\\"better reconstruction performance\\\". \\n\\n2. \\\"However, having a model that does not over-fit the dataset can be useful, but in this case the decoder of a standard\\nVAE should not be regarded as a generative model\\u2014that is not its purpose. If we wish to generate\\nrealistic looking images we need to imagine the information discarded by the encoder\\\"\\n\\nI am not sure I understand this characterization. The decoder is by definition a generative model. Depending on the decoder/encoder capacities (e.g. PixelCNN decoder vs DeConvNet decoder), different types of information will be encoded in the latent space.\\n\\n3. \\\"The job of the decoder in a variational autoencoder is to reconstruct the image only using information\\nthat can be compressed. Image specific information is ignored. For example, information about the\\nprecise shape of an object is probably not compressible. As a result the decoder tends to hedge its\\nbets and has a blurry outline.\\\"\\n\\nAgain, all of this is dependent on how the encoder/decoder is parameterized. I do not agree that \\\"information about the precise shape of an object is probably not compressible\\\". For example see https://hal.archives-ouvertes.fr/hal-01676326/document\\n\\n4. \\\"VAEs are often taken to be a pauper\\u2019s GAN. That is, a method for generating samples that is easier\\nto train than a GAN, but gives slightly worse results.\\\"\\n\\nMy view on VAEs is that they are a way of training latent variable models with likelihood training. One potential application of this is to generate samples, but that is not the only (nor the primary) application. \\n\\n\\nFinally, I hope I am not coming across as nitpicking or overly combative, but I am genuinely confused as to the problem that this paper is addressing. I look forward to discussing further among other reviews and the authors during the rebuttal period.\\n\\n[Response to author rebuttal]\\n\\nThank you very much for your thoughtful response. However, I must say that I just fundamentally disagree with motivations and some of the statements made in the paper. In particular, there seems to be some conflation (I could be misunderstanding) of \\\"good reconstruction\\\" vs \\\"good generation\\\". For example:\\n\\n- \\\"Our point is that it does not generate a realistic image by design (in contrast to a GAN which is designed to generate a realistic image).\\\" \\n\\nI am not sure I understand. Would you say an autoregressive model (e.g. PixelCNN) does not generate realistic images by design? (Clearly, they do!). Would you say that autoregressive language models (e.g. GPT2) do not generative realistic language by design? The argument seems to be that likelihood-based training of generative models does not, by design, encourage realistic-looking images (indeed it is true that good likelihood does not *necessarily* imply good generative models). However, there is massive empirical evidence that likelihood training does result in good generation.\\n\\nAfter reviewing the other reviews and the author rebuttals, I am maintaining my original score.\"}" ] }

[ "David Salinas", "Huibin Shen", "Valerio Perrone" ]

[ "Hyperparameter optimization", "Bayesian Optimization", "Gaussian Process", "Copula", "Transfer-learning" ]

{ "note_id": [ "_4hYzqt6tc", "HJg70jIujS", "SJg7RcUdjH", "r1gHWqIOjB", "rkxI0KIuoS", "rJgxW_iQ5S", "HkeNJPwAFS", "BkecWwptYr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731876, 1573575627113, 1573575370705, 1573575164643, 1573575118371, 1572218871553, 1571874524458, 1571571457827 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1762/Authors" ], [ "ICLR.cc/2020/Conference/Paper1762/Authors" ], [ "ICLR.cc/2020/Conference/Paper1762/Authors" ], [ "ICLR.cc/2020/Conference/Paper1762/Authors" ], [ "ICLR.cc/2020/Conference/Paper1762/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1762/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1762/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper tackles the problem of transferring learning between tasks when performing Bayesian hyperparameter optimization. In this setting, tasks can correspond to different datasets or different metrics. The proposed approach uses Gaussian copulas to synchronize the different scales of the considered tasks and uses Thompson Sampling from the resulting Gaussian Copula Process for selecting next hyperparameters.\\n\\nThe main weakness of the paper resides in the concerns raised about the experiments. First, the results are hard to interpret, leading to a misunderstanding of performances. Moreover, the considered baselines may not be adapted (they may be trivial). This might be due to a misunderstanding of the paper, which would align with the third major concern, that is the lack of clarity. These points could be addressed in a future version of the work, but it would need to be reviewed again and therefore would be too late for the current camera-ready.\\n\\nHence, I recommend rejecting this paper.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer 1\", \"comment\": \"We thank the reviewer for the two relevant references on related methods to achieve transfer learning. However, the two approaches differ importantly in our opinion. The first approach [1] fits a GP on observations of all tasks that are mapped through a rank estimator to avoid scale issues. We agree that the motivation for the ranking estimator is very similar: enable transfer learning by avoiding scaling issues. However, this approach is not applicable in our case as the runtime scales cubically in the number of observations, which is always greater than 50K. For this reason, [2] also claims that this baseline cannot be applied to a larger dataset. Furthermore, this approach cannot adapt to discrepancies of the current task as all observations are seen through an estimator fit on other tasks. For instance, a rank predictor always predicting zero will feed only zeros as observations to the GP, whereas our approach allows the posterior to deviate from the prior given new observations. Finally, the rank estimate is a single point prediction and uncertainty is not estimated from other tasks.\\n\\nThose issues motivated the second paper that you referenced [2], which proposes linear mixture of experts where the experts are GPs fitted on each dataset. To avoid numerical scale issues, the acquisition function is adapted to measure the weighted improvement on all tasks rather than just the current one. However, this approach only supports Gaussian noise, requires a non-standard acquisition function as well as hand-designed meta-features to estimate the mixture weights. In addition, it still requires fitting a GP on each dataset where the number of observations can be greater than 50K. Our parametric approach does not suffer from these shortcomings. We discussed these papers in the revision.\\n\\n[1] Collaborative hyperparameter tuning\\n\\n[2] Scalable Gaussian Process based Transfer Surrogates for Hyperparameter Optimization\\n\\nWhen comparing against warm-start GP, we took 100 evaluations to get a representative number of samples from the closest task without making the cubical scaling prohibitive. Following on the reviewer\\u2019s feedback, we ran additional experiments by taking the best evaluation from each related task to initialize the GP. The results confirm that our approach still outperforms this baseline, pointing to the benefits of mapping the tasks to a comparable space. We added these results to the paper.\\n\\nAs suggested, we also checked the CDFs of the error metrics of these tasks and included the plots in the appendix. From the CDFs, we observe in XGBoost that more hyperparameter configurations lead to higher objective value than in FCNet. In addition, while the slice dataset has the largest proportion of good hyperparameters, it is not the case that the benefits of transfer learning is less evident; in fact, the improvement over random search is large and only second to the one observed on naval. As a result, we still believe that low RMSE are correlated with performance as our method works significantly better when the RMSE scores are low. The advantage of our method increases as the RMSE tends to zero, and the prior becomes perfect in the limit case.\\n\\n\\nWe also highlight that a key benefit of our method is the adaptivity to new tasks. As a result, when plain GP performs best on some datasets for XGBoost, our method is always at least the second best and beats all other transfer methods (Table 2).\\n\\n* \\\"Table 6 contains more datasets than Table 2. Why did you drop some tasks?\\\"\\n\\nWe kept only public datasets in the final evaluation that were inadvertently left in this table.\\n\\n* \\\"The aggregated results in Table 2 are nice but actually we are interested in the outcome after the search after a given budget. Can you add such a table?\\\"\\n\\n We included such tables at iteration 10, 50 and 100 in the appendix. In short, at the 10th iteration transfer learning methods, especially our CGP and Box RS, performed much better than GP. When looking at results at 50 and 100 iterations, CGP significantly outperforms all other transfer methods because of its improved adaptivity.\\n\\n* \\\"A discussion of related work on autoML for multiple objectives is missing.\\\"\\n\\nOne key novelty of our method is to use Copula transformation for transfer learning. This makes metrics in different units (such as time) comparable. While proposing another multi-objective BO method is out of the scope of the paper, we aimed to demonstrate how the Copula transformation can be seamlessly used in a multi-objective setting. Indeed, many existing multi-objective methods can potentially also be combined with our Copula transformation as an extension, which we believe is an interesting venue for future work. We made this point clearer in the paper.\\n\\nWe thank the reviewer and have improved the readability of the paper with the proposed suggestions.\"}", "{\"title\": \"Response to reviewer 2\", \"comment\": \"* \\\"In equation 1, the last term seems to be constant. [...] So why is it included?\\\"\\n\\nThe term $\\\\psi\\u2019(\\\\psi^{-1}(z))$ is not constant over $z$ and is needed when performing the change of variable. It weights more the error (and gradients) occurring when the quantile function of $y$ is changing rapidly. In other words, error in regions where the quantile function changes rapidly should cost more compared to flat regions.\\n\\n* \\\"Before equation 1, the text says 'by minimizing the Gaussian negative log-likelihood on the available evaluations (x, z)' But then, equation 1 is not the NLL on z but on y.\\\"\\n\\nThe NLL is minimized in $z$ and there is indeed no $y$ in equation 1.\\n\\n* \\\"Why modeling the residuals this way and not the observations themselves? The split of modeling between a parametric and non-parametric part is not justified.\\\"\\n\\nThis is in fact a key point of the paper. Modeling the residuals allows us to set a parametric prior that can transfer well across tasks by avoiding the numerical issues caused by different task and metric scales. Note that using a parametric prior also avoids the cubic complexity in the number of evaluations that makes it prohibitive to fit a GP on all task evaluations.\\n\\n* \\\"Given that the authors are trying to aggregate information about the optimal hyperparameters from several tasks, they should not compare with single-task approaches, but with the simplest way to combine all the tasks. For instance: a) Normalize the outputs of every task. [...] b) Collect the z of all tasks and feed them into an existing GP black-box Bayesian optimizer. This is a very simple way to get \\\"transfer learning\\\" and it's unclear that the extra complexities of this paper (copulas, changes of variable with proper renormalization when the transformation is parameter free, etc) are buying much else.\\\"\\n\\na) We asked ourselves the same question when designing our approach. For this reason a simple standardization to normalize tasks (where mean and std computed on each task separately) was evaluated in our ablation study in Section 5.2 and Table 3, where we probe the benefits of the Copulas and parametric priors against simpler alternatives. The results shows that standard normalizations are not competitive with our proposal, and these baselines perform poorly due to the scale issues met across different tasks and metrics. Given that this method was not seen in your first review, we renamed it and referenced it better in the ablation description to make it more visible.\\n\\nb) Following on the reviewer\\u2019s feedback, we ran additional experiments standardizing the outputs of every task and feeding these into warm-start GP. While this approach scales cubically in the number of observations, which makes it prohibitive with a large number of observations per task, our proposal of learning a prior does not have this limitation. Nonetheless, we applied this method by selecting the best hyperparameter evaluations from each related task for a total of 100 evaluations (so as to match the number of observations used in the warm-start GP baseline in the paper). The results show that our approach significantly outperforms this variant of warm-start GP, confirming as in Table 3 that standardization is not sufficient to transfer information successfully across heterogeneous tasks. We thank the reviewer for the suggestion and have incorporated this additional baseline into the paper.\\n\\n* \\\"One of the main points of the variable changes is to normalize the scales of the different tasks. However, equations 1 adds together the samples of the different tasks (which, as pointed out by the authors might have different sizes)... the different dataset sizes will bias the solutions towards larger datasets... In fact, this will have the same effect as if the cost had different scales for different tasks, which is precisely the problem that the authors are trying to avoid.\\\"\\n\\nIt is true that we do not normalize by the dataset size. Note that this is orthogonal to the scale issues we focus on: larger tasks will have larger gradient contributions but the scaling we propose still allows us to learn tied parameters across tasks as their scales are made similar. We thank the reviewer for the observation and revised the manuscript to discuss this option.\\n\\n* \\\"Is d really the number of hidden nodes of the MLP? Or the number of outputs? Given that d is also the size of w, it seems it's actually the latter.\\\"\\n\\nWe clarified in the main text that the number of outputs and nodes coincide.\\n\\n* \\\"Explain why the EI approach is used for the second model (with the GP), but not for the first model.\\\"\\n\\nAs the first model is stateless, the same hyperparameter minimizing EI would be sampled at each step (as opposed to the GP-based model). This is why we perform Thompson sampling instead.\"}", "{\"title\": \"Response to Reviewer 3\", \"comment\": \"We would like to thank you for the positive feedback.\"}", "{\"title\": \"Response to reviewers\", \"comment\": \"We thank all the reviewers for their valuable feedback and uploaded a new version of our paper addressing their comments. With respect to Reviewer 2's comments on baselines, we would like to highlight that a baseline with standard normalization was present in the first version of our manuscript to demonstrate the benefits of the Copula transformation. Following on the feedback, we added the proposed variant of warm-start GP using standardization as well as the warm-start GP baseline suggested by Reviewer 1. Our approach outperforms both these additional baselines. In addition, we discussed the given references and performed the extra analysis proposed by Reviewer 1 to improve the understanding of task distributions.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\nThis paper proposes a new Bayesian optimization (BO) based hyperparameter searching method that can transfer across different datasets and different metrics. The method is to build a regression model based on Gaussian Copula distribution, which maps from hyperparameter to metric quantiles. The paper shows that by leveraging this estimation using some specific (sampling) strategies, it is able to improve over other BO methods.\\n\\nThe high-level idea of this paper seems sound to me -- that improves standard BO to generalize across datasets and metrics by learning a mapping between the space of hyperparameters and metrics. While I am not an expert in this area, the derivation looks sound to me, and the evaluation results of this paper are comprehensive to show that CGP seems to outperform a number of methods.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper tackles the problem of black-box hyperparameter optimization when multiple related optimization tasks are available simultaneously, performing transfer learning between tasks. Different tasks correspond to different datasets and/or metrics. Gaussian copulas are used to synchronize the different scales of the tasks.\\n\\nI have several reservations with this paper. First and foremost, it seems to be lacking a fair and trivial baseline (I will describe it below) that justifies the apparently unnecessary complicated path followed in this paper. Second, there are a few small incorrect or improperly justified technical details throughout the paper.\\n\\n\\n1) Mistaken/unjustified technical details:\\n\\n- In equation 1, the last term seems to be constant. For each task, the function psi is not parametric, so its gradient is also not parametric and the input is the inverse of z, i.e., y, which is also fixed. So why is it included in the cost function? This sort of probabilistic renormalization is important in e.g. warped GPs because the transformation is parametric. In this case, I don't see the point. It can be treated as a normalization of the input data, prior to its probabilistic modeling.\\n\\n- Before equation 1, the text says \\\"by minimizing the Gaussian negative log-likelihood on the available evaluations (x, z)\\\" But then, equation 1 is not the NLL on z but on y.\\n\\n- In section 4.2 the authors model the residuals of the previous model using a powerful Matern-5/2 GP. Why modeling the residuals this way and not the observations themselves? The split of modeling between a parametric and non-parametric part is not justified.\\n\\n- One of the main points of the variable changes is to normalize the scales of the different tasks. However, equations 1 adds together the samples of the different tasks (which, as pointed out by the authors might have different sizes). Even if the scales of the outputs are uniform, the different dataset sizes will bias the solutions towards larger datasets. Why would that be a good thing? This is not mentioned and doesn't seem correct: there should not be a connection between a dataset size and the prior influence of the corresponding task. In fact, this will have the same effect as if the cost had different scales for different tasks, which is precisely the problem that the authors are trying to avoid.\\n\\n\\n2) Trivial baseline\\n\\nGiven that the authors are trying to aggregate information about the optimal hyperparameters from several tasks, they should not compare with single-task approaches, but with the simplest way to combine all the tasks. For instance:\\n a) Normalize the outputs of every task. This can be accomplished in the usual way by dividing by the standard deviation, or even better, by computing the fixed transform z = psi(y), separately for each task.\\n b) Collect the z of all tasks and feed them into an existing GP black-box Bayesian optimizer.\\n\\nThis is a very simple way to get \\\"transfer learning\\\" and it's unclear that the extra complexities of this paper (copulas, changes of variable with proper renormalization when the transformation is parameter free, etc) are buying much else.\", \"minor_improvements\": [\"Page 2: \\\"is the output of a multi-layer perceptron (MLP) with d hidden nodes\\\" Is d really the number of hidden nodes of the MLP? Or the number of outputs? Given that d is also the size of w, it seems it's actually the latter.\", \"Explain why the EI approach is used for the second model (with the GP), but not for the first model.\"], \"edit_after_rebuttal\": \"\\u201cThe term is not constant over z\\u201d -> Sure, it\\u2019s not constant over z. But z is constant. So the term is constant.\\n\\n\\u201cThe NLL is minimized in z and there is indeed no y in equation 1.\\u201d -> Sure, there\\u2019s no y in the equation, that\\u2019s correct. But it is still the NLL of y, and not the NLL of z.\", \"about_the_new_baseline\": \"Instead of simply renormalizing using mean and standard deviation, I suggested above using the same z=psi(y) that is used in the paper for the normalization. Is that where the advantage of the proposed method is coming from?\\n\\n\\\"Note that this is orthogonal to the scale issues we focus on: larger tasks will have larger gradient contributions but the scaling we propose still allows us to learn tied parameters across tasks as their scales are made similar. \\\" Both issues affect the scaling of the task, so I don't see how they can be orthogonal. Their scales are not made similar precisely because of the different sample sizes.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"The authors propose a new way of normalizing the labels of the meta-data. They propose a Thompson sampling strategy as a new hyperparameter optimization warmstarting strategy and an optimization method that leverages transfer learning.\\n\\n\\nThe transfer learning baselines are surprisingly weak and show no improvement over the random search baseline for two of your tasks. You should consider stronger baselines. One interesting work is \\\"Collaborative hyperparameter tuning\\\" by Bardenet et al. which overcomes the problem of different scales by considering the problem as a ranking problem. You discussed further works in your related work. Another interesting work based on GPs is \\\"Scalable Gaussian Process based Transfer Surrogates for Hyperparameter Optimization\\\".\\nThe warm-start GP currently seems to be your strongest baseline. However, I have doubts that it is implemented correctly. You say that you estimate the most similar dataset and then evaluate its best 100 hyperparameter configuration. First of all, I don't understand why you decided to choose 100 (DeepAR only has 220 configurations in total!). Second, this is not how this method works. Instead, you estimate the k most similar tasks and evaluate its best hyperparameter configuration. In your case k is upper bounded by 10 (number of tasks). This is notably smaller than 100 and gives more time to the Bayesian optimization which will likely improve your results.\\n\\n\\nYou observe worse results on XGBoost and good ones on FCNet. You refer to Table 6 and connect it to the RMSE. This might be true but I have a much simpler explanation: the search space of FCNet is orders of magnitudes larger and it contains many configurations that lead to very high MSE and only few with low. Therefore, a random search will on average provide poor results where a warmstarted search will obtain decent results for the first iterations. XGBoost is less sensitive to hyperparameters such that the overall variance in the losses is smaller. Maybe you can provide some insights into the complexity of the optimization tasks (plot the distributions )and add it to the appendix?\\n\\n\\nTable 6 contains more datasets than Table 2. Why did you drop some tasks?\\n\\n\\nThe aggregated results in Table 2 are nice but actually we are interested in the outcome after the search after a given budget. Can you add such a table?\", \"i_have_few_suggestions_to_improve_the_readability_of_the_paper\": \"That there is a choice of copula estimators is mentioned at the very end of the paper. Can you add it to the section where you describe them first? In section 5.1 you already argue by referring to Table 6. However, Table 6 is explained first in section 5.2 which makes it hard to follow your argumentation.\\nYou use the term \\\"metric\\\" to refer to objectives. This is confusing, you might consider changing this. You propose to use scalarization to address the multi-objective optimization problem. Why would an average of both objectives be the optimal solution? Does the unit you use to measure the time matter? What if you use machine learning algorithms that scale super-linear in the number of data points? How is this novel and why can't your baselines employ the same idea? A discussion of related work on autoML for multiple objectives is missing.\\nThe second paragraph of section 5 is confusing. You cite some work, discuss it and then conclude that your setup is entirely different. Would any information be lost if you say that you precomputed the values?\"}" ] }

[ "Mingyang Yi", "Huishuai Zhang", "Wei Chen", "Zhi-Ming Ma", "Tie-Yan Liu" ]

[ "adversarial training", "robustness", "separable data" ]

{ "note_id": [ "N9d8kA54k", "B1lw6TI3iS", "rkgmGXzusS", "rkxcy7zdoS", "Syevazz_or", "r1gP-GPSqr", "BygvfpMCFS", "Bkeuqk7NtS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731844, 1573838270752, 1573557003163, 1573556961778, 1573556927261, 1572332031213, 1571855630865, 1571200911563 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1761/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1761/Authors" ], [ "ICLR.cc/2020/Conference/Paper1761/Authors" ], [ "ICLR.cc/2020/Conference/Paper1761/Authors" ], [ "ICLR.cc/2020/Conference/Paper1761/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1761/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1761/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper studies adversarial training in the linear classification setting, and shows a rate of convergence for adversarial training of o(1/log T) to the hard margin SVM solution under a set of assumptions.\\n\\nWhile 2 reviewers agree that the problem and the central result is somewhat interesting (though R3 is uncertain of the applicability to deep learning, I agree that useful insights can often be gleaned from studying the linear case), reviewers were critical of the degree of clarity and rigour in the writing, including notation, symbol reuse, repetitions/redundancies, and clarity surrounding the assumptions made.\\n\\nNo updates to the paper were made and reviewers did not feel their concerns were addressed by the rebuttals. I therefore recommend rejection, but would encourage the authors to continue refining their paper in order to showcase their results more clearly and didactically.\", \"title\": \"Paper Decision\"}", "{\"title\": \"After Rebuttal\", \"comment\": \"I don't think the response addresses my questions, nor the paper is ready for publication.\\n\\nJust to list a few here. In the response Lemma 1 is used for multiple questions. Note that in Lemma 1 the dataset is assumed to be \\u201c\\\\eps strongly linear separable\\u201d. Why can it still be used in section 4? Since w* may not be unique, how do I guarantee the convergence of the algorithm? I don't think clipping can be seen as rescaling. Also, will the claims in the paper still hold given the clipping?\"}", "{\"title\": \"Thanks for your review, the following are our responses to your questions\", \"comment\": \"\\u201cFor linear classifiers, I do not know if there are existing work on their robustness when perturbation of samples are being trained on, but to be well-placed in the literature, the authors must either claim there is none, or cite those papers\\u201d\\nTo the best of our knowledge, theoretical analysis to the limited point of adversarial training only appeared in paper Ilyas et al., 2019 which we give a discussion between our difference in the last of section 4.1.2. The robustness of limited point is a hard problem, even under linear framework and standard training. Some recent works study it since 2017 i.e. Soudry et al. (2017), Ji & Telgarsky (2018). But this is the first paper to discuss in for adversarial training. \\n\\n\\u201cThat regular training is susceptible to outliers is proof by example and people already know that\\u2026\\u201d\\nThe standard training is susceptible to outliers is provided in proposition 1 and illustrated by example 1 and example 2. The stability of adversarial training is not build on the assumption \\u201cp_{i}^{1} has same scale with p_{j}^{2}\\u201d. On the contrary, we show that p_{i}^{1} has same scale with p_{j}^{2} in the first paragraph of section 4.1.2. This is because the limited point of adversarial training w^{*} has a finite norm, otherwise it will end up with a infinite loss due to the non-separability of data.\\n\\n\\u201cI am not very sure if I understand it correctly, but the logic seems to be that the logits are bounded\\u201d\\nEquation (16) can be viewed as the confidence e^{|w^{T}x_{k,i}|} will be upper bounded by the average loss among the opposite class (smaller than some constant). Our result is e^{|w^{T}x_{k,i}|} can be upper bounded rather than |w^{T}x_{k,i}|.\\nThe assumption of large \\\\epsilon is a technical assumption in order to give a quantitively description of confidence. A relatively smaller \\\\epsilon can also correspond with low confidence in practice.\", \"definition_2\": \"\\\"logit\\\" <--- people call this probability estimates; logits are wTx Sec. 4.1.2 needs to clarify what k in x_i^k means -- it's continued from proposition 1 which at first glance is irrelevant.\\n\\tWe will revise the notations accordingly to make this paper more friendly to read in the next version.\"}", "{\"title\": \"Thanks for your review, the following are our responses to your questions\", \"comment\": \"\\u201cSimulations are constructed to verify the arguments in the paper but there is no experiments on real dataset.\\u201d\\n\\tA series of experiments on CIFAR10 related to our results are delegated to appendix E.2.\\n\\u201cTheorem 1 assumes the \\\\eps strongly linear separable, and Theorem 2 assumes a large \\\\eps\\u201d\\nThe assumption \\u201c\\\\eps strongly linear separable\\u201d is equivalent to linearly separable which implies there must exist a \\\\epsilon>0 such that w^{*T}x_{i} \\\\geq \\\\epsilon \\\\|w^{*}\\\\|. We can accordingly adjust the adversarial radius, then it is a medium assumption. The assumption of large \\\\epsilon in theorem 2 tries to give a quantitively characterization to confidence of classifier obtained by adversarial training. A slightly small \\\\epsilon can also correspond with similar observation in practice. We will try to get rid of this technical assumption in the next version. \\n\\n\\u201cHow do you guarantee by \\u201crescale the norm of w\\u201d (from the footnote) to make sure that c_1 is not -\\\\infty\\u201d\\nWe should highlight that assumption only appears in theorem 1. The condition is used to ensure L-smoothness of adversarial loss l(w). For every classifier w, we can rescale w as w/\\\\|w\\\\| \\\\max{\\\\|w\\\\|, c} for some small constant c. Then, \\\\|w\\\\| can not equal to zero. Besides that, we can similarly clip the norm of w if w^{T}x_i - \\\\epsilon\\\\|w\\\\| smaller than some negative constant. Then we can ensure c_1 > -\\\\infty. As a matter of fact, we can omit the rescale procedure if the loss function l(u) is \\\\log{1+e^{-u}} or initial learning is smaller than 1/L(w(0)), where L(w(0)) is the local Lipschitz constant of point w(0). \\n\\n\\u201cLemma 1, this is for every x_i, or some particular x_i\\u201d\\nThe third equation in lemma 1 is for every x_i. It shows that w_t will converge to a point with zero adversarial training loss according to our assumption to the loss function l(u).\\n\\n\\u201cWhat is w(t). What is the condition on \\\\eta in Theorem 1? Why it is O(\\\\eta) in equation (11) or equation (24)\\u201d\\nw(t) is defined in equation (6), which is the gradient flow iterates of adversarial training. The behavior of gradient flow iterates can approximate the real iterates obtained by gradient descent w_t here. In theorem 1, we characterize the w(t) to reveal behavior of the real gradient descent iterates w_t. The error bound is given by the learning rate \\\\eta. Hence the O(\\\\eta) in equation (11) and (24) are estimation error between gradient flow iterates w(t) and gradient iterates w_t.\\n\\n\\u201cThe claims in section 4.1 seem to be depended on carefully designed examples. Would it still true rigorously for general cases?\\u201d\\nThe two examples are used to illustrate the unstability of standard training, a general description is presented in proposition 1. The stability of adversarial training is delegated to equation (15).\\n\\n\\u201cIn section 4.1.2 first paragraph, why \\\\|w_t\\\\| can not go to infinity? in the third paragraph, how assumption 2 implies p^k_i / p^k_j = o(1), or later p^k_i / p^k_j = O(1)?\\u201d\\nAssumption 2 informs that \\\\lim_{u\\\\to\\\\infty} l(u) = 0, since we assume the data are not separable, then there exists a x_i such that w^{T}x_i - \\\\epsilon \\\\|w\\\\|\\\\leq 0 for each w. The iterates w_t will converge to a minimum according to lemma 1, while \\\\|w_t\\\\| goes to infinity will rends w^{T}(t)x_i - \\\\epsilon \\\\|w(t)\\\\| goes to -\\\\infinity. Then we will ends up with infinite loss. A more detailed description is referred to paragraph 1 in section 4.1.2. \\nWithout loss of generality, we use l(u)=e^{-u} as an illustration. p^{k}_{i} = e^{-w^{*T}x_{i}^{k} + \\\\epsilon\\\\|w^{*}\\\\|}. Then p^{k}_{i} / p^{k\\u2019}_{j} = \\\\exp{-w^{*T}(x_{i}^{k} \\u2013 x_{j}^{k})}. The minimum w^{*} has infinite norm which makes the loss goes to zero. Then, p^{k}_{i} / p^{k\\u2019}_{j} = o(1), since x_{i}^{k} is non-support vector and x_{j}^{k\\u2019} is support vector (w^{*T}(x_{i}^{k} \\u2013 x_{j}^{k})<0). But for adversarial training \\\\|w^{*}\\\\| is not infinity then p_{i}^{k} will have the same scale for each I and k.\\n\\n\\u201cIn section 4.2 second paragraph, what is \\u201ck_th category\\u201d?\\u201d\\n\\tIt represents the data is from the first or the second class.\\n\\n\\u201cIs w^* unique in Theorem 2?\\u201d\\n\\tw^{*} is the hard margin solution, it can be non-unique.\"}", "{\"title\": \"Thanks for your review, the following are our responses to your questions\", \"comment\": \"\\u201cHere \\\"fast\\\" is not the standard 1/T fast rates but, rather, a rate of o(1/log T)\\u201d\\nWe try to reveal the actual convergence rate of adversarial training to robust solution rather than developing a faster algorithm to obtain a robust solution. Hence, we think the result is valuable, even though the promotion is not surprising large to O(1/T).\\n\\u201cthe authors decided to use x_i to replace the product y _i x_i in order to \\\"simplify\\\" notation.\\u201d\\nWe would like to clarify the notations here and accordingly revise them in the paper. First, representing y_i x_i by x_i is out of the consideration that x_i and y_i are always appear as x_i y_i and w^{T}x_i y_i means the data is correctly classified. The symbol will not mislead the notations. We will substitute the \\\\epsilon in definition 2 with some other symbol. \\n\\u201cThird, some of the mathematical results do not make sense\\u201d\\nThe two conditions in definition 1 are used to emphasis that w^{*} can not only make a correct classification for each data but also ensure each data is away from the margin larger than \\\\epsilon. We will get rid of the first condition in the next version. The condition that union data is linearly separable is necessary to make proposition meaningful. We will then add it. The example 1 is used to give an intuitively explanation to proposition 1. It can be concluded as the hard margin solution can be sensitive to outliers closely with each other, but standard training can converge to hard margin solution which informs the unstability of standard training. The equation (12) is a direct induction to reveal the x_{i}^{1} will locate closely to the new hard margin solution (\\\\hat{w}_{1}, \\\\hat{w}_{2}) due to some outliers \\\\x_{i}^{2} are added, while they locate away from their original hard margin solution (\\\\hat{\\\\w}_{1}, 0). \\n\\u201cForth, the authors make many assumptions about the loss without mentioning one example that satisfies them\\u201d\\nWe enumerate some loss functions satisfy our assumptions in the footnote of page 3. For example, choosing loss function l(u) as e^{-u} or \\\\log{1+e^{-u}} can satisfy with our assumptions. Also please notice that assumption 2 exponential tail focuses on the behavior of loss function l(u) when u is large. We chose l(u)=\\\\log{1+e^{-u}} which will closely to e^{-u} when u is large. A simple inequality that \\\\log{1+x} \\\\geq x \\u2013 x^{2}/2 can give the exact conclusion. \\n\\u201cHow did the authors arrive at Eq 4\\u201d\\nIt is a directly result according to the linear classifier and monotonically decreasing property of loss function l(u).\\n\\n\\u201cunder a very limited setting (homogenous linear classifiers under a strong separability assumption)\\u201d\\nOur assumptions about loss function can be achieved by the most generally used loss functions i.e. e^{-u}, \\\\log{1+e^{-u}}. Also about the assumption \\u201cstrongly separability\\u201d, it is equivalent to the data are linearly separable, because if the data are separable, then there must exist a \\\\epsilon>0 such that w^{*T}x_{i} \\\\geq \\\\epsilon\\\\|\\\\w^{*}\\\\|. Then the data can satisfy our \\u201c\\\\epsilon-strongly linearly separable\\u201d. Hence, our core assumption is the data are themselves linearly separable. Besides that, we also give a discussion to adversarial training on non-linearly separable data.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The aim of this paper is to provide a theoretical analysis of adversarial training under the linear classification setting. The main result states that, under many technical assumptions, adversarial training using gradient descent may converge to the hard margin SVM classifier with a fast rate. Here \\\"fast\\\" is not the standard 1/T fast rates but, rather, a rate of o(1/log T) (in comparison to recent results that looked into the convergence of gradient descent with logistic loss to the hard-margin SVM solution).\\n\\nOverall, the paper is not recommended for publication for many reasons. \\n\\nFirst, the notation used is sometimes imprecise and in some cases it is entirely wrong. For example, the authors decided to use x_i to replace the product y _i x_i in order to \\\"simplify\\\" notation. This makes things really hard to follow. The authors need to use a different symbol, such as z to stand for the product yx. Second, some equations do not appear to be correctly typed (e.g. in Page 1, standard learning should have xi not x). Third, the authors use epsilon to denote two different things (one for the definition of linearly separable data and one for the robustness radius), and so on. \\n\\nSecond, the paper needs to be proof-read. It has a lot of typos and grammatical errors that make sentences difficult to understand. Examples include: \\n- \\\"while there several outliers are not or even not linearly separable\\\", \\n- \\\"a simple generalization error bound informs the high loss on test set\\\"\\n\\nThird, some of the mathematical results do not make sense. For example, Definition 1 has two conditions, the first one is immediately satisfied once the second condition is satisfied, so why both? Also, in Proposition 1, the authors should mention that both sets are subsets of a linearly separable superset. In that case, the conclusion of Proposition 1 is obvious. Definitely, if data are linearly separable, then the hard margin will shrink as more examples are added (it cannot increase by definition of \\\"maximum\\\" margin). Moreover, in Example 1, the authors conclude with an inequality of norms and I don't see how this follows from the description of the example. The example is generic and there is nothing that indicates one norm would be larger than the other. \\n\\nForth, the authors make many assumptions about the loss without mentioning one example that satisfies them. In fact, the loss function they used later in the experiments violates Assumption 2.\", \"some_additional_comments\": \"- How did the authors arrive at Eq 4? I don't see how this follows from the assumptions. Can you please elaborate? \\n- I am not aware of any book written by Vapnik in 1995 called \\\"Convex Optimization\\\". I think the authors meant the SVM paper. \\n\\nGiven all of these issues and the fact that the main result is incremental and holds under a very limited setting (homogenous linear classifiers under a strong separability assumption) and relies on very strong assumptions about the loss function that may not be achievable to begin with, I do not recommend acceptance. \\n\\n=========== \\n#Post Rebuttal Remarks\\n\\nRegarding the rate of convergence, just to be clear, I was only clarifying what the term \\\"fast\\\" meant in the paper, not complaining. So, this did not affect my score. The notation issues need to be fixed and the authors need to ensure that mathematical equations are written precisely, as stated in my review. \\n\\nThanks for clarifying the issue about the loss functions. I agree that the example they mentioned satisfy it but it is not a common loss used in practice.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper provides some analyses of the difference between adversarial training and standard training for linear classification problem. In particular, it proves that when the data is \\\\eps linearly separable, adversarial training converges faster than standard trading. It also argues that when the data is not \\\\eps linearly separable, adversarial training is more robust to outlier. Simulations are constructed to verify the arguments in the paper but there is no experiments on real dataset.\\n\\nThe first result of this paper is interesting, that adversarial training converges faster than standard training. Studying the difference between the convergent points between adversarial training and standard training is also an interesting research problem. However, I still have two main concerns about the current version of the paper.\\n1. The paper is trying to develop rigorous results, but its writing is arguably not rigorous. Many statement are not clear and some notations are used without definition. Section 4.1 has many vague statements. See more concrete comments below. \\n2. I am not sure about the significance of the results in the paper. The results highly depend on the linear setting with convex losses. More than that, Theorem 1 assumes the \\\\eps strongly linear separable, and Theorem 2 assumes a large \\\\eps (if the statement is that |w* x_{k,i}| is less than a large number, it seems much less interesting). These are very strong assumptions that are usually not true in practice. Experimental results only cover carefully designed simulation as well.\", \"detailed_comments_for_item_1_above\": \"1. Assumption 3, what is the quantifier for w? Is it for every w? There exists some w? How do you guarantee by \\u201crescale the norm of w\\u201d (from the footnote) to make sure that c_1 is not -\\\\infty?\\n2. Lemma 1, this is for every x_i, or some particular x_i?\\n3. What is w(t)?\\n4. What is the condition on \\\\eta in Theorem 1? Why it is O(\\\\eta) in equation (11) or equation (24)?\\n5. The claims in section 4.1 seem to be depended on carefully designed examples. Would it still true rigorously for general cases? \\n6. In section 4.1.2 first paragraph, why ||w_t|| can not go to infinity? in the third paragraph, how assumption 2 implies p^k_i / p^k`_j = o(1), or later p^k_i / p^k`_j = O(1)?\\n7. In section 4.2 second paragraph, what is \\u201ck_th category\\u201d?\\n8. Is w^* unique in Theorem 2?\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"TL;DR: The paper gives interesting, theoretical results to adversarial training. The paper only uses linear classifiers, which are hardly the same problem as deep networks where adversarial attacks are problematic. Some conclusions from theorems can be vague or informal, and therefore are not very convincing. I vote for rejecting this paper since it is hard to claim it informs deep learning research (the motivating reason for doing adversarial training). However, I am not familiar with theoretical analysis of adversarial attack/defense, so I am open to counter-arguments.\\n\\n=================\\n1. What is the specific question/problem tackled by the paper?\\nThe paper gives a theoretical analysis to the theoretically less-studied procedure of adversarial training, and shows properties of adversarial training in comparison to regular training, for both linearly separable data or inseparable data. The paper sheds light on some empirical behavior of adversarially trained networks, namely that they are more robust to outliers and lower in performance.\\n\\n2. Is the approach well motivated, including being well-placed in the literature?\\nI am not an expert of adversarial samples, so I am ill-equipped to judge the novelty of the paper. The research direction itself is well motivated, and in the realm of deep learning, it is posed as a first paper to theoretically analyze adversarial training.\\nHowever, the authors only analyzed linear classifiers. This makes the results of the paper ill-suited for deep networks, whose non-linearity is arguably the reason why adversarial samples are such a problem. The motivation of the paper is thus greatly diminished. For linear classifiers, I do not know if there are existing work on their robustness when perturbation of samples are being trained on, but to be well-placed in the literature, the authors must either claim there is none, or cite those papers.\\n\\n3. Does the paper support the claims? This includes determining if results, whether theoretical or empirical, are correct and if they are scientifically rigorous.\", \"claims_and_novelties_in_this_paper_include\": \"(1) Adversarial training converges faster than regular training if samples are \\u03b5-strongly linearly separable,\\n(2) If samples are not \\\"\\u03b5-strongly linearly separable\\\", adversarial training is robust to outliers, while regular training is not,\\n(3) Confidence is low for all (training) samples if \\u03b5 is large. \\n\\nOnly (1) seems to be sufficiently proved. I am not certain that this is a very useful result, and I am open to counter-arguments.\\n(2) and (3) have steps that are vague and informal:\\n\\n(2) That regular training is susceptible to outliers is proof by example and people already know that. Also the claim relies on the assumption that pi^1 and pj^2 are on the same scale, while those samples that violate the decision boundary can have arbitrarily large pi^1 or pj^2. Since outliers are often the violators, and inliers often are not, a small number of carefully placed outliers can make ||p2|| quite large while each pi^1 can be very small. It is also worth noting the logic seems to boil down to \\\"N1>>N2, so inliers should overwhelm outliers, making the training robust\\\". The claim is not guaranteed.\\n\\n(3) I am not very sure if I understand it correctly, but the logic seems to be that the logits are bounded, so they cannot be too large, and so the confidence is low. However, the bound also involves the magnitude of w and x from the other class, so the final step of the proof is either unclear or the bound can indeed be quite large. Note that |wx| does not need to be very large for the confidence to be high (e.g. if logit is 5, the confidence is 1/(1+e^-5)=99.3%). The claim also relies on the assumption that epsilon is larger than the distance between the farthest points in the dataset, which is extreme since you can find an adversarial sample that can be considered to be simultaneously \\\"close enough\\\" to the two most dissimilar samples in the dataset.\\n\\nIn the end, the results are not very convincing or useful for informing deep learning research.\\n\\n=============\", \"to_improve_paper\": [\"Clarify motivation and how this would inform adversarial training highly non-linear classifiers;\", \"Add related work for robustness to perturbation of linear models, or state that they don't exist;\", \"Clarify weaknesses in the claims.\"], \"editorial_changes\": \"\", \"definition_2\": \"\\\"logit\\\" <--- people call this probability estimates; logits are wTx\\nSec. 4.1.2 needs to clarify what k in x_i^k means -- it's continued from proposition 1 which at first glance is irrelevant\\n\\n\\n=================\\nPost rebuttal\\n\\nApologies for not interacting earlier due to deadlines. \\nThe rebuttal does not address my major concern (motivation), nor does it discuss its relationship with related work. The math questions are not answered very clearly; I do not see how section 4.1.2 proves p_i^1 has the same scale as p_j^2, except maybe that they are all smaller than some constant for certain samples. And other discussions in the rebuttal simply confirms my concern. \\nIn summary, I think this paper is not yet ready for publication.\"}" ] }

[ "Jamie Hayes", "Krishnamurthy Dvijotham", "Yutian Chen", "Sander Dieleman", "Pushmeet Kohli", "Norman Casagrande" ]

[ "watermarking", "provenance detection" ]

{ "note_id": [ "KKPWxNnVBQ", "rJe59wI7sS", "H1gouI8QoH", "BygcMIUmjr", "SyeSWv57cB", "rklOwGk0YH", "ryeRqiZaKH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731816, 1573246866009, 1573246579352, 1573246481625, 1572214525133, 1571840608054, 1571785621712 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1760/Authors" ], [ "ICLR.cc/2020/Conference/Paper1760/Authors" ], [ "ICLR.cc/2020/Conference/Paper1760/Authors" ], [ "ICLR.cc/2020/Conference/Paper1760/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1760/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1760/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper offers an interesting and potentially useful approach to robust watermarking. The reviewers are divided on the significance of the method. The most senior and experienced reviewer was the most negative. On balance, my assessment of this paper is borderline; given the number of more highly ranked papers in my pile, that means I have to assign \\\"reject\\\".\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for your detailed review\", \"comment\": \"We thank the reviewer for their review and detailed feedback. We will attempt to address each of the highlighted concerns.\\n\\n1. Re: how the detector and watermarking mechanisms are jointly obtained: We apologize for the confusion around this. Indeed, in our scheme the watermark generation and watermark detection are jointly optimized. We only refer to a fixed watermark below Definition 2.1 as an instructive motivating example for how we can train the detector, before explaining how the joint optimization is solved in Section 3. We can think of the game of learning both a watermark and a watermark detection model as adversarial training, where the classification task is to recognise the watermark. At each step of training, given an input, x, we find a perturbation (via gradient descent) that would cause the detector model to classify this input as watermarked, we then add this perturbation to the input, x+\\\\delta, and then update the detection model based on how strongly it classified x as not watermarked, and classified x+\\\\delta as watermarked. \\n\\n2. Re: motivation: The reviewer is entirely correct that deep fake creators would not want to add a watermark to their creations. Our threat model is designed to protect an individual or organization against abuse of data or media released by them. For example, consider an organization that has a corpus of photos/audio samples/videos, or a generative model that they allow the public access to query. These samples could then be used to create deep fakes by another party. If the organization has watermarked these samples with our scheme, the watermark is designed to be robust to post-processing transformations that an adversary might introduce, and so the watermark will persist when the deep fake creator manipulates the sample. The organization can then verify if the deep fake was originally created using samples that they own.\\n\\n3. Re: transferability: An attacker that has access to content that contains both non-watermarked and watermarked versions, is currently outside of our threat model. As far as we can tell, this is a relatively uncommon assumption in zero-bit watermarking, as if an attacker has a procedure to identify if a piece of content was or was not watermarked they have practically already broken the scheme. Furthermore, to increase the cost of this detection attack, in potential practical applications an organization can limit the exposure to watermarked content by prohibiting any single party from making a large number of queries to access content. Our experiments in Section 4.2.2 are designed to show that an attacker that does not have this information, but has other information about the watermarking procedure, cannot break the scheme. The attacker has full knowledge of the architecture of the detector model, the algorithm that is used to create watermarks and the detector, hyperparameters used in the algorithm, and the exact training set used, and so represents a relatively strong attack.\\n\\n4. Re: certified robustness: Section 4.2.3 attempts to show that a detector model that has been trained with post-processing transformations (eq. 1) is more robust than a model trained without post-processing transformations (eq. 1 with max over T omitted from the formulation). The certification procedure provides a certified region of robustness to Gaussian noise (correspondingly an L2 norm ball). Since Gaussian noise is included in the set of post-processing transformations, the model that has been trained to be robust to these transformations is naturally more robust to Gaussian noise (and so has a larger certified area with respect to the L2 norm) than the model that has not been trained on these set of transformations.\\n\\n5. Re: Validity of comparing with Broken Arrows: Please refer to our response to Blind Reviewer 1 regarding the concern of the comparison with Broken Arrows.\\n\\n6. Re: No numerical results in Section 4.2.4: The results on the SVHN test set is a single number, reporting that no mistakes were made by the detector. Results on BigGAN also had a false positive and false negative rate of <1.2% when attacked with various post-processing transformations. We have updated the paper now detailing these results.\\n\\nThank you again for the helpful review, we would be more than happy to act on any specific recommendations that the reviewer would consider important to revising the score for this paper.\"}", "{\"title\": \"Thank you for your comments\", \"comment\": \"We thank the reviewer for their careful review and thoughtful feedback. We agree that an exciting direction for future work would be to compare this scheme to other key-based watermarking schemes. In this work, we decided to focus on zero-bit watermarking as it most naturally aligns with the goal of provenance detection. However, it is true that the scheme could be extended in the future to multi-bit watermarking schemes and in that case it would be natural to compare against other key-based multi-bit watermarking schemes. Please refer to our response to Blind Reviewer 1 regarding the concern of the comparison with Broken Arrows.\"}", "{\"title\": \"Thank you for your comments\", \"comment\": \"We thank the reviewer for their careful review and thoughtful feedback.\\n\\nA concern was raised in the reviews that since Broken Arrows was proposed ten years ago, it might not reflect the state of the art. We had performed an extensive literature survey of zero-bit watermarking algorithms during our research. We used Broken Arrows to compare against our method for the following reasons:\\n1. To the best of our knowledge, no publicly available zero-bit watermarking scheme has been published that claims superiority over Broken Arrows and is still used as the state-of-the-art watermarking method by other works. For example, Quiring et al. (2018) [1] used Broken Arrows in their recent work (see Section 4.1 in [1]).\\n2. Broken Arrows won an international competition designed for defending against watermarking attacks [2]. \\n3. It has provable robustness to Gaussian noise attacks (see Furon & Bas (2008) [3]).\\n\\n[1] Quiring, Erwin, Daniel Arp, and Konrad Rieck. \\\"Forgotten siblings: Unifying attacks on machine learning and digital watermarking.\\\" 2018 IEEE European Symposium on Security and Privacy (EuroS&P). IEEE, 2018.\\n[2] http://bows2.ec-lille.fr/\\n[3] Furon, Teddy, and Patrick Bas. \\\"Broken arrows.\\\" EURASIP Journal on Information Security 2008.1 (2008): 597040.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper is about a novel method to add watermarks to images and audio that is highly robust to several transformations that is closely related to gan methods. The idea is that the watermark signal is learned concurrently to the detector network, which share similarities to a generator and detector networks. Five standard attack transformations are considered and a specific optimization to reduce the transferability of the watermark is considered. The method is compared against Broken arrows on Cifar10 and Imagenet. It shows similar or better performance for Gaussian noise attack given the same amount of perturbance allowed in a signal while very much better performance for the other attacks. Going beyond the five attacks, the paper also includes an estimation of the probability of confidence of finding watermarked images given a fixed l2 norm radius. Finally the method is also tested on audio on a proprietary dataset with a deepspeaker architecture which still shows very good performance and it is confirmed by a human evaluation where participants found the watermarked audio not significantly worse or degraded.\\n\\nThe paper is well written and of excellent presentation. The proposed method is novel and goes in the interesting direction of learning watermarks using adversarial training techniques. Experiments shows that the method has good performance but it is only compared to Broken Arrows (i.e. a zero bit watermarking) which is the state of the art of this type of watermarking. It would have been interesting a comparison to other key based watermarking methods to also deeply evaluate the signal transformation attacks. All in all, I found the paper to be significant and I would like to see it accepted.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"The authors introduce ReSWAT, a method for transformation-resilient watermarking of images via adversarial training. The high level idea is to learn a watermark/detector pair (W,D). W can be any transformation (in this paper, an l-infty bounded perturbation) that imputes an imperceptible distortion to a given input, while D is a detector that distinguishes watermarked from non-watermarked images. There is an additional requirement that the detector should be robust to simple transformations such as rotations, cropping, flipping, and contrast enhancement.\\n\\nThe authors pose a min-max style learning problem for learning (W,D) that leads to a natural adversarial training scheme. In particular, it can be viewed as creating an adversarial defense to the Expectations-over-transformations attack of Athalye et al. Experimental results on a bunch of different datasets confirm that the method works,\\n\\nThe paper is well-written and the contributions are clear. In terms of conceptual or theoretical novelty, the paper is limited (the method essentially boils down to regular adversarial training with a slightly non-standard loss function) but the connection to watermarking seems novel and is nicely executed.\\n\\nMy only concern is the lack of comparisons with baselines. I am not a digital forensics expert -- so I don't know what the state of the art is -- but the only comparisons made are with the Broken Arrows (BA) watermarking , which seems to be over 10 years old, so I am not sure how to evaluate the results.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper proposes a watermark design algorithm that is based on an adversarial training paradigm where a watermark detector and a watermark generator are jointly trained to maximize the likelihood of watermark detection while bounding the manipulation of individual pixels.\\n\\nI have several concerns about the clarity of the paper, its applicability to the motivating problem, and the completeness of the results that motivate my \\\"reject\\\" recommendation.\\n\\nThe description does not make it clear how the detector and watermark mechanism are jointly obtained in (1) - intuitively, a detector should know the type of watermark to be observed in order to pose a feasible detection model. The discussion focuses on \\\"fixed detector\\\" and \\\"fixed watermark\\\" cases only, but it is not clear how an iterative optimization would be initialized, and no discussion of how to set up the problem appears in the manuscript. \\n\\nThe motivation for the paper is also unclear. The authors discuss the need to be able to detect deep fakes, which is a relevant and meritorious problem. However, it seems that the application of the proposed approach in this setting would require deep fake creators to add watermarks to their creations, which would be a naive assumption. Watermarks are used to demonstrate provenance, but deep fake creators would not want to have such provenance verified.\\n\\nThe watermark transferability premise seems to rely on the random initialization of the underlying deep learning watermark detector. One could posit that if a third party is interested in creating false positives, they could train the deep learning network using a database of watermarked and non-watermarked images. Another concern here is that the formulation provided in (3) would potentially lead to low transferability for the specific classifiers provided there, but one does not know which classifier an adversary would construct - or why an adversary would construct an adversary that does not have knowledge of the specific watermark being used, or of samples of the watermark. It would have been interesting to see if it is possible to construct a good watermark detector this way (given enough training data being available). It would also be good to have some intuition as to why this robustness formulation also improves \\\"certified robustness\\\".\\n\\nFinally, the performance comparison only considers a 10-year old watermarking algorithm; some discussion as to why this is sufficient should have been included.\\n\\nMinor comments\\nThe norm infinity subindex is missing from the minimization conditions in (1).\", \"figure_3_caption\": \"watermark's -> watermarks\\nIt is not clear why no numerical results were given for Section 4.2.4.\"}" ] }

[ "Lanxin Lei", "Zhizhong Li", "Xiaoyang Li", "Cong Qiu", "Dahua Lin" ]

[ "Reinforcement Learning", "Regulatory Focus", "Promotion and Prevention", "Exploration" ]

{ "note_id": [ "teAI1SstlZ", "r1eAZ3hCFB", "SkxnCx6Ytr", "H1xOY14FtB" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731786, 1571896326305, 1571569875735, 1571532672358 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1759/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1759/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1759/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors take inspiration from regulatory fit theory and propose a new parameter for policy gradient algorithms in RL that can manage the \\\"regulatory focus\\\" of an agent. They hypothesize that this can affect performance in a problem-specific way, especially when trading off between broad exploration and risk. The reviewers expressed concerns about the usefulness of the proposed algorithm in practice and a lack of thorough empirical comparisons or theoretical results. Unfortunately, the authors did not provide a rebuttal, so no further discussion of these issues was possible; thus, I recommend to reject.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper presents a modification to policy gradient methods that are computed from advantage function estimates. For a given trajectory of n steps, there are n different advantage function estimates: from 1-step to n-step. GAE (Schulman, 2016) proposes to take an exponentially weighted average of these estimates to compute the policy gradients. This paper proposes instead to use order statistics to compute the policy gradient; e.g. the most optimistic estimate, the most pessimistic estimate, or the most extreme estimate. The paper introduces a regulatory ratio: the probability of using the averaged advantage estimate vs using the order statistic, for computing the policy gradients. This hyper-parameter is justified on the optimistic case (max advantage), as a way to prevent overtly optimistic estimates. The paper conducts experiments on different domains (sparse and dense rewards, discrete and continuous actions, fully observable and partially observable environments) which show the effect of choosing different order statistics and regulatory ratio on the policy performance.\\n\\nThis paper could be accepted as it presents an interesting idea with extensive experiments showing where it works and where it fails, along with some justification for the hyperparameter choices. But there are a couple concerns about the validity of the method.\\n\\nOne concern is that the regulatory ratio is only justified for the max case, but not for the min or max-abs case. In addition, the choice of regulatory ratio seems to have a wildly varying impact depending on the choice of the order statistic and environment, for which we get little insights from the paper.\\n\\nAnother concern is that most of the results are reported only for an ensemble of 4 n-step estimators. In the appendix, the papers reports a comparison with a larger ensemble (12 n-steps estimators), which results in lower performance. This is a bit confusing: using the order statistics (max, min and max-abs), and following the reasoning presented in the paper, I would expect that increasing the ensemble size would result in better performance ( the max/min of the 4 ensemble is n upper/lower bound of the max/min of the 12 ensemble). This paper could be improved with more detailed results on the effect of the ensemble size. While the paper provides arguments for why small ensembles might suffice, it does not explain how the ensemble should be chosen, and what would happen as the ensemble size increases.\\n\\nFinally, the parallels to human psychology are bit superfluous. I understand how it might serve as an inspiration for algorithm design, but it is a bit distracting from the technical contribution of the paper.\", \"things_to_improve\": \"Why is max-abs missing from Figure 2?\\nDid you try a Rainbow-style training (changing the regulatory focus periodically over training)?\\n\\nIn some sentences, it looks as if insure was written in place of ensure. In general, this paper would benefit from proof reading by a proficient English speaker.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper studies the problem of advantage estimation for actor-critic RL algorithms. The key observation is that the advantage can be computed using 1-step returns, 2-step returns, etc. The paper suggests that, instead of choosing a fixed n, we should aggregate these advantageous together. If the maximum is taken, the resulting policy will be exploratory (i.e., have a \\\"promotion focus\\\"); if the minimum advantage is taken, the resulting policy will be risk sensitive (i.e., have a \\\"regulatory focus\\\").\\n\\nThe paper presents results on a number of tasks. A few toy examples show instances where the proposed method works better. Experiments on sparse reward environments show that taking the max advantage provides for exploration that outperforms a baseline (EX2). On a walking talk show that the min-approach can outperform state-of-the-art on-policy RL (PPO); the paper suggests this is because the min-approach is implicitly risk sensitive. Experiments on some of the Mujoco control tasks and some of the Atari games show that taking the advantage with the maximum absolute value performs well, as compared to PPO.\\n\\nOverall, I think the main contribution of this paper is the finding that there exist smarter ways of computing the advantage. A second contribution is connecting the ideas of risk-sensitive and risk-seeking control with ideas from psychology (regulatory fit theory). \\n\\nI am leaning towards rejecting this paper. On the one hand, the paper is well written, well motivated, and the experiments quite thorough. On the other hand, I don't think the paper is particularly useful, either from a theory or algorithmic perspective. It is not surprising that if we add an additional hyperparater to an existing RL algorithm and tune that hyperparameter, we'll do better than the existing RL algorithm. The ablation experiments in Fig 6 and Fig 7 suggest that how the advantage estimates are combined, and how they are \\\"mixed\\\" with a standard advantage estimate, matter a lot. Further, the fact that using a smaller number of advantage estimates worked better (point #2 on pg 5, Effect of Ensemble Size in Appendix A) suggests that the ensemble size is an important hyperparameter, and that risk-seeking / risk-aversion (i.e., regulatory vs promotion focus) cannot alone explain why the proposed method works. I think that, for this idea to be useful, it must be equipped with a fixed value for these hyperparameters that works well across a wide range of tasks (e.g., a learning rate of 3e-4 works well for Adam on most tasks), or an automated strategy for choosing this hyperparameter (e.g., the automatic entropy tuning in SAC). I would consider increasing my review if either of these were included.\\n\\nA second concern is that, despite the close connections with risk-sensitive and risk-seeking control (discussed in Section 2), none of these prior works are compared against. Many of these prior works include a temperature parameter for trading off risk-seeking vs risk-aversion (e.g., \\\\beta in Eq 11 of [Mihatsch 2002]), which is arguably a more transparent (to the user) and easier to analyze (for the RL researcher) than the order statistics used in this paper. Given the wealth of prior work on risk-sensitive and risk-seeking control, I think it's important to know whether these prior methods already solve the problem at hand (choosing between risk-seeking and risk-aversion). I would consider increasing my review if one of these methods were included as a baseline.\\n\\nMinor comments\\n* \\\"that humans own\\\" -> \\\"that humans' own\\\"\\n* The max strategy seems quite closely related to UCB-based exploration. I'd be curious to learn about some discussion on the similarities/differences\\n* Fig 6 -- Can you add error bars to show the variance across random seeds?\\n* Appendix D -- How does this approach and the results shown in Fig 9 differ from standard GAE? \\n\\n----------------------- UPDATE AFTER (NO) AUTHOR RESPONSE ----------------------------\\nThe authors did not post a response, so I will maintain my vote to \\\"weak reject\\\" this paper. I would encourage the authors to incorporate the feedback in all reviews and submit the paper to a future conference.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper focuses on risk-aware reinforcement learning, where an agent could be encouraged to take more risk (high reward, high variance states) or avoid risk (low variance states). Risk control is instantiated by different ways of estimating the advantage of a state (max/min instead of average). Experiments on several environments show good performance of the proposed algorithm.\\n\\nThe paper is written clearly and the approach is straightforward. However, it's unclear what is the objective function the algorithm is optimizing by using a biased estimation of the advantage. It seems to work pretty well in practice, but I wonder how it compares to other risk-sensitive RL algorithms (e.g. those cited in the related work section).\\n\\nOverall, this paper presents a simple heuristic to steer the policy towards risk-seeking / risk-avoiding directions, but could benefit from either more theoretical analysis or more empirical comparison with other methods.\"}" ] }

[ "serrurier Mathieu", "Loubes Jean-Michel", "Edouard Pauwels" ]

[ "fairness", "adversarial networks", "properties", "wasserstein adversarial networks", "major topic", "machine learning", "questions", "context", "deep learning" ]

{ "note_id": [ "ypqPVEMrjG", "rkgtvbuH9S", "BJerxTNAYr", "HygO3_32YH" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731757, 1572335969366, 1571863788945, 1571764400277 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1758/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1758/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1758/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper presents an approach to enforce statistical fairness notions using adversarial networks. The reviewers point out several issues of the paper, including 1) their approach does not provably enforce criteria such as demographic parity, 2) lack of novelty and 3) poor presentation.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a variant of adversarial learning to achieve some of the popular group fairness definitions. The main novelty is the idea of minimizing Wasserstein distance between the conditional distributions of classifier predictions given different values of the protected attribute.\\n\\nMy main concern is the approximation of a simple 1d Wasserstein distance with a neural network. Wasserstein distance between two discrete distributions in 1d can be computed in closed form (simple function of order statistics). That is, eq. (1) is simple to evaluate for two empirical distributions. There is no need to use a neural network for approximation, and even if authors choose to do so, some discussion on how well it approximates actual Wasserstein distance is needed. I think the proposed algorithm could be more interesting if authors can work out the optimization problem with the actual Wasserstein distance.\\n\\nOn the theoretical/motivation side, it is not enough to say that demographic parity is achieved when the corresponding Wasserstein distance is 0. What is needed is that demographic parity difference is bounded from above by the corresponding Wasserstein distance (I don't know if it is true or not, but would like to know). Then minimizing Wasserstein distance to achieve demographic parity could be justified.\\n\\nFinally, the paper is quite poorly written. The description of fairness in the introduction is very vague. Authors essentially describe demographic parity as fairness, while it is simply one of the several definitions of group fairness. There is also individual fairness (the paper by Dwork et al. is cited, but not properly discussed) and prior work emphasizing certain deficiencies of group fairness [1] along with several recent papers studying individual fairness [2,3], some also utilizing Wasserstein distance [4].\\nAuthors also provided incorrect definition of disparate impact. Equation in the bottom of page 2 corresponds to statistical parity difference, while disparate impact is the ratio.\\n\\\"Equality of opportunity\\\" on the top of page 3 seems to be a typo\\n\\\"the mathematical properties of the disparate impact measure are not favorable, in particular it lacks robustness and smoothness features which would be necessary to blend algorithmic practice and mathematical theory\\\" - I don't think this claim makes sense. There are many prior works studying disparate impact and proposing algorithms to achieve it, e.g. the cited work of Feldman et al. Authors should be more specific regarding what mathematical properties they consider not favorable.\\n\\nThere are a lot of typos and grammatical mistakes, e.g.\\nin the 1st paragraph of section 2.2, the sentence \\\"Hence the aim in this case is to\\u201d is unfinished.\\nin the 1st paragraph of section 3, the first sentence seems to be unfinished.\\n\\n[1] Kleinberg, J., Mullainathan, S., & Raghavan, M. (2016). Inherent trade-offs in the fair determination of risk scores.\\n[2] Kearns, M., Roth, A., & Sharifi-Malvajerdi, S. (2019). Average Individual Fairness: Algorithms, Generalization and Experiments.\\n[3] Jung, C., Kearns, M., Neel, S., Roth, A., Stapleton, L., & Wu, Z. S. (2019). Eliciting and Enforcing Subjective Individual Fairness.\\n[4] Yurochkin, M., Bower, A., & Sun, Y. (2019). Learning fair predictors with Sensitive Subspace Robustness.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors propose a method for adding an approximate disparate impact loss to a classification objective, and show that optimizing a classifier for this loss leads to \\\"fairer\\\" predictions with little or no accuracy loss.\\n\\nThe authors first formulate two notions of fairness in terms of earth mover distance between the distribution of scores conditioned on either value of a protected variable. They then show that the dual formulation of the earth mover distance can be approximated (specifically, lower-bounded) by optimizing the parameters of a neural network under spectral norm constraints. This leads to a min-max global optimization scheme to learn a fair classifier.\", \"strengths\": \"The proposed method does better on the considered fairness metric than a GAN model with similar accuracy.\", \"weaknesses\": \"The paper is difficult to read, glosses over some important details, and contains some inaccuracies.\\n\\n-- Clarity: \\n--- The authors need to better describe the assumptions (or lack thereof) made on the joint distribution of X, S, and Y. \\n--- Measures such as the quantiles or probability laws need to be formally defined before they are used in definitions. \\n--- The \\\\mathcal{L} notation is overloaded (it is used for probability laws, conditional and unconditional, as well as marginals, with \\\\mathcal{L}_1 referring to both!), leading to potential confusion.\\n--- The domain of X and Y in equation (2) is not defined anywhere, neither is the distance.\\n--- Similar lack of consistency with the use of F / \\\\mathcal{F} / \\\\hat{f}, without any explicit parameterization\\n--- Figure 2 needs to be in Section 4, and Table 1 needs to be trimmed to size\\n\\n-- Overlooked problems:\\n--- In the dual formulation, the optimization is done over a sub-set of Lipschitz function, hence approximation of \\\\mathcal{W} is a lower bound at every step. Minimizing a lower bound on a loss can be justified, but requires more discussion\\n--- The trade-off inherent in the choice of n_w in algorithm 1 needs to be further discussed, especially in the case of large datasets where a full epoch of SGD in the inner loop of the optimization process is impractical\\n\\n-- Inaccuracies:\", \"the_graphical_models_in_figures_1_and_2_and_conditional_independences_written_in_the_text_are_not_consistent\": \"--- In Figure 1, X is NOT independent of S given Y (neither is Y*) (see: V structures in a directed graphical model)\\n--- In Figure 2, X* is independent of S regardless of conditioning on Y\\n\\nConsidering all of the above issues, the paper is not currently ready for publication\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"In this paper, the authors proposed a fairness-aware learning method.\\nIn particular, the authors considered two kinds of fairness problem and designed two regularizers accordingly.\\nEssentially, both of these two strategies learn classifiers and calibrate the distributions conditioned on protected variables jointly. \\nThe calibration of the distribution is achieved in the framework of optimal transport.\\n\\nThis work is a natural extension of the optimal transport-based method shown in (Barrio et al, 2019a,b). The main differences include 1) instead of calibrating distributions after learning classifiers, the proposed method achieves calibration and learning jointly, replacing the primal Wasserstein barycenter problem with the dual form of Wasserstein distance (Arjovsky et al. 2017); 2) the proposed method considers two types of fairness problem. \\n\\nCompared with vanilla GAN, the potential advantage of WGAN on distribution matching is well-known. It seems unfair that the authors compared the vanilla GAN-based regularizer with the proposed WGAN-based regularizer just on EMD because EMD corresponds to the proposed regularizer directly. In Table 1, although the DI of vanilla GAN is higher than that of WGAN, its ACC is also higher than that of WGAN as well. In Figure 5 (a, b), if we set lambda=0.6 for WGAN and lambda=1 for vanilla GAN, both of them can achieve ~0.838 ACC and ~0.100 DI. In Figure 5(c), what do the points represent? Why not use DI as the x-axis? Because of the issues in experiments, it is hard to evaluate the improvements of the proposed method.\\n\\nAdditionally, the proposed method always causes the degradation of ACC when improving DI. However, the method in (Barrio et al, 2019a) just applies a Wasserstein barycenter-based post-processing but can suppress the degradation on ACC greatly. Could the authors discuss the differences and the advantages of the proposed method in detail? Could the authors consider more recent work as their baselines?\\n\\nIn summary, the method makes sense, but its novelty is limited and the improvements are incremental.\", \"minors\": \"Page 6, Line 3: Figure 3 \\u2014> Figure 2.\\nI suggest swapping Figure 2 and Figure 3.\"}" ] }

[ "Jie Zhang", "Yuxiao Dong", "Jie Tang" ]

[ "data mining", "graph convolutional networks" ]

{ "note_id": [ "EGb5NlF92", "rkxOTkZg5S", "H1lFVxr0Fr", "S1gSMbtpuH" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731729, 1571979200173, 1571864624518, 1570767116757 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1757/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1757/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1757/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"All three reviewers are consistently negative on this paper. Thus a reject is recommended.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This work proposes node adaptive rescaling and node adaptive encoder to enhance GCN. The proposed method is well explained and simple to implement. The experimental results indicate improvement and use t-test to show it's statistically significant.\", \"cons\": \"1. I am not convinced that GCN does not consider node importance. When aggregating information, the degree information is used. In the example of opinion leaders, the nodes corresponding to opinion leaders usually have large degree.\\n2. The proposed method is very similar to the attention method in \\\"Hierarchical Attention Networks for Document Classification\\\", where the query vector is a learnable vector. The authors should at least cite it and make appropriate discussion.\\n3. The node adaptive encoder is more suitable to be called a new activation function. Then ablation study on different activation functions should be made.\\n4. The experiments should cover comparison to more recent works, as well as graph classification tasks. For example, comparison to GIN or extending the proposed method to GIN would be interesting.\\n5. Examples of how the proposed method learn the node importance correctly should be provided.\", \"minor_cons\": \"1. Figure 1 has a part of low resolution.\\n2. The updated node importance in eqn.(6) should use a modified notation to differ from eqn.(3).\\n3. The position of Table 2&3 is strange.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper is well-written. The authors propose a new graph convolution neural network framework (DiagGCN) with an adaptive neighborhood aggregation step to adaptively scale the output signal for each node. In addition, they also propose an adaptive nonlinear encoder for the node's signal. Interestingly, sophisticated propagation models like graph networks\\n with node/edge/path attention can also be formulated as the permutations of the adjacency matrix and diagonal matrices in the DiagGCN.\\n\\nGenerally, from my perspective, the idea in the paper is to reformulate the attention mechanism used in the GCN and this new diagonal mechanism is simpler and more general. One of my concerns is the \\\"node adaptive encoder\\\". Based on the experimental results, I did not see an obvious improvement between DiagGCN and DiagGCN with rescaling alone in Table 2 and Table 3. The classification improvement is 0, 0.1, 0.1, 0.3 with the node adaptive encoder. In addition, when comparing DiagGCN with baseline GAT, which is proposed in 2018, the performance approvement is also limited for the transductive learning tasks. \\n\\nIn conclusion, the idea proposed in the paper is interesting to me but the effectiveness of each module is not very well-supported by the experiments.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposed a so-called diagonal GCN method with adaptive neighborhood aggregation rules., namely, each node will be associated with an individual importance factor to re-scale the output signal of the graph convolutional operator, and also an adaptive encoder function is adopted. This is achieved by adding a left diagonal matrix P and a right diagonal matrix Q to the standard message passing of GCN H\\u2019 = sigma(AHW), as H\\u2019(PAQHW); P and Q are parameterized based on the HW part (node features). Multihead attention and multi-hop versions are also derived. Connection of this formulation to the existing GAT method is also discussed. Empirical results on transductive node classification and inductive node classification are reported against state-of-the-arts.\\n\\nOverall I feel that the novelty of the proposed method is quite limited. Adding a rescaling on each node seems a minor structural change based on existing GCN framework. Although the authors claim that the GAT method can be deemed as a special case of the proposed methods, the main contribution of inductive attention in GAT should not become a credit of the proposed method. Actually the empirical performance of the proposed method has only a small margin compared with the GAT, on the three node classification benchmark datasets, where the accuracies are around 80%, the improvement of the best version of the proposed methods is only around 0.3%, 0.7%, and 0.1%, respectively. This is somehow predictable since the proposed method does not exploit any extra useful information (such as graph structure) but simply base the new parameters on the existing components of the GCN message passing procedure. \\n\\nThe motivation of the paper is that the diffusion process on the graph should depend on the nodes\\u2019 status and influence (for example, some opinion leader nodes should have an amplified spread than other nodes), while existing GCN dies bit examine such ``difference\\u2019\\u2019 between the nodes. However, the roles or status of the nodes in a network are usually unavailable in practice; and if they can be ever be inferred, one has to use structural clues, Unfortunately, the extra parameters added here, namely the matrix P, Q, are dependent only on the features of the nodes (namely HW part), but not the graph structures, therefore I do not think such parametrization would lead to significant benefit in improving the diffusion process. Indeed, the original Gat method (graph attention network) has already achieved adaptive aggregation of the neighborhood information, by learning an inductive attention function based on the node features; therefore the authors claim that existing GCN \\u201cignores the difference between nodes\\u201d sounds very vague, since their new parameters are not capable enough to reflect \\u201cnode differences\\u201d either (as a function of the node features).\\n\\nSome minor comments \\n(1)\\tFigure~1 is not easy to read while it is supposed to visualize the main idea clearly in one picture\\n(2)\\tIs there proof that if multi-hop version is used, the results can be improved?\\n(3)\\tWhat is the extra number of parameters compared with the standard GCN message passing\\uff1f\"}" ] }

[ "Igor Lovchinsky", "Alon Daks", "Israel Malkin", "Pouya Samangouei", "Ardavan Saeedi", "Yang Liu", "Swami Sankaranarayanan", "Tomer Gafner", "Ben Sternlieb", "Patrick Maher", "Nathan Silberman" ]

[ "Evaluation Metrics", "Medical Imaging" ]

{ "note_id": [ "OOXxnJSnER", "r1eMkC3mnS", "rylN2D7DoB", "SygxEwQvsH", "r1xu9IQwsH", "Syx3rIXDor", "HkexErQvjH", "rJxRqQToFr", "SJxfhzwjtS", "BJg_rIUitr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731697, 1574321626499, 1573496748399, 1573496615694, 1573496463992, 1573496388246, 1573496104328, 1571701653990, 1571676841689, 1571673664214 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1756/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1756/Authors" ], [ "ICLR.cc/2020/Conference/Paper1756/Authors" ], [ "ICLR.cc/2020/Conference/Paper1756/Authors" ], [ "ICLR.cc/2020/Conference/Paper1756/Authors" ], [ "ICLR.cc/2020/Conference/Paper1756/Authors" ], [ "ICLR.cc/2020/Conference/Paper1756/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1756/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1756/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper tackles an interesting problem: \\\"How should we evaluate models when the test data contains noisy labels?\\\". This is a particularly relevant question in the medical imaging domain where expert annotators often disagree with each other. The paper proposes a new metric \\\"discrepancy ratio\\\" which computes the ratio how often the model disagrees with humans to how often humans disagree with each other. The paper shows that under certain noise models for the human annotations the discrepancy ratio can exactly determine when a model is more accurate than humans, whereas commonly used baselines such as comparing with the majority vote do not have this property. Reviewers were satisfied with the author rebuttal, particularly with the clarification that the goal of the metric is to accurately determine when model performance exceeds that of human annotators, and not to better rank models. The metric should be quite useful, assuming users are cautious of the limitations described by the authors.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Clarification of my comment\", \"comment\": \"The authors asked me that:\\n\\\"We would like to ask for clarification from the reviewer on this comment. Is the reviewer asking for an additional panel to demonstrate how the Relative F1 score scales with the label swap probability?\\\"\\n\\nThe short answer is yes. If I understand correctly, there is no F1 score scales with the label swap probability. So it may be nice to add it for completeness if possible. Nevertheless, if the authors believe that this is not important because we do not use this information to interpret the result or highlighting the importance of the relative F1 or the discrepancy ratio. Then I believe it is acceptable to not include it.\\n\\nI read through the updated paper and found that the clarity is improved. I increased my score to accept.\"}", "{\"title\": \"Response to Official Blind Review #3 (2/2)\", \"comment\": \"\\u201c10. I think reporting the performance of the mean-per-class discrepancy ratio without defining it in the main text (it is defined in a natural language (English) in Appendix E) makes the main body of the paper not self-contained.\\u201d\\n\\nWe have removed the reporting of the mean-per-class discrepancy ratio from the main text of the paper. We have retained it in Appendix E since it may still be a useful extension of the discrepancy ratio in cases where the dataset is highly imbalanced.\\n\\n\\u201c11. Can we use F1 to evaluate whether a model is better than the average human performance?\\u201d\\n\\nYes, as we state in the second paragraph of Section 4.1.4 that \\u201cRelative F1 does allow for evaluating the model by comparing it to annotator performance but the discrepancy ratio has several advantages\\u201d. The advantages are subsequently summarized.\\n\\n\\u201c1. The authors may consider using \\\\begin{definition} when defining three key values: the mean pairwise deviation (MPD), annotator discrepancy, and model discrepancy. So the readers can easily look up when getting lost in definitions. And it may help to make a definition more precise (e.g., given label Y, data points, MPD is defined as follows).\\u201d\\n\\nWe thank the reviewer for this suggestion. This will be changed in the final version of the paper.\\n\\n\\u201c2. Eqs (1), (2), (3) may be easier to understand if we put a denominator out of the summation as much as we can. For example, for Eq. 2, N can be outside of the two sums, A^i(A^i-1) can be between the two sums.\\u201c\\n\\nThis will be changed in the final version of the paper.\\n\\n\\u201cOn the experiments: \\n1. Why we use model 1.0, 0.95, 0.9 for Figures 1c, 1d, 2c, 2d but 1.0, 0.9, 0.8 for Figures 1a, 1b, 2a, 2b? Perhaps full results with many models should also be reported for completeness.\\u201d\", \"the_reason_for_this_was_purely_aesthetic\": \"in Figures 1c and 1d the curves for models 1.0, 0.95 and 0.9 are closer together, thus making it easier to see the dependence on the number of annotators and the single annotator strictness.\\n\\n\\u201c2. Isn't it more natural to show a graph that has x-axis as label swap probability in the same figure for comparison? For example? For example, Figure 1 only report performance with respect to each measure (Naive Evaluation, Majority Vote, Relative F1, Discrepancy ratio) with respect to label swap probability. We can fix the number of annotators to 9 for relative F1 and discrepancy ratio, to validate whether it also becomes increasingly difficult to differentiate model improvements or not for them. Then after suggesting that Naive evaluation is not good and other metrics look fine. Majority vote cannot evaluate that the model is better than human average performance, which we want to know.\\u201d\\n\\nWe would like to ask for clarification from the reviewer on this comment. Is the reviewer asking for an additional panel to demonstrate how the Relative F1 score scales with the label swap probability?\\n\\n\\u201c3. For the MNIST benchmark dataset, it is possible to show the performance on the test ground truth too. And I think it is interesting to compare different models with the same label noise rate and confirm that discrepancy ratio is successful to evaluate the best model with respect to the ground truth accuracy. It would be very impressive to see if the best model with respect to the discrepancy ratio can successfully be the best also on the ground truth test set.\\u201d\\n\\nFor the MNIST dataset, the performance on the test ground truth data corresponds to a label swap probability of 0. Thus this information is already included in Figures 1 and 2. The three models presented in each panel on these plots are also all evaluated with respect to the same label noise. Specifically, each point on the x-axis corresponds to a specific level of label noise. The discrepancy ratio is indeed successful at finding the best model - notice that in Figure 2b, the best model (Model 1.0) has the lower discrepancy ratio for all levels of label noise, indicating that it performs best as compared to annotators.\\n\\n\\u201cOn space usage:\\nI feel there are many important things to be included in the main body but it is instead in the appendix. On the other hand, I think there are several parts that can be shortened in the main body. And it seems that the binary classification example (3.3) also consumed a lot of space. The derivation may be omitted and make it a Theorem and then let's suggest the readers to read the appendix for the proof if they are interested. Then we can have more space to add important components in the main body of the paper.\\u201d\\n\\nAs requested by the reviewer, some of the details of the derivation of the example in Section 3.3 has been moved to Appendix B.\", \"in_the_final_version_of_the_paper_we_will_move_the_following_information_from_the_appendices_to_the_main_text\": \"1). Some details on the model training and architecture presented in Appendix D.\\n2). Tables S1 and S2 detailing the dataset sizes for the PLAX and LVEF tasks.\"}", "{\"title\": \"Response to Official Blind Review #3 (1/2)\", \"comment\": \"We thank the reviewer for their positive review of our work. We specifically appreciate their highlighting the fact that \\u201cThe problem that this paper addresses is highly important\\u201d, that our proposed solution is \\u201cintuitive and is easy-to-use\\u201d and that the reviewer is \\u201cconvinced that discrepancy ratio is potentially useful for the community.\\u201d We have addressed in detail all comments and suggestions made by the referee and feel that this has substantially strengthened our paper. We hope that the resulting paper (uploaded) now merits clear acceptance. We strongly feel that our work is of critical importance in allowing the community to evaluate models in real-world settings where we have no access to unambiguous ground truth labels. To the best of our knowledge, our work represents the only approach in the literature that provides a robust, task-independent, interpretable metric to answer the crucial question of how a machine learning model performs at a task relative to humans.\\n\\n\\u201c1. While MPD is defined with an argument (Y, Y'), the definitions annotator discrepancy, model discrepancy and discrepancy ratio ignore the inputs of the function. It might be better to keep the arguments of the function for clarity. When people want to adopt the discrepancy ratio in their work and there are many ratios to discuss, then the proposed definition can be used conveniently.\\u201c\\n\\nThis has been added to the revised paper. \\n\\n\\u201c2. In Eq. (5) \\\\sigma seems to be undefined. I believe it is a variance but it is kinder to readers to explicitly state it (I found the authors explicitly stated it in Appendix B).\\u201d\\n\\nGood catch. We explicitly note that \\\\sigma^2 is variance in the revised version of the paper.\\n\\n\\u201c3. In Eq. (8), is it intentional to have exists over m and y as written in the paper?\\u201d\\n\\nYes, the subscript i refers to the sample index and the covariance is taken over this index. In order to avoid confusion this has been clarified in the revised version.\\n\\n\\u201c4. Does the discussion of the simplicity of the discrepancy ratio hold for other \\\\delta, i.e., not the squared loss?\\u201d\\n\\nYes, the interpretation of the discrepancy ratio is exactly the same for any measure of agreement (delta), as described in the last paragraph of Section 3.1. To summarize, the discrepancy ratio can be thought of as a \\u201cmeta-metric\\u201d that measures how the model performs relative to humans, as compared to how they perform relative to each other for any measure of agreement. We elaborate further on this point in Section. 4.2.2.\\n\\n\\u201c5. In Section 4.1 (experiment with MNIST), did the authors binarize the data or simply discuss the multi-class problem? Because this follows section 3.3 which dedicates the whole page to discuss binary classification. It may be a bit sudden to move to multi-class classification and an additional sentence to clarify this might help.\\u201d\\n\\nWe completely agree with the reviewer that this point was not clear in our initial submission. To clarify, in Section 4.1 we simply calculate the accuracy using the MNIST labels without binarizing. This has been clarified in the revised version.\\n\\n\\u201c6. I could not understand what is \\\"single annotator strictness in Figures 1d and 2d. It is not explained if I did not miss it.\\u201d\\n\\nIn Section 4.1.3, we state that \\u201cWe find that as we increase the strictness of a single annotator (the fraction of the dataset they skip), the relative F1 score is correspondingly reduced.\\u201d The reviewer is absolutely correct that the exact phrase \\u201csingle annotator strictness\\u201d, which refers to the statement above was never defined. This has been fixed in the revised paper.\\n\\n\\u201c7. I think we should not resize the figure/table caption. The text size of captions are much smaller than the main text.\\u201d\\n\\nThis has been fixed.\\n\\n\\u201c8. Is PLAX binary classification? (I understood as yes from Figure (3a) but I think it's also nice to clarify it to avoid confusion (also for MNIST)).\\u201d\\n\\nYes, the PLAX task is binary classification. This has been clarified in the revised paper.\\n\\n\\u201c9. What is the y-axis for figure 3a?\\u201d\\n\\nFigure 3a is a normalized histogram of the average labels on the PLAX classification dataset. This has been clarified in the caption.\"}", "{\"title\": \"Response to Official Blind Review #1 (2/2)\", \"comment\": \"\\u201cIn section 4.1.4, the claim that \\\"It is in this regime that the model performance can be said to be better than the average human performance\\\" seems questionable. When the ratio is less than 1, it simply means the annotator-model agreement is larger than the annotator-annotator agreement. What's the definition of the \\\"average human performance\\\" in this context?\\u201d\\n\\nWithout access to unambiguous ground-truth labels, a natural way to define human performance is by comparing an annotator\\u2019s labels to those provided by other humans on the same samples. Consequently, we use the annotator discrepancy (Eq. 2) as a measure of average human performance. We have stated this explicitly in the revised paper immediately following Eq. 2 so that it is absolutely clear what we mean by this term.\\n\\nDespite not having access to unambiguous ground-truth in real-world problems, we have provided empirical support for the aforementioned choice using MNIST. The MNIST experiments demonstrate that the point at which the discrepancy ratio falls below 1 corresponds precisely to the point at which the model performance exceeds that of the (simulated) humans. The red and blue dots in Figure 2b occur at label swap probabilities of 0.1 * 10/9 and 0.2 * 10/9, respectively, which correspond precisely to when the accuracies of the simulated human annotators equal that of the models (0.9 and 0.8, respectively).\\n\\n\\u201cIn the work, multiple assumptions were made regarding different annotators. First, multiple annotators are assumed to be conditionally independent given the ground-truth label. Second, each annotator should provide high-quality labels. Third, different experts should have similar expertise. However, in medical applications, any of these assumptions could be violated. For example,1) multiple experts could provide similar labels given their similar knowledge background. It's more likely to observe multiple groups of experts, where experts in the same group are similar. 2), experts could provide random labels when they are extremely uncertain 3) different experts could have varying levels of expertise. Could the author provide some ideas on how to relax some of these assumptions?\\u201d\\n\\nWe completely agree with the reviewer regarding all three of these points! Thanks for helping us clarify the paper on these issues:\\n\\n1). The reviewer is absolutely correct that there can be situations where annotators adhere to several distinct \\u201cschools of thought\\u201d and their annotations could be similar within each group but different between groups. In these cases it may make more sense to evaluate the discrepancy ratio independently with each group rather than lumping everyone together. We have added this point to the Limitations section of our paper (Section 5).\\n\\n2). As we mentioned in Section 4.1.3, in many tasks annotators must be allowed to voluntarily skip examples they feel uncertain about precisely so that they don\\u2019t provide random labels. It is this skipping that can lead to the strict annotator bias that is resolved through the use of the discrepancy ratio (see Figure 2d). We have added a remark to the Limitations section that emphasizes that in difficult tasks annotators must be allowed to skip examples where they are unsure about the label.\\n\\n3). We explicitly state in Limitations section that our approach assumes that the experts have been vetted as capable for this task. We have provided details of our vetting procedure in Appendix A and have added a reference to a comprehensive review paper on gathering reliable labels (Hovy and Lavid, Int. Journal of Translation (2010)) to the Limitations section.\"}", "{\"title\": \"Response to Official Blind Review #1 (1/2)\", \"comment\": \"We thank the reviewer for their helpful comments and suggestions and for bringing up several important issues. We have addressed in detail all of the reviewer\\u2019s comments below.\\n\\n\\u201cIn the experimental results on MNIST classification, for Naive Evaluation and Majority Vote, as the author pointed out, the gap between three models (model 1.0, model 0.9, model 0.8) becomes narrower when the label swap probability increases, indicating it's becoming more difficult to differentiate model improvements. However, this is also true for the proposed discrepancy ratio, as is shown in Figure 2(b). Therefore, the discrepancy ratio didn't overcome those \\\"drawbacks\\\" associated with these two baselines, correct? In other words, in the high-noise regime, it would be natural to observe narrower gaps between those models.\\u201c\\n\\nWe would like to clarify an important point that may have been unclear in our initial submission.The reviewer is referring to a comment made in Section 4.1.1 that \\u201c...as the label noise increases, the observed accuracy converges and it becomes increasingly difficult to differentiate model improvements\\u201d. This remark was meant to support the use of the Majority Vote method over the Naive Evaluation method, since majority vote can to some extent improve the ability to distinguish models by averaging away the label noise. We have clarified this point explicitly in the revised version of the paper.\\n\\nTo be sure, we never claimed that the discrepancy ratio itself removes the narrowing gap between models as the label noise increases, and this was not our goal. By definition the performance gap between models has to narrow and converge to zero for *any* metric since when the swap probability is 1 the labels are completely random and hence all models are equivalent. If one simply wants to rank models (determine which one is better) than *any* of the baseline methods are perfectly suitable. On the other hand, as discussed in Section 2, in many real-world applications one instead needs to know how a model performs at a task relative to how humans perform. To answer this question, the Naive Evaluation and Majority Vote methods are no longer useful, whereas the discrepancy ratio is quite insightful. \\n\\n\\u201cIn MNIST classification, the performances gap of those three models evaluated by Naive Evaluation and Majority Vote didn't decrease rapidly as the noisy level increases. This is consistent with the constant gap between their ground truth accuracies (gap=10% between 100%, 90%, 80%). However, for the proposed discrepancy ratio, the gap between these three models decreases rapidly when the noise level increases from 0 to 0.4. This is an undesirable behavior since it over-estimate the performance gap in the low-noise regime. Since most real-world datasets could belong to this low-noise regime, this rapid change of performance gaps seems problematic.\\u201d\\n\\nThe reviewer is quite correct that the performance gaps between the majority vote and discrepancy ratio metrics are different. Ultimately, our goal is a metric that captures how the model performs relative to annotator disagreement. To this end, any gap in quantitative performance between two models should capture the degree to which the model predictions are closer or further to annotators than annotators are to one another. To clarify this, Figures 1b and 2b demonstrate a subtle but crucial feature of metrics sensitive to annotator disagreement: models exhibiting similar accuracies (even within 10%) can actually perform *very* differently when compared to human variability. In this sense, the discrepancy ratio is *correctly* picking up on real performance differences that the baseline metrics are simply not sensitive to. As a case-in-point, consider the example of Models 1.0 and 0.9 when the label swap probability is equal to 0.05. When looking at the baseline metrics (Native Evaluation and Majority Vote) the models only appear to be different by 10% (Figures 1a and 1b). However, when taking into account the annotator disagreement (Figure 2b) a completely different story emerges and one immediately sees that the models are in fact dramatically different: despite only a slight difference in accuracy, model 0.9 is twice as far from annotators as they are from each other, while model 1.0 is twice as close, thus resulting in a performance gap of 400%. We have clarified this very important point in Section 4.1.4.\"}", "{\"title\": \"Response to Official Blind Review #2\", \"comment\": \"We thank the reviewer for their helpful comments and suggestions. We have addressed in detail all of the reviewer\\u2019s concerns and hope the resulting paper (uploaded) merits unambiguous acceptance. Our point-by-point response is detailed below.\\n\\n\\u201cNovelty: The mathematical formulation of the metric consists of simple sums and averages, which in itself is not novel. However, the authors' choice of using such formulations to assess model performance is novel.\\u201d\\n\\nThe derivation of the model and annotator discrepancies in Section 3.3 was simply meant to serve as a concrete example of how these metrics are calculated in a particularly simple case. We hope that this example clarifies to the reader the more general formalism presented in Section 3.1 and helps to build intuition for the properties of the discrepancy ratio (see for example the last paragraph of Section 3.3). We would also like to point out that the novelty of our paper is not in this simple example but in providing a robust, task-independent and highly-interpretable metric that allows us to quantify model performance in the context of how humans perform relative to each other. \\n\\n\\n\\u201cClarity: The authors could simplify their notations, and could periodically remind the readers about the notations. For example, I was stuck by the sigma^2 notation in (6) (undefined) and the m notation in (7) defined in the past page in small texts. It would help me a lot if the authors reminded me the definition of those notations when they appeared.\\u201d\\n\\nThanks for pointing this out. We have added to the revised manuscript the definition of the variance to Eq. 5 and included several additional reminders of the meaning of symbols when they occur far from their definitions.\\n\\n\\n\\u201cNumerical Experiments: This is my biggest complain. I wasn't able to conclude that the authors proposal is better than majority voting based on Figures 1b and 2b. To me they look qualitatively the same. Is there any reason that the discrepancy ratio is superior to the majority vote?\\n\\nThis is a subtle point that could have been made more clear in our initial submission. The discrepancy ratio is indeed superior to the majority vote method, as discussed in detail in Section 2, Section 4.1.2 and Section 4.2.1, though this may not be entirely clear from simply looking at Figures 1b and 2b. The curves in these figures look qualitatively similar only in the sense that both allow us to decide which model is better. Indeed, if the goal is simply to rank models, then *any* of the baseline metrics are perfectly suitable. If, however, one needs to know how a particular model performs *relative to humans*, the majority vote method simply does not answer that question. This is the main motivation for using the discrepancy ratio and has been clarified in Section 4.1.2.\\n\\nIn other words, evaluating against majority vote does not give a clear indication of whether a machine learning model is reasonably substitutable for humans who are currently performing a given task in the real world. This is especially important in medical applications since decisions in the field are usually made by a single individual, not majority consensus. The discrepancy ratio, on the other hand, unambiguously answers the question of how a model performs against the humans by considering human-to-human disagreement. An additional property that makes the discrepancy ratio superior to the majority vote method is that the majority vote method gives a metric that is dependent on the number of annotators that annotated the dataset (as can be inferred from Figures 1a and 1b), while the discrepancy ratio is independent of the number of annotators (Figure 2c).\\n\\n\\n\\u201cIn addition, I didn't see whether the curves in Figures 1 and 2 are an average of numerous simulated samples. if not, the authors should use the average to mitigate randomness; and if the curves are averages, then the distribution of the metrics should be described, because I think ultimately, we don't care about whether the average black curve is above the average red curve, but the chance of black curve being above the red curve.\\u201c\\n\\nThe curves in Figures 1 and 2 are generated by taking the MNIST test set and randomly swapping the labels with probability p, as described in Section 4.1. Each point on these plots is generated with an independent swapping of the labels resulting in the small fluctuations that are evident in these curves. As requested by the reviewer, in the revised version of the paper we have replaced the curves with their averages as well as overlaid 95% quantiles to indicate the distributions of all the metrics.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors presented an interesting paper which tries to solve a practically important question. Biomedical and tech industries usually hire human labelers for machine learning tasks, whose labels are usually noisy. Therefore, it is important to have a metric that can distinguish the performance of models with noisy labels.\\n\\nIn this paper, the authors proposed to measure the model performance based on the ratio of the discrepancy between the model prediction and the labeler, with the discrepancy between the labelers. The authors showed that in binary classification settings, their proposed metric can be reduced to simple to analyze quantities. The authors demonstrated the performance of their proposal in synthetic data as well in two real-world medical image datasets.\\n\\nOverall, the paper is well motivated with a reasonable amount of novelty. The numerical experiments are well conducted, but I am not totally convinced by their results.\", \"detailed_comments\": \"\", \"significance\": \"The paper is trying to address a practically important issue in machine learning.\", \"novelty\": \"The mathematical formulation of the metric consists of simple sums and averages, which in itself is not novel. However, the authors' choice of using such formulations to assess model performance is novel.\", \"clarity\": \"The authors could simplify their notations, and could periodically remind the readers about the notations. For example, I was stuck by the sigma^2 notation in (6) (undefined) and the m notation in (7) defined in the past page in small texts. It would help me a lot if the authors reminded me the definition of those notations when they appeared.\", \"numerical_experiments\": \"This is my biggest complain. I wasn't able to conclude that the authors proposal is better than majority voting based on Figures 1b and 2b. To me they look qualitatively the same. Is there any reason that the discrepancy ratio is superior to the majority vote? In addition, I didn't see whether the curves in Figures 1 and 2 are an average of numerous simulated samples. if not, the authors should use the average to mitigate randomness; and if the curves are averages, then the distribution of the metrics should be described, because I think ultimately, we don't care about whether the average black curve is above the average red curve, but the chance of black curve being above the red curve.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposed to evaluate model performance when the ground truth labels were not available and noisy labels provided by multiple uncertain experts were provided instead. The proposed evaluation metric, called discrepancy ratio, is defined as the ratio between the average model-annotator discrepancy and the average annotator-annotator discrepancy. It can be applied to compare 1) the relative performance of different models; and 2) the relative performance of average annotators and the model. Experimental results on the MNIST classification task showed the proposed metric can compare model performance under different noise levels and is robust w.r.t. the number of annotators and a single annotator's strictness. Experimental results on a real-world medical image classification was also presented.\\n\\nThis work is well motivated. In real-world applications, such as medical image classifications, obtaining ground-truth labels is difficult or even infeasible, only multiple noisy labels are available to evaluate the model performance. Therefore, the setup of this problem seems novel (besides Rajpurkar et al. 2018).\\n\\nIn the experimental results on MNIST classification, for Naive Evaluation and Majority Vote, as the author pointed out, the gap between three models (model 1.0, model 0.9, model 0.8) becomes narrower when the label swap probability increases, indicating it's becoming more difficult to differentiate model improvements. However, this is also true for the proposed discrepancy ratio, as is shown in Figure 2(b). Therefore, the discrepancy ratio didn't overcome those \\\"drawbacks\\\" associated with these two baselines, correct? In other words, in the high-noise regime, it would be natural to observe narrower gaps between those models. \\n\\nIn MNIST classification, the performances gap of those three models evaluated by Naive Evaluation and Majority Vote didn't decrease rapidly as the noisy level increases. This is consistent with the constant gap between their ground truth accuracies (gap=10% between 100%, 90%, 80%). However, for the proposed discrepancy ratio, the gap between these three models decreases rapidly when the noise level increases from 0 to 0.4. This is an undesirable behavior since it over-estimate the performance gap in the low-noise regime. Since most real-world datasets could belong to this low-noise regime, this rapid change of performance gaps seems problematic.\\n\\nIn section 4.1.4, the claim that \\\"It is in this regime that the model performance can be said to be better than the average human performance\\\" seems questionable. When the ratio is less than 1, it simply means the annotator-model agreement is larger than the annotator-annotator agreement. What's the definition of the \\\"average human performance\\\" in this context?\\n\\nIn the work, multiple assumptions were made regarding different annotators. First, multiple annotators are assumed to be conditionally independent given the ground-truth label. Second, each annotator should provide high-quality labels. Third, different experts should have similar expertise. However, in medical applications, any of these assumptions could be violated. For example,1) multiple experts could provide similar labels given their similar knowledge background. It's more likely to observe multiple groups of experts, where experts in the same group are similar. 2), experts could provide random labels when they are extremely uncertain 3) different experts could have varying levels of expertise. Could the author provide some ideas on how to relax some of these assumptions?\\n\\nComments after reading authors' reply\\n-----------------------------------------------------\\nI would like to thank the authors for their detailed reply to my questions. The key novelty of this work, i.e., evaluating model performance (in terms of comparing different models and model vs average annotator performance) in the presence of noisy labels, should be appreciated. I'll raise my score to 6: Weak Accept. In the future work, it would be interesting to apply the proposed metric (or develop new metrics) to study whether any of the assumptions (conditional independence, reasonably high accuracy of each individual expert, different expert accuracies) can be relaxed.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary: This paper proposes a novel metric \\\"discrepancy ratio\\\" for evaluating the performance of a model where the ground truth for each data point comes from many expert-yet-imperfect annotators. The authors suggested that this problem has remained largely unexplored. The proposed metric is intuitive and is easy-to-use. The authors suggested that this metric can be used for many applications. For example, to evaluate if a model is better than the annotators on average, to evaluate the annotator, to compare between models.\\n\\n========================================================\", \"comments_on_clarity\": \"The writing is good overall and I am able to understand the contribution and the key idea of this paper.\", \"questions_and_comments_on_clarity\": \"1. While MPD is defined with an argument (Y, Y'), the definitions annotator discrepancy, model discrepancy and discrepancy ratio ignore the inputs of the function. It might be better to keep the arguments of the function for clarity. When people want to adopt the discrepancy ratio in their work and there are many ratios to discuss, then the proposed definition can be used conveniently. \\n\\n2. In Eq. (5) \\\\sigma seems to be undefined. I believe it is a variance but it is kinder to readers to explicitly state it (I found the authors explicitly stated it in Appendix B).\\n\\n3. In Eq. (8), is it intentional to have $i$ exists over m and y as written in the paper?\\n\\n4. Does the discussion of the simplicity of the discrepancy ratio hold for other \\\\delta, i.e., not the squared loss?\\n\\n5. In Section 4.1 (experiment with MNIST), did the authors binarize the data or simply discuss the multi-class problem? Because this follows section 3.3 which dedicates the whole page to discuss binary classification. It may be a bit sudden to move to multi-class classification and an additional sentence to clarify this might help.\\n\\n6. I could not understand what is \\\"single annotator strictness in Figures 1d and 2d. It is not explained if I did not miss it.\\n\\n7. I think we should not resize the figure/table caption. The text size of captions are much smaller than the main text.\\n\\n8. Is PLAX binary classification? (I understood as yes from Figure (3a) but I think it's also nice to clarify it to avoid confusion (also for MNIST)).\\n\\n9. What is the y-axis for figure 3a?\\n\\n10. I think reporting the performance of the mean-per-class discrepancy ratio without defining it in the main text (it is defined in a natural language (English) in Appendix E) makes the main body of the paper not self-contained. I strongly suggest to either (1) remove mean-per-class discrepancy ratio entirely from the main body or (2) define it mathematically in the main body. Also, I don't see reporting mean-per-class discrepancy ratio adds values to a paper for current experiments because the reported mean-per-class ratio is very similar to that of discrepancy ratio (Figures 3b (almost overlap) and Table 2 (same trend and the values are very close to the discrepancy ratio for all cases)).\\n\\n11. Can we use F1 to evaluate whether a model is better than the average human performance?\\n\\nMinor points on clarity, which may just be a personal preference of a reviewer:\\n\\n1. The authors may consider using \\\\begin{definition} when defining three key values: the mean pairwise deviation (MPD), annotator discrepancy, and model discrepancy. So the readers can easily look up when getting lost in definitions. And it may help to make a definition more precise (e.g., given label Y, data points, MPD is defined as follows).\\n\\n2. Eqs (1), (2), (3) may be easier to understand if we put a denominator out of the summation as much as we can. For example, for Eq. 2, N can be outside of the two sums, A^i(A^i-1) can be between the two sums. \\n\\n========================================================\", \"significance\": \"The problem that this paper addresses is highly important. I like the idea and I am convinced that discrepancy ratio is potentially useful for the community. And it also opens many possibilities to analyze the discrepancy ratio and to find potential applications for it. \\n\\n========================================================\", \"other_comments\": \"\", \"on_the_limitation\": \"I am satisfied that the authors clearly discuss the limitations of the proposed method and I agree with the discussion.\", \"on_the_experiments\": \"1. Why we use model 1.0, 0.95, 0.9 for Figures 1c, 1d, 2c, 2d but 1.0, 0.9, 0.8 for Figures 1a, 1b, 2a, 2b? Perhaps full results with many models should also be reported for completeness.\\n\\n2. Isn't it more natural to show a graph that has x-axis as label swap probability in the same figure for comparison? For example, Figure 1 only report performance with respect to each measure (Naive Evaluation, Majority Vote, Relative F1, Discrepancy ratio) with respect to label swap probability. We can fix the number of annotators to 9 for relative F1 and discrepancy ratio, to validate whether it also becomes increasingly difficult to differentiate model improvements or not for them. Then after suggesting that Naive evaluation is not good and other metrics look fine. Majority vote cannot evaluate that the model is better than human average performance, which we want to know. Then Figure 2 highlights the contribution of the discrepancy ratio and shows that it outperforms relative F1 in some perspectives.\\n\\n3. For the MNIST benchmark dataset, it is possible to show the performance on the test ground truth too. And I think it is interesting to compare different models with the same label noise rate and confirm that discrepancy ratio is successful to evaluate the best model with respect to the ground truth accuracy. It would be very impressive to see if the best model with respect to the discrepancy ratio can successfully be the best also on the ground truth test set.\", \"on_space_usage\": \"I feel there are many important things to be included in the main body but it is instead in the appendix. On the other hand, I think there are several parts that can be shortened in the main body. And it seems that the binary classification example (3.3) also consumed a lot of space. The derivation may be omitted and make it a Theorem and then let's suggest the readers to read the appendix for the proof if they are interested. Then we can have more space to add important components in the main body of the paper.\\n\\n========================================================\", \"decision\": \"Although there are issues on the clarity of the paper, I believe these issues are not too difficult to fix in the final version. The idea is interesting and potentially impactful. For these reasons, I vote a weak accept for this paper.\", \"update\": \"I would like to thank the authors for the clarification. I have read the rebuttal.\\nI read through the updated version of the paper. I believe that although there is still room for improvement, the contribution of this paper is sufficient as an important step towards solving this difficult yet highly relevant problem. This paper also stimulates several future directions. The authors clearly stated the limitations of the paper, which is highly useful for researchers to improve this work or find a better solution. The clarity of the paper is also improved. For these reasons, I increased my score from weak accept to accept.\"}" ] }

[ "Zoe Piran", "Naftali Tishby" ]

[ "optimal prediction learning", "exponential families", "critical points", "information theory" ]

{ "note_id": [ "wArIlIZqV", "ByeFBnz3sH", "S1lGehG3iS", "rJgaJbzhsH", "rkewlAW2jr", "HyxXpukrjH", "S1xNKglJcH", "Byl0gW8RFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731666, 1573821505466, 1573821418275, 1573818596554, 1573817839453, 1573349563446, 1571909755968, 1571868917667 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1755/Authors" ], [ "ICLR.cc/2020/Conference/Paper1755/Authors" ], [ "ICLR.cc/2020/Conference/Paper1755/Authors" ], [ "ICLR.cc/2020/Conference/Paper1755/Authors" ], [ "ICLR.cc/2020/Conference/Paper1755/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1755/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1755/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"Main content:\\n\\nBlind review #1 summarizes it well:\\n\\nThis paper introduces a variant of the Information Bottleneck (IB) framework, which consists in permuting the conditional probabilities of y given x and y given \\\\hat{x} in a Kullback-Liebler divergence involved in the IB optimization criterion.\\n\\nInterestingly, this change only results in changing an arithmetic mean into a geometric mean in the algorithmic resolution.\\n\\nGood properties of the exponential families (existence of non-trivial minimal sufficient statistics) are preserved, and an analysis of the new critical points/information plane induced is carried out.\\n\\n--\", \"discussion\": \"The reviews generally agree on the elegant mathematical result, but are critical of the fact that the paper lacks any empirical component whatsoever.\\n\\n--\", \"recommendation_and_justification\": \"The paper would be good for ICLR if it had any decent empirical component at all; it is a shame that none was presented as this does not seem very difficult.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"We thank the reviewer for the\\u00a0helpful comments.\\u00a0We were delighted to see that the reviewer\\u00a0agrees with\\u00a0our understanding of\\u00a0the contribution and innovations\\u00a0that the formalism\\u00a0suggest in comparison to the known $\\\\rm{IB}$.\", \"we_will_now_relate_to_specific_points_raised_be_the_reviewer\": [\"Motivation:\\u00a0 We refer the reviewer to the general official comment posted.\", \"Interpretation of the results: Following the reviewers comment we have added examples to previous application of the$\\\\rm{IB}$ (sec. 1.1 \\u201cThe Information Bottleneck method\\u201d) to tie the general framework into context and believe that now the dualIB framework can be tested and applied to on similar tasks.\", \"Figures explanation:\\u00a0 In addition we understood that the text was lacking an explanation of the experiments and the interpretations of the results, Thus we added an explanation regarding the results and settings for the problem presented in figure 1(a),2(a).\", \"We thank\\u00a0the reviewer for pointing out the typo at Equation 3 (fixed in the newly submitted version).\"]}", "{\"title\": \"Response to Review #2\", \"comment\": \"We wish to thank to reviewer for his response, while we stand behind it being a theoretical paper (a point we soon relate to in detail) we agree that the paper did not convey the motivation behind the new framework presented and with that the advantages it has. We have revised the paper accordingly as elaborated in the general official comment posted.\", \"we_will_now_relate_to_specific_points_raised_be_the_reviewer\": [\"Theory-based paper:\\u00a0\\u00a0We agree with the\\u00a0reviewer\\u00a0that this paper is theory-based as it aims to reveal a new framework, the\\u00a0Dual Information Bottleneck\\u00a0($\\\\rm{dualIB}$), and lay down the grounds for future applications of it. As such we find it necessary to analyse it\\u2019s analytical/theoretical properties and compare it to the original form, the\\u00a0Information Bottleneck ($\\\\rm{lIB}$).\", \"Having said that, we understand the need of empirical experiments to demonstrate the claims and have thus added a new example that is analysed in the text, see sec. 3 \\u201cThe Exponential Family dualIB\\u201d. This is still not an application of a\\u00a0concrete representation learning task yet it provides a visual evidence to the claimed behaviour.\", \"Motivation:\\u00a0 We refer the reviewer to the general official comment posted.\", \"Computational complexity:\\u00a0\\u00a0The computation complexity reduction emerges from the analytical result (Theorem 7) in which we show that the encoder-decoder are given by the sufficient statistics of the input distribution. Thus, the task is now to track these rather than the full distribution. This result is appealing to many problems which have a natural parametrization (i.e a physical problem defined by a given Hamiltonian thus a Gibbs distribution) but are intractable using the IB.\", \"Output prediction:\\u00a0We clarify that in this we refer to the maximization of the information on the\\u00a0newly added\\u00a0prediction variable, $\\\\hat{Y}$, as it rises naturally from the new Lagrangian. Thus the theory behind this lays in the formalism of the problem. in the revised version a visualization of the superiority is shown for the new setting and the phenomenon is clearly, see Figure 4(b).\", \"As suggested by the reviewer we further evaluate now the prediction ability over the examined cases by displaying the $\\\\mathcal{L}_1$-expected loss which makes it evident that the $\\\\rm{dualExpIB}$ outperforms the IB minimizing the loss for all values of the trade off parameter, see Figure 4(c).\", \"Data from exponential families or\\u00a0\\u201calmost\\u201d: \\u00a0We thank the\\u00a0reviewer\\u00a0for pointing out that the motivation for choosing\\u00a0the exponential families was unclear and made an\\u00a0attempt to\\u00a0clarify this in the text (see Sec. 3 \\u201cThe Exponential Family dualIB\\u201d ). From a statistical point of view a natural\\u00a0way of\\u00a0parametrizing data is\\u00a0obtained by a distribution from the\\u00a0exponential family as\\u00a0most of the commonly used distributions form an exponential family or subset of an exponential family.\", \"Moreover the exponential family form emerges when asking what is the\\u00a0maximum-entropy\\u00a0distribution consistent with given constraints on expected values.\", \"Within the scope of the paper we did not relate to the quality of the exponential family approximation of the data\\u2019s distribution.\", \"Application to deep learning:\\u00a0We thank the referee for pointing the lack of further discussion. We have clarified that in this we mean that current applications of the $\\\\rm{IB}$ in the DL regime can be done using the new $\\\\rm{dualIB}$ formalism, with the latter conserving natural parametrization and symmetries of the original problem.\", \"Deterministic IB: The Deterministic IB aims to better capture the notion of compression whereas our framework focuses on the parametrization of the data and prediction, thus they concern a different task.\", \"\\u201cpreserves the low dimensional sufficient statistics of the data\\\": \\u00a0As stated by the reviewer the $\\\\rm{dualExpIB}$ decoder (and encoder) are now of exponential form and given by sufficient statistics which are given by the one\\u2019s of the original distribution, $A_{r}(x)$, as given in equation 16. Thus it is given that the dimensionality is preserved, with $d$ such variables. The dependance on $\\\\hat{x}$ is obtained by considering the expectations, $\\\\sum_{x} A_{r}(x) p_{\\\\beta}(x\\\\mid\\\\hat{x})$, of the same functions, hence preserves the values mean.\", \"We realize that the $\\\\rm{dualIB}$ framework is currently given from a theoretical point of view, yet we hope that having made the motivation for it along with it\\u2019s advantages clearer now, including the addition of an experiment visualizing superiority of it, that the reviewer finds it relevant to the ICLR conference.\"]}", "{\"title\": \"Response to Review #1\", \"comment\": \"We were glad to see that the conceptual idea along with the innovation presented in the Dual Information Bottleneck ($\\\\rm{dualIB}$) framework was well understood.\\nWe thank the reviewer for raising critical points we failed to address and/or properly explain, we have revised the paper accordingly as elaborated in the general official comment posted and will now relate to specific points raised by the reviewer:\\n\\n- Motivation: We refer the reviewer to the general official comment posted.\\n\\n- Definitions and notations: Following the\\u00a0reviewer's remark we have added more background defining the Mutual Information (sec. 1.1 \\u201cThe Information Bottleneck method\\u201d) and adapted the notations of the expected value to the ML preferred formalism ($\\\\mathbb{E}$).\\n\\n- The $\\\\rm{dualIB}$form: The\\u00a0reviewer's comment has brought to our attention that the rationale behind the obtained\\u00a0dual\\u00a0form, that is the permutation\\u00a0of the probabilities distribution in the rate distortion function was not made clear in the text, we have added an explanation in the text (sec. 3 \\u201cThe Exponential Family dualIB\\u201d) and would like to elaborate it here:\\nGiven the above motivation, we looked for the minimization problem preserving the\\u00a0structure on the decoder, $p(y | \\\\hat{x})$, an exponential form (not the Bayes optimal we get by minimizing the IB Lagrangian).\\nA natural alteration achieving this goal is the geometric mean (rather than an arithmetic one) from which it follows that the distortion function minimizing it is the geometric dual distortion; the KL divergence given by the permutation of the distributions.\\u00a0\\n\\n- VAEs: We agree that this could be interesting but unfortunately within the scope of the paper we could not include this analysis.\\n\\n- New empirical results: While this work is mainly theoretical and intends to lay down the formalism of the $\\\\rm{dualIB}$ framework we agree that an empirical result could better convey it\\u2019s advantages and thus we added a new experiment in sec. 3 \\u201cThe Exponential Family dualIB\\u201d:\\n\\nIn the new experiment we apply the $\\\\rm{IB}$ and $\\\\rm{dualExpIB}$ to data sampled from a Normal distribution. \\nThe visual results, given in Figure 4(a) now showcase the benefits and improvements of the framework in this scenario.\\u00a0\\nIn Figure 4(a) we provide examples of the predicted distributions in which one can see the ability of the $\\\\rm{dualExpIB}$ to better generalize and capture the original distribution even though it is given the empirical sufficient statistics of the sample.\\nNext, we aim to visualize the superiority in prediction; we wish clarify that this in this we referred to the maximization of the information on the\\u00a0newly added\\u00a0prediction variable, $\\\\hat{Y}$, as it rises naturally from the new Lagrangian.\\u00a0\\nIn Figure 4(b) we show the $\\\\hat{Y}$\\u2019s information plane it for the new example. The superiority of $\\\\rm{dualExpIB}$ over the $\\\\rm{IB}$ is apparent.\\nFurthermore, we evaluate the prediction ability over the examined cases using the $\\\\mathcal{L}_1$-expected loss, testing the accuracy between the obtained predictions of each method to the original population distribution, see Figure 4(c).\\nWe see that for all values of the trade off parameter, $\\\\beta$, the loss is minimized by the $\\\\rm{dualExpIB}$. \\n\\n\\nTo conclude, We hope that we were able to portray the advantages and interest in the $\\\\rm{dualIB}$\\u00a0framework. \\nThe framework is given from a theoretical point of view but we hope that with a clearer view of the motivation along with visual examples and elaboration on\\u00a0possible applications,\\u00a0it\\u2019s importance and relevance to the ICLR conference is apparent.\"}", "{\"title\": \"A general response clarifying revisions\", \"comment\": \"We thank the reviewers for their thoughtful reviews. While we were glad to see that generally the conceptual idea behind the new Dual Information Bottleneck ($\\\\rm{dualIB}$) framework was conveyed we feel that that the reviewers pointed out a major flaw in out presentation - lack of properly explaining the\\u00a0motivation\\u00a0behind the dualIB framework and from there on it's\\u00a0superiority over\\u00a0the original information\\u00a0bottleneck ($\\\\rm{IB}$).\", \"we_have_thus_modified_the_text_as_follows\": \"- Revised the abstract\\n- Added a subsection to the introduction, sec 1.5 \\u201cdrawbacks of the IB\\u201d.\\n- Extended the discussion in sec. 1.6 \\u201ccontributions of this work\\u201d.\\n- Performed an additional experiment specifically relating to the suggested advantages (see sec. 3, Figure 4) of our framework. \\n\\nWith that, we would like to bring the gist of the motivation here as well:\\n\\nThe main theoretical and practical drawback of the $\\\\rm{IB}$\\u00a0is its reliance on the explicit knowledge of the joint distribution between the input variable $X$ and label $Y$, $p(x,y)$, in a completely non-parametric manner.\\u00a0Often times, the data has natural parametrization (for example biological data-set that is known to follow an analytical model) or that given a finite sample of the data assumptions regarding the \\\"features\\\" can be made,\\u00a0finding the sufficient statistics\\u00a0representation.\\u00a0\\nSuch parametrization allows for a\\u00a0dimensionality reduction (instead of tracking the entire joint\\u00a0distribution\\u00a0$p(x,y)$\\u00a0we may track the\\u00a0sufficient statistics) and for generalization to patterns we haven\\u2019t observed in the given sample.\\u00a0\\nHaving that the IB is non-parametric, our motivation was to find an\\u00a0approximation to it\\u00a0that\\u00a0uses\\u00a0this information to (a) simplify the task and (b) obtain a better generalization as it follows the inner structure of\\u00a0the distribution.\\nGiven this aim, by requiring that the decoder will preserve a parametric structure a new\\u00a0rate-distortion problem has emerged, the Dual Information\\u00a0Bottleneck ($\\\\rm{dualIB}$).\\nAs we show in the text the\\u00a0framework is\\u00a0applicable to any problem; obtaining a good approximation to the IB and it follows the same\\u00a0characteristic behaviour. \\u00a0\\nThe dualIB's superiority emerges in the\\u00a0case that the rule distribution, $p(y|x)$, can be approximated by an exponential family.\\nIn the\\u00a0exponential dual information bottleneck\\u00a0scenario ($\\\\rm{dualExpIB}$) we show that the obtained solutions,\\u00a0encoder-decoder, follow the same\\u00a0exponential form with the parameters defined by the original parameters through expectations. This achieves the desired property of preserving the natural form of the distribution thus obtaining a better generalization, as we now emphasize by providing new experimental results.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Description:\\n\\nThe information bottleneck (IB) is an information theoretic principle for optimizing a mapping (encoding, e.g. clustering) of an input, to trade off two kinds of mutual information: minimize mutual information between the original input and the mapped version (to compress the input), and maximize mutual information between the mapped input and an output variable. It is related to a minimization of an (expected) Kullback-Leibler divergence betwen conditional distributions of an output variable.\\n\\nIn this paper, instead of the original IB, authors consider a previously presented dual problem of Felice and Ay, where the Kullback-Leibler divergence is minimized in the reverse direction: from the conditional distribution of output given encoding, p(y|hat x), to the conditional distribution given the original input, p(y|x).\\nThe \\\"dual problem\\\" itself has a more complicated form than the original IB, authors claim this is \\\"a good approximation\\\" of the original bottleneck formulation, and aim to prove various \\\"interesting properties\\\" of it. \\n\\n- An iterative algorithm (Algorithm 1) similar to the original IB algorithm but with a few more steps is provided.\\n\\n- A theorem about critical points where cardinality of the representation changes is given, similar to the IB critical points, and another theorem about difference of curves on an information plane between the IB and dual-IB solutions.\\n\\n- Authors also show that if the true conditional distribution of outputs given inputs has an exponential-family form, the dual-IB decoder also has a form in the same family, which is said to reduce computational complexity of the algorithm.\", \"evaluation\": \"This is an entirely theory-based paper; although an algorithm is given, it is not instantiated for any concrete representation learning task, and no experiments at all are demonstrated.\\n\\nOverall, I feel the motivation is not clear and strong enough. The abstract does not illustrate the importance of the mentioned \\\"interesting properties\\\" well enough for concrete tasks. Reading the paper, the clearest motivations seem to be improving computational complexity, and having a clearer connection to output prediction in cases where the predictor may be sub-optimal. However, authors do not quantify these well:\\n\\n- The computational complexity improvement is not made clear (quantified) in a concrete IB optimization task: it seems it is only for exponential families, and even for them only affects one part of the algorithm, reducing its complexity from dim(X) to d; the impact of this is not tried out in any experiment.\\n\\n- For output prediction, authors motivate that dual-IB could have a more direct connection e.g. \\\"due to finite sample, in which it can be very different from the one obtained from the full distribution\\\"). Authors further claim that the dual-IB formulation can \\\"improve the generalization error when trained on small samples since the predicted label is the one used in practice\\\". However, this is not tested at all: no prediction experiments, no quantification of generalization error, and no comparisons are done, thus the impact of the clearer connection to output prediction is not tested at all, and no clear theorems are given about it either.\\n\\nThe property that the algorithm \\\"preserves exponential form of the original data distribution, if one exists\\\" is interesting in principle, but it is unclear if any real data would anyway precisely have such a distribution; what happens if the data is not exactly in an exponential family?\\n\\nIn its current state the paper, although based on an interesting direction, in my opinion does not make a sufficient impact to be accepted to ICLR.\", \"other_comments\": \"\\\"Application to deep learning\\\" mentioned in Section 1.5 is only a sincle sentence in the conclusions.\\n\\nThere have been some other suggested alternative IB formulations, for example the Deterministic IB of Strouse and Schwab (Neural Computation 2017) which also claim improved computational efficiency. How does the method compare to those?\\n\\nSection 1.5 claims the algorithm \\\"preserves the low dimensional sufficient statistics of the data\\\": it is not clear what \\\"preserves\\\" means here, certainly it seems the decoder in Theorem 7 uses the same kinds of sufficient statistics as in the original data, but it is not clear that hat(x) would somehow preserve the same values of the sufficient statistics.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper introduces a variant of the Information Bottleneck (IB) framework, which consists in permuting the conditional probabilities of y given x and y given \\\\hat{x} in a Kullback-Liebler divergence involved in the IB optimization criterion.\\n\\nInterestingly, this change only results in changing an arithmetic mean into a geometric mean in the algorithmic resolution.\\n\\nGood properties of the exponential families (existence of non-trivial minimal sufficient statistics) are preserved, and an analysis of the new critical points/information plane induced is carried out.\\n\\nThe paper is globally well written and clear, and the maths are rigorously introduced and treated. Two minor comments would concern the definition of the Mutual Information (MI), which could be recalled to help the unfamiliar reader and improve self-containedness, and the notation of the expected value (\\\\langle \\\\rangle_p), unusual in Machine Learning where \\\\mathbb{E} is often preferred.\", \"another_point_that_could_be_enhanced_is_the_intuition_behind_the_ib_and_dualib_criteria\": \"a small comment on their meaning/relevance as well as the rationale/implication of the probabilities permutation would be valuable addition.\\n\\nExhibiting a link with variational autoencoders (VAEs), and expliciting the differences in a VAE framework between the two criteria could also represent an interesting parallel, especially for machine learning oriented readers.\\n\\nOne of the main drawback of the present paper is however the lack of convincing experiments. Graphs showing the good behavior of the introduced framework are ok, but the clear interest of using dualIB rather than IB could be emphasized more. In particular, it does not seem that the issues about IB raised in the abstract (curse of dimensionality, no closed form solution) are solved with dualIB. Furthermore, dualIB optimizes the MI of the predicted labels, which is claimed to be beneficial contrary to the MI on the actual labels. However, no empirical demonstration of this superiority is produced, which is a bit disappointing.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a new \\\"dual\\\" variant of the Information Bottleneck framework. The IB framework has been the subject of many papers in past years, with a focus on understanding the inner workings of deep learning. The framework poses machine learning as optimizing an internal representation to tradeoff between retaining less information about the input features and retaining more information about the output label (prediction). The existing framework measures the retained information about the prediction via mutual information, which can be expressed as a KL divergence. The new dual framework reverses the arguments of this divergence.\\n\\nThe paper shows that this dual IB closely mirrors the original IB while having several additional nice mathematical properties. In particular, it can be shown that for exponential families the dual IB representation retains the exponential form.\\n\\nOverall, I think this paper adds a meaningful new perspective to the IB framework and the analysis appears to be thorough. As such, I support acceptance. \\n\\nThe paper could be improved by giving more interpretation of the formal results and experiments - i.e. tying this framework back to the higher-level questions and explaining what the quantities mean.\\n\\nFigures 1, 2 part a: The meaning of both axes is unclear. The horizontal axis beta is a lagrangian parameter with no meaning outside the framework. The vertical axis Pr[y=0|\\\\hat x] has no semantics since the problem has not been defined. What is the reader meant to take away from these figures?\\n\\nEquation 3, part i: The numerator should have a beta subscript. The meaning of the denominator is not clear (it is a normalizing constant).\"}" ] }

[ "Sungsoo Ahn", "Younggyo Seo", "Jinwoo Shin" ]

[ "deep reinforcement learning", "combinatorial optimization" ]

{ "note_id": [ "dkKrN3OvEp", "BylIEntsjS", "rJl15Er9or", "H1lMS4B5oS", "S1lJkEr9jH", "Skl3tmS9iB", "rklWm7H5oB", "H1g_kS26YH", "HkxCFnOTFS", "Hkghsm4cFr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731638, 1573784621988, 1573700743454, 1573700666387, 1573700567108, 1573700484493, 1573700376936, 1571828959967, 1571814534377, 1571599268131 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1754/Authors" ], [ "ICLR.cc/2020/Conference/Paper1754/Authors" ], [ "ICLR.cc/2020/Conference/Paper1754/Authors" ], [ "ICLR.cc/2020/Conference/Paper1754/Authors" ], [ "ICLR.cc/2020/Conference/Paper1754/Authors" ], [ "ICLR.cc/2020/Conference/Paper1754/Authors" ], [ "ICLR.cc/2020/Conference/Paper1754/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1754/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1754/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes a new way to formulate the design of the deep reinforcement learning that automatically shrinks or expands decision processes.\\n\\nThe paper is borderline and all reviewers appreciate the paper and gives thorough reviews. However, it not completely convince that it is ready publication. \\n\\nRejection is recommended. This can become a nice paper for next conference by taking feedback into account.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Common response to all reviewers: summary of revision\", \"comment\": \"Dear all reviewers,\\n\\nFirst of all, we appreciate your valuable time and effort to review our paper. \\n\\nIn the revised draft, we enhanced our manuscript with a substantial amount of materials, including:\\n\\n- a separate description of related works (Section 2), suggested by R3\\n- a graphical illustration of our framework (Appendix A), suggested by R3 \\n- experimental results for two more combinatorial problems (Appendix B), suggested by R2 and R3\\n- comparison with the state-of-the-art supervised learning (Appendix C), suggested by R1 and R3\\n- evaluation of CPLEX with different hyperparameters (Appendix D), suggested by R1\\n\\nAny modifications from the original draft are highlighted by \\u201cred\\u201d texts. \\n\\nThank you very much.\"}", "{\"title\": \"Response to R3 (2/2)\", \"comment\": \"Q5. A graphical representation of the whole framework would be helpful. Notion of diversification reward is missing in the description of the proximal policy optimization.\\n\\nA5. Following your suggestion, we have added the graphical representation of our framework in Appendix A of the revised draft. The notion of diversification reward was also added to the description of proximal policy optimization in the last equation of Section 3.2.\\n\\nQ6. ReduMIS [2] is nowadays dominated by more efficient kernelization/heuristic techniques for solving MIS. It is suggested to replace ReduMIS with solvers of better performance, e.g., OnlineMIS [4]. \\n\\nA6. Our implementation of ReduMIS [3] incorporates the recent graph kernelization technique by Hespe et al. [5]. We believe our choice of implementation is a strong baseline due to the following facts:\\n\\n1) Hespe et al. [5] report better performance than Chang et al. [6] under incorporating their kernelization technique into ReduMIS. \\n2) Chang et al. [6] in turn report better performance than OnlineMIS [4] under incorporating their kernelization technique into ARW [7]. \\n3) ARW [7] itself was shown to underperform compared to ReduMIS in the original ReduMIS paper [2]. \\n\\nWe have made changes in our revised draft to highlight this fact. In Section 4, first paragraph, 17th line, we state: \\u201cNote that we use the implementation of ReduMIS equipped with graph kernelization proposed by Hespe et al. 2019.\\u201d. Since specification on the state-of-the-art MIS solver can be debatable, we removed any claims on our implementation to be the state-of-the-art. \\n\\nQ7. It is suggested to compare with the supervised learning framework of Li et al. [8].\\n\\nA7. Following your suggestion, we include the comparison to Li et al. [8] in Appendix C of the revised draft. From the experiments, one observes that our framework outperforms the supervised learning framework when given a similar computational budget. We also refer our response A2 for R2.\\n\\nQ8. It is suggested to compare with the SMT solver Z3 [9] that was used for solving the MIS problem by Li et al. [8].\\n\\nA8. Thank you for the constructive feedback. We would like to point out that Z3 has been used as a SAT solver by Li et al. [8], instead of being used as an MIS solver. Specifically, they compare the performance of Z3 for solving the SAT problems with other MIS solvers for solving the MIS problems (reduced from the same SAT problems). We believe that using Z3 for solving arbitrary MIS problem is non-trivial, e.g., Li et al. [8] states in Section 5.2, 2nd paragraph, 11th line of their paper as follows: \\u201cNote that Z3 directly solves the SAT problem and cannot solve any transformed MIS problem on the SATLIB dataset\\u201d. \\n\\n[2] Sebastian Lamm, Peter Sanders, Christian Schulz, Darren Strash and Renato F. Werneck, \\u201cFinding Near-Optimal Independent Sets at Scale\\u201d, In Journal of Heuristics, 2017 \\n[3] http://algo2.iti.kit.edu/kamis/\\n[4] Jakob Dahlum, Sebastian Lamm, Peter Sanders, Christian Schulz, Darren Strash and Renato F. Werneck, \\u201cAccelerating Local Search for the Maximum Independent Set Problem\\u201d, In International symposium on experimental algorithms. Springer, 2016\\n[5] Demian Hespe, Christian Schulz and Darren Strash, \\u201cScalable Kernelization for Maximum Independent Sets\\u201d, In JEA, 2019\\n[6] Lijun Chang, Wei Li and Wenjie Zhang. \\u201cComputing a near-maximum independent set in linear time by reducing-peeling.\\u201d In SIGMOD, 2017.\\n[7] Diogo V. Andrade, Mauricio G.C. Resende and Renato F. Werneck \\u201cFast Local Search for the Maximum Independent Set Problem\\u201d, In Journal of Heuristics, 2012\\n[8] Zhuwen Li, Qifeng Chen and Vladlen Koltun, \\u201cCombinatorial Optimization with Graph Convolutional Networks and Guided Tree Search\\u201d, In NeurIPS, 2018\\n[9] Leonardo Mendon\\u00e7a de Moura and Nikolaj Bj\\u00f8rner, \\u201cZ3: An efficient SMT solver\\u201d, In TACAS, 2008\"}", "{\"title\": \"Response to R3 (1/2)\", \"comment\": \"Thank you for your valuable efforts and time spent on reading our paper. We believe our framework proposes an interesting idea (as pointed out by you and R1) with promising empirical results (as pointed out by you and R2), based on clear writing (as pointed out by R1 and R2). In the revised draft, we added a description of related works (in Section 2), a graphical illustration of our framework (in Appendix A) and additional experiments (in Appendix B, C, D) to incorporate all reviewers\\u2019 comments. Any modifications from the original draft are highlighted by \\u201cred\\u201d texts. We think the revised draft has much improved by clarifying the contributions, describing the related works and validating the empirical evaluation, thanks to your comments. Our responses to all your questions are provided below.\\n\\nQ1. It would be nice to use a spelling or grammar checker. \\n\\nA1. Thank you for your feedback. We have run the grammar checking software on the revised draft and removed any possible typos. \\n\\nQ2. It is an overstatement to claim the framework to be applicable to any combinatorial problems. The paper should (a) focus the paper on the maximum independent set (MIS) problem or (b) include other APX-hard combinatorial optimization problems. \\n\\nA2. Thank you very much for the valuable suggestion regarding the direction of our paper. Following your suggestion (a), we have revised the whole draft (including the title) towards focusing on solving the MIS problem. Under the decision, we also provide experimental results on applying our framework for two variants of the MIS problem in Appendix B: the maximum weighted independent set (MWIS) problem and the prize collecting maximum independent set (PCMIS) problem. In the new applications, our framework shows its general applicability by achieving better performance than the generic integer programming solver, i.e., CPLEX. Note that ReduMIS and ES are not applicable to the considered variants of the MIS problem, which highlights the advantage of our scheme. \\n\\nQ3. Related-work section is needed to be included in the paper. \\n\\nA3. Thank you for such a constructive and informative suggestion. Following your suggestion, we have added Section 2 of the revised draft for describing the related works. We included all the papers that were mentioned in your comments. \\n\\nQ4. It is not clear whether a policy with exponentially large action space converges in polynomial time. \\n\\nA4. Thank you for making an important point. While it is non-trivial to prove the polynomial convergence of our policy, we empirically observe it converge very fast. Specifically, our policy usually takes 10 000 steps to converge, e.g., see Figure 5 (a) in the revised draft. This requirement is similar to the convergence reported for S2V-DQN [1] that has a polynomially small action space. \\n\\n[1] Elias B. Khalil, Hanjun Dai, Yuyu Zhang, Bistra Dilkina and Le Song, \\u201cLearning combinatorial optimization algorithms over graphs\\u201d, In NeurIPS, 2017\"}", "{\"title\": \"Response to R1\", \"comment\": \"Thank you for your valuable efforts and time spent on reading our paper. We believe our framework proposes an interesting idea (as pointed out by you and R3) with promising empirical results (as pointed out by R2 and R3), based on clear writing (as pointed out by you and R2). In the revised draft, we added a description of related works (in Section 2), a graphical illustration of our framework (in Appendix A) and additional experiments (in Appendix B, C, D) to incorporate all reviewers\\u2019 comments. Any modifications from the original draft are highlighted by \\u201cred\\u201d texts. We think the revised draft has much improved by clarifying and improving the contributions of our work, thanks to your comments. Our responses to all your questions are provided below.\\n\\nQ1. Some statements of the paper appear a little bit too strong.\\n\\nA1. Thank you for the suggestion. In the revised draft, we softened several sentences, e.g., \\n\\n- Abstract, 4th line: \\n(before revision) Recently\\u2026 they can automatically learn the design of a good solver without using any sophisticated knowledge or hand-crafted heuristic specialized for the target problem.\\n(after revision) Recently ... they can automatically train a good solver while relying less on sophisticated domain knowledge of the target problem.\\n\\n- Introduction, 3rd paragraph, 4th line:\\n(before revision) Then the corresponding agent can be trained (without any human guidance) based on existing training schemes of DRL.\\n(after revision) Then, the corresponding agent can be trained based on existing training schemes of DRL, e.g., ...\\n\\nQ2. Why is encoding NP-hard problem using a (discrete, finite) Markov decision process (MDP), which is rather efficient to solve, is a good idea? \\n\\nA2. Thank you for raising a very important question. While a NP-hard problem is hard to solve exactly, it is often good enough to use a heuristic for obtaining an approximate solution under a reasonable amount of time. This motivates to express heuristic as a MDP with \\u201cfinite\\u201d episode length, which is able to be optimized for generating high-quality solutions under a given time constraint. \\n\\nFurthermore, the MDP can be used to learn new problems while relying less on problem-specific domain knowledge, which is often expensive to acquire. To illustrate, we introduce the prize collecting maximum independent set (PCMIS) problem in Appendix B of the revised draft. In the PCMIS problem, the \\u201chard\\u201d constraints of a MIS, i.e., forbidding two adjacent vertices, are replaced by a \\u201csoft\\u201d penalty term. For this problem, existing MIS heuristics are not applicable without substantial changes relying on problem-specific knowledge provided by experts. On the contrary, our MDP can be trained to solve the problem with a very small modification, i.e., excluding the clean-up phase. \\n\\nQ3. Experiments on problems other than the maximum independent set (MIS) problem are missing for a complete picture. \\n\\nA3. Following your suggestion, we applied our framework to other problems in Appendix B of the revised draft, i.e., the maximum weighted independent set (MWIS) problem and the PCMIS problem. Such choices are due to our decision to focus on the MIS-like problems in the revised draft, following the suggestion of R3. In the new experiments, our framework shows its general applicability by achieving better performance than the generic integer programming solver, i.e., CPLEX. Note that ES and ReduMIS are not applicable to the considered variants of the MIS problem, which highlights the advantage of our scheme.\"}", "{\"title\": \"Response to R2 (2/2)\", \"comment\": \"Q3. Is there a case where the policy learns to fix all variables in just one step? It is not clear what encourages the policy to defer decisions.\\n\\nA3. Thank you for your interesting question. Empirically, the policy was not observed to fix all variables in just one step. Instead, it usually takes more than 30 steps to fix all variables in all our experiments. Our policy learns to defer decisions through higher rewards coming from better solving the problem. Intuitively, the deferring of decisions on vertices helps to solve the MIS problem by prioritizing making decisions on the \\u201ceasier\\u201d vertices first. Whenever easier decisions are made, the surrounding values of \\u201chard\\u201d vertices are fixed. Then, the decisions on \\u201chard\\u201d vertices become easier to be made. For example, when a \\u201chard\\u201d vertex becomes surrounded by \\u201ceasy\\u201d vertices that were excluded from the final independent set in the previous iteration, it should be included in the independent set. In Section 3, the first paragraph, the 12th line of the revised draft, we added an explanation of such intuition for a better understanding of readers.\\n\\nQ4. Please point out specific results on outperforming the existing solver.\\n\\nA4. Thank you for your suggestion. In the revised draft, we have modified the 15th line (last sentence) of the abstract to point out the specific results. It is now stated as follows: \\u201cThe reported performance of our generic DRL scheme is also comparable with that of the existing solvers for MIS, e.g., AutoDP outperforms them for the Barabasi-Albert graph with two million vertices.\\u201d. To avoid any confusion, we would also like to remark that the same statements are mentioned in the last paragraph of Section 1 and the last paragraph of Section 4.3 of both the original and the revised draft. \\n\\nQ5. Hyper-parameters of CPLEX should be adjusted for a fairer comparison.\\n\\nA5. To address your concerns, we tried setting the hyper-parameter of CPLEX to prioritize the optimization of objective in various ways. For your information, we report the performance of CPLEX under varying the optimality gap used for the stopping criterion, as in the last paragraph of Appendix E of the revised draft. However, we found that such a change of hyper-parameter tends to degrade in performance with marginal speedups. Therefore, we still keep the current settings of CPLEX in the revised draft. \\n\\nNevertheless, to incorporate your comments, we have revised the paper to clarify this point for the readers. Specifically, we added the following description as a footnote in the first paragraph of Section 4: \\u201cNote that CPLEX is also able to prove an upper bound of the solution for the MIS problem.\\u201d We also describe our algorithm to \\u201cachieve a better objective\\u201d rather than to \\u201coutperform\\u201d over CPLEX.\\n\\n\\nQ6. It would be insightful to explore the tradeoff between making many decisions and better credit assignment. \\n\\nA6. Thank you for the interesting question. We did not observe any credit assignment problem even when training on quite large graphs, e.g., a graph with 500 vertices requires action with 500 dimensions. Instead, S2V-DQN (which makes one-dimensional decision per step) suffers more from unstable training in large graphs. We think this is due to the long episode length required for the execution of S2V-DQN (known as the temporal credit assignment problem [4]). We still think your question is very interesting and hope to explore this point in the future.\\n\\nQ7. Acronym \\u201cADP\\u201d is already known in the optimization community as Approximate Dynamic Programming. \\n\\nA7. Thank you very much for your information. Following your suggestion, we have changed our acronym to \\u201cAutoDP\\u201d. \\n\\n[4] Richard S. Sutton, \\u201cLearning to predict by the methods of temporal differences\\u201d, In Springers, 1988\"}", "{\"title\": \"Response to R2 (1/2)\", \"comment\": \"Thank you for your valuable efforts and time spent on reading our paper. We believe our framework proposes an interesting idea (as pointed out by R1 and R3) with promising empirical results (as pointed out by you and R3), based on clear writing (as pointed out by you and R1). In the revised draft, we added a description of related works (in Section 2), a graphical illustration of our framework (in Appendix A) and additional experiments (in Appendix B, C, D) to incorporate all reviewers\\u2019 comments. Any modifications from the original draft are highlighted by \\u201cred\\u201d texts. We think the revised draft has much improved by validating the empirical fairness and clarifying the intuition behind our framework, thanks to your comments. Our responses to all your questions are provided below.\\n\\nQ1. The proposed framework uses hard constraints of the maximum independent set (MIS) problem, which makes it unfair to compare with S2V-DQN [1]. \\n\\nA1. We run the code [2] provided by the authors of S2V-DQN, which also utilizes similar hard constraints. Hence, we think our empirical comparison is fair. To explain in more detail, we first point out that S2V-DQN solves the minimum vertex cover problem (identical to the MIS problem). During its greedy selection of vertices, any vertices that are already fully covered by the intermediate vertex cover are automatically excluded. Such a process is very much similar to our use of hard constraints. \\n\\nQ2. Comparison with the supervised learning framework proposed by Li et al. [3] is needed since they also solve the MIS problem.\\n\\nA2. Following your suggestion, we include the comparison to Li et al. [3] in Appendix C of the revised draft. From the experiments, one can observe that our reinforcement learning framework outperforms the supervised learning framework when given a similar computational budget. Here, we remark that Li et al. [3] additionally introduced the idea of combining local search and graph reduction between intermediate steps. Although such an idea is applicable to both the supervised and reinforcement learning frameworks, we disable them in our experiments for a fair comparison, i.e., to use a similar amount of domain knowledge. We found that the strong results reported by Li et al. [3] can be reproduced when the local search and graph reduction are added to the supervised learning framework. \\n\\n[1] Elias B. Khalil, Hanjun Dai, Yuyu Zhang, Bistra Dilkina and Le Song, \\u201cLearning combinatorial optimization algorithms over graphs\\u201d, In NeurIPS, 2017\\n[2] https://github.com/Hanjun-Dai/graph_comb_opt\\n[3] Zhuwen Li, Qifeng Chen and Vladlen Koltun, \\u201cCombinatorial Optimization with Graph Convolutional Networks and Guided Tree Search\\u201d, In NeurIPS, 2018\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper proposes a Deep RL approach called Auto-Deferring Policy (ADP) to learning a policy for constructing solutions for the Maximum Independent Set (MIS) problem. Rather than constructing a solution one variable per episode step, the policy can make decisions about multiple variables per step, as well as defer decisions to later steps. At each step the MIS constraints of a valid solution are checked and decisions that violate the constraints are reverted, and any additional decisions that are automatically implied by the decisions so far and the MIS constraints are taken. The policy and value function are parameterized as a graph convolutional network to make use of the graph structure of the problem. A diversity regularizer that encourages the policy to generate diverse final solutions is used for training. Results show that the approach is able to match the objective value of the state-of-the-art solvers on several datasets, and is able to outperform S2V-DQN on several datasets when neither approach is trained on them.\", \"pros\": [\"Extensive experiments have been done on several datasets, both synthetic and real. The ablation and generalization experiments are particularly valuable for getting insight into the performance of the algorithm.\", \"Comparison to CPLEX and ReduMIS strengthens the results.\", \"The paper is reasonably well-written and easy to follow. Figures 2 and 3 are especially useful for quickly understanding key ideas.\"], \"cons\": [\"The update and clean-up phases inject significant domain knowledge about MIS into the policy. At a high level the idea is similar to unit propagation in Boolean Satisfiability (SAT) solvers or domain propagation in Mixed Integer Programming (MIP) solvers, i.e., the hard constraints of the problem can be used to make decisions in the search for a solution. Both SAT and MIP solvers have been shown to rely on it to significantly improve their search. Improvements over S2V-DQN could be due to this extra built-in domain knowledge, rather than any improvements related to learning. For a fairer comparison, either this knowledge should be removed from ADP as an additional ablation study, or it should somehow be given to S2V-DQN as well. Without such a comparison, it is not clear that the improvements are really due to the new ideas like auto-deferring, diversity regularizer, etc.\", \"A comparison to Li, Chen, and Koltun, NeurIPS\\u201918 is needed since they have shown strong results for MIS and they have made the code available online. Although their approach is supervised learning, it would still be useful to assess an RL approach\\u2019s performance relative to SL.\"], \"additional_comments\": [\"It is not clear what encourages the policy to defer decisions, rather than making all the {0,1} decisions for all variables in one step or a small number of steps. Does this behaviour of the policy arise naturally via training, or does some regularization need to be applied to ensure that the policy doesn\\u2019t prematurely fix all variables? For example, as a way to avoid/reduce variance in the reward signal, the policy may learn to fix all variables in just one step -- does this issue arise?\", \"The abstract says \\\"The reported performance of our generic DRL scheme is also comparable with that of the state-of-the-art solvers specialized for MIS, e.g., ADP outperforms them for some graphs with millions of vertices.\\\" Can you please point out the specific results in the paper that this sentence is referring to?\", \"Comparison to CPLEX: The paper mentions that \\u201cwe observe that our algorithm outperforms the CPLEX solver on ER-(100, 200) and ER-(400, 500) datasets, while consuming a smaller amount of time\\u201d, and \\u201cIt is remarkable to observe that ADP outperforms the CPLEX solver on both datasets under reasonably limited time.\\u201d Note that CPLEX not only optimizes the objective, but also proves a bound on the objective, while ADP only does the former. So this is not a fair comparison. A fairer comparison would be to set CPLEX hyperparameters to give higher priority to optimizing the objective and then measure the time CPLEX took to first reach a particular objective value rather than the time to \\u201csolve\\u201d an instance. It can be the case that a given objective value is achieved quickly but then proving a bound requires much longer. Also there is no mention of the optimality gap used as a stopping criterion for CPLEX.\", \"One potential advantage of making one decision per step is that the credit assignment problem may be simpler compared to making many decisions simultaneously. It would be insightful to explore this tradeoff in more detail.\", \"The acronym ADP is somewhat well-known in the optimization community as Approximate Dynamic Programming (http://adp.princeton.edu/), so it would be helpful to use a different one.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper introduces auto-deferring policies (ADPs) for deep reinforcement learning (RL). ADPs automatically stretching or shrinking their decision process, in particular, deciding whether to finalize the value of each vertex at the current stage or defer to determine it at later stages. ADPs are evaluated on maximum independent set problems.\\n\\nThe paper is in principle well written and structured. Some statements of the paper appear a little bit too strong. For instance, saying that deep RL approaches \\\"can automatically learn the design of a good solver without using any sophisticated knowledge or hand-crafted heuristic specialized for the target problem\\\" is misleading as thee designer of the RL setup is putting a lot of knowledge into the design. Likewise the statement \\\"without any human guidance\\\" is not true, at least at the current stage. It would be great to acknowledge this by softening this statement. \\n\\nThe basic idea is also fine, learning to expand or not nodes in the current fringe of the combinatorial solver. However, being an informed outsider, I would like to understand more why encoding NP-hard problem using an (discrete, finite) MDP, which is rather efficient to solve, is a good idea. Moreover, while the focus on MIS is justified in the paper, showing the (potential) benfit on other tasks such as TSP is useful to convince the reader that ADPs apply across different problem classes. Without, it is not clear whether ADPs work fine on other problem classes (even if one may expect that this is the case). \\n\\nAnyhow, the main idea of implementing two independent agents/networks to implement a reward signal that rewards deviation is interesting. Moreover, the experimental results show that the approach works. It actually manages to be on par with ReduMIS. However, it does not really improve upon this well-known heuristic. Hence some experiments across different problems would really be beneficial. Without, it is just interesting to see that RL can also come up with good heuristics but the existing heuristic already works pretty well.\\n\\nTo sum up, a very nice idea that shows promise but experiments on other problems is missing for a complete picture. Also some statements should be soften.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper aims at solving graph-based combinatorial optimization problems using a new paradigm for Deep Reinforcement Learning. In contrast with standard Markov Decision Processes used for combinatorial DRL, the authors advocate the use of \\u201cdeferred\\u201d MDPs capturing more complex actions (which can choose a subset of nodes), and more sophisticated transitions, which combine an update phase together with a cleanup phase. For the DRL architecture, an Actor-Critic framework is trained using a proximal policy optimization approach. The paper specifically focuses on the Maximum Independent Set (MIS) problem. Comparative results on various instances reveal that this approach clearly outperforms the S2V-DQN algorithm and is competitive with some usual combinatorial solvers (CPLEX, ReduMIS).\\n\\nThe overall approach is interesting and the experimental results look promising, but it is quite difficult to accept the paper in its current state due to the following reasons:\\n\\n(a) Though the paper is comprehensible, it is littered with spelling mistakes (just take the first sentence: problem -> problems). For the camera-ready version, it would be nice to use a spell/grammar checker.\\n\\n(b) The authors focus on the MIS problem, but in the introduction, they claim that their approach can be applied to \\u201cany\\u201d combinatorial optimization task. Well, that is a bit of an overstatement, due to the immense variety of NP-hard optimization tasks! On the one hand, the MIS problem is already taking an important place in many areas of computer science, and I would suggest concentrating on this task - by changing the title and the summary of the paper, and by just pointing out in the conclusion that the present DRL approach might be applicable to other problems. On the other hand, if the authors are convinced that their DRL approach is generic, then the paper should include other well-known APX-hard problems such as, for example, Min Feedback Node Set, Min Balanced Cut, or Min Set Cover. \\n\\n(c) The paper is missing a related-work section. To this point, the MIS problem should be presented in more detail with some theoretical results about its approximability (to within a linear factor, as shown by Boppana and Halldorsson, 1992), and its fixed-parameter tractability (it is W[1]-hard). From an algorithmic viewpoint, the ReduMIS technique is mentioned, but there are many other heuristic approaches and kernelization methods. Notably, OnlineMIS (Dahlum et. al. ESA\\u201916) is orders of magnitude faster than ReduMIS on various large instances. Furthermore, the kernelization technique developed by Chang et. al. (SIGMOD\\u201917) and its parallel version (Hespe et. al., JEA\\u201919) is known to be much more efficient than the kernelization algorithm (VC-Solver) suggested in ReduMIS. Besides heuristic algorithms, very little is said about deep learning approaches for solving MIS. To this point, a more detailed presentation of Dai et. al.\\u2019s approach (NIPS-17) would be relevant, by explaining the similarities and differences with your approach. Finally, nothing is said about the recent deep architecture proposed by Li et. al. (NIPS\\u201918) for solving MIS. Though they adopt a supervised learning approach, their GCN architecture shares strong similarities with the Actor-Critic framework in the present paper.\\n\\n(d) The deep auto-deferring policy, presented in Section 3 looks interesting, but it is quite unclear. The main idea lies in deferred MDPs, for which the size of the action space is in $\\\\Omega(3^N)$ because any subset of nodes can be chosen at each state. For such huge MDPs, it is not clear that we can converge to a stable policy in polynomial time. Furthermore, the authors are mixing Actor-Critic (using a graph convolution net) and Proximal Policy Optimization, together with diversification rewards, which makes the overall framework unintelligible. A graphical representation for explaining this complex framework would be welcome. Furthermore, I don\\u2019t see how the notion of diversification reward is implemented in the PPO approach for training the agent.\\n\\n(e) In the experiments, some competitors are missing. As mentioned above, ReduMIS is nowadays dominated by more efficient kernelization/heuristic techniques for solving MIS. I would suggest to instead use OnlineMIS. Moreover, the choice of S2V-DQN is relevant as an RL competitor, but it would be interesting to also use the GCN architecture (Li et. al., NIPS\\u201918) as a supervised competitor. Finally, the SMT solver Z3 is known to be quite efficient for MIS instances (as already observed by Li et. al.), and a comparison with it would indicate whether ADP is competitive (or not) with the best generic solvers.\"}" ] }

[ "Christian Tomani", "Florian Buettner" ]

[ "deep learning", "uncertainty", "calibration", "domain shift", "robustness" ]

{ "note_id": [ "NMzDAf6POZ", "SJlh3Z5FoS", "SkxbtWOKsS", "HklDuBPFsS", "H1lvk4vYoB", "Hyggr7vFoH", "HyeT9fvFoS", "H1xalm-RKr", "BklNA9L6tS", "r1leOMl3tH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731610, 1573654964006, 1573646713245, 1573643631038, 1573643231122, 1573643064319, 1573642901168, 1571848949286, 1571805900453, 1571713639998 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1753/Authors" ], [ "ICLR.cc/2020/Conference/Paper1753/Authors" ], [ "ICLR.cc/2020/Conference/Paper1753/Authors" ], [ "ICLR.cc/2020/Conference/Paper1753/Authors" ], [ "ICLR.cc/2020/Conference/Paper1753/Authors" ], [ "ICLR.cc/2020/Conference/Paper1753/Authors" ], [ "ICLR.cc/2020/Conference/Paper1753/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1753/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1753/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes an algorithm to produce well-calibrated uncertainty estimates. The work accomplishes this by introducing two loss terms: entropy-encouraging loss and an adversarial calibration loss to encourage predictive smoothness in response to adversarial input perturbations.\\n\\nAll reviewers recommended weak reject for this work with a major issue being the presentation of the work. Each reviewer provided specific examples of areas in which the paper text, figures, equations etc were unclear or missing details. Though the authors have put significant effort into responding to the specific reviewer mentions, the reviewers have determined that the manuscript would benefit from further revision for clarity. \\n\\nTherefore, we do not recommend acceptance of this work at this time and instead encourage the authors to further iterate on the manuscript and consider resubmission to a future venue.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to all reviewers\", \"comment\": \"We thank all reviewers for their constructive comments and we have addressed individual concerns in point-by-point responses below.\\nFollowing the suggestions of all reviewers we have revised the manuscript to improve clarity and mathematical rigor, in particular in section 2 where we introduce ECE and our loss terms. \\nIn addition, we have also added the additional experiments suggested by the reviewers: We have now extended our ablation study to analyse the influence of both loss terms and the robustness of FALCON wrt hyperparameter choice. These experiments illustrate that while individual loss terms yield improvements over many baselines, the combination results in consistently better calibration under domain drift. \\nFurthermore, we have added experiments on two additional baselines, namely temperature scaling and SVI and show that FALCON also yields a better calibration under domain drift also compared to these modelling approaches. \\nFinally, we have added further evaluations and first demonstrated a better OOD perfromance of our method over baselines using the entropy and boxplots illustrating overall distribution of confidence scores. Second, we further assessed limitations of our method related to under-confident scores using empirical CDFs.\\n\\nWe hope we could address all concerns in a satisfactory manner and hope that the reviewers will re-evaluate their scores in light of our response and revision.\", \"our_revisions_have_resulted_in_the_following_changes_in_the_manuscript\": \"\", \"section_1\": [\"minor changes to make separation in related work and our contribution explicit (following Reviewer2).\", \"Section 2.1.1:\", \"Add mathematical rigor and elaborate on eq. 1 and 2 (R2)\", \"Move description of specific perturbations to Experiments (R2)\", \"Add comment on limitations of ECE for high noise levels and change in label set (R3)\", \"Section 2.2.1:\", \"Replace the formulation on misleading evidence and clarify that loss term only operates on wrong/negative predictions (R3, R1)\", \"Remove our claim that the \\u201closs surface remains largely unchanged\\u201d and instead clarify that the loss terms do not directly affect the reconstruction loss and speculate that this is a reason why in contrast to e.g. EDL our model generalises to more complex learning problems/architectures (R2, R3)\", \"Section 2.2.2:\", \"Add mathematical rigor and more details when introducing loss term L_adv (R1, R2, R3)\", \"Add reference to ablation studies and robustness analysis (R3,R1)\"], \"section_3\": \"Add reference to additional baselines\\n\\nSection 3.1: Move former Table 1 to Appendix (now table S2) to make space for clarifications suggested by the reviewers. \\n\\nSection 3.2: \\n- Add discussion on underconfidence of EDL and FALCON for small perturbations and reference appendix A.6 (R3) \\n- Add discussion on OOD performance and quantification via entropy (R3)\", \"section_4\": \"Shorten the discussion to make space for the clarifications suggested by the reviewers.\", \"appendix\": [\"Algorithm 1: add new states to clarify computation of L_adv\", \"A.4: Extend ablation study to second loss terms and add robustness analysis (R3, R2)\", \"A.5: Add results for additional baselines\", \"A.6: Add discussion on under-confidence of EDL and FALCON for small perturbations; add empirical CDF and boxplots of confidence scores for perturbation 0 (test set) to further quantify and illustrate extend of under-confidence and demonstrate it only affects a small fraction of samples for FALCON (R3)\", \"A.7 Add figure S7 showing entropy of all baselines under max. perturbation (OOD samples) for MNIST, CIFAR and sequential MNIST to demonstrate FALCON also outperforms baselines in terms of entropy (R3). Add Figure S8 to further illustrate this with boxplots depicting the overall distribution of OOD confidence scores.\"]}", "{\"title\": \"We have now improved the presentation and revised the manuscript following the suggestions of the reviewer\", \"comment\": \"We thank the reviewer for the constructive suggestions on how to improve the presentation and are glad to read the reviewer found our submission was \\\"good\\\" and \\\"has an interesting idea\\\". We have now improved the presentation and revised the manuscript following the suggestions of the reviewer.\\n\\n1. In the introduction we now clarify where we describe related work and focus more on the \\\"pitch\\\". \\n\\n2. We have now clarified and elaborated on all entities in both equations. Eq. 1 is the formal definition of calibration, leading to the definition of miscalibration/calibration error (eq 2). The expectation in eq 2 is then approximated as the expected calibration error (ECE) eq 3, which in turn is used throughout the manuscript to quantify calibration under domain drift and, as pointed out by the reviewer, motivates the FALCON loss term L_{adv}. \\n\\n3. The reviewer is correct. ECE is defined independent of any type of specific domain shift and was indeed recently popularised when used to quantify in-domain calibration. Since in practical applications it is a priori not clear what type or magnitude of domain drift will occur, we evaluate calibration under domain drift for our approach (and numerous baselines) on 10 different perturbations and 10 different noise levels not seen during training that were chosen to represent different ways of domain drift. We have now clarified this and moved details to Experiments as suggested. \\n\\n4. We agree that our statement regarding the loss surface was not very clear and worded too strongly. We have rephrased this section to reflect that in comparison to probabilistic methods such as EDL or MNF, the reconstruction loss remains unchanged from the standard categorical cross entropy loss. Our additional loss terms do not directly affect accuracy and we hypothesize that this may be a reason why in contrast to EDL our approach yields SOTA accuracies also for more difficult learning problems (e.g. sequential MNIST with LSTMs, Tables S2 and S3 in Appendix A.7). We also agree that application to LSTMs/GRUs/more complex architectures depends on the type of BNN \\u2013 while MC-dropout is straight-forward to apply, more complex state-of-the-art approaches such as multiplicative normalizing flows are very hard to apply to architectures that contain layers that are more complex than dense layers or convolutional layers. For example for LSTM cells but also ResNET architectures, first, forward passes using local reparameterizations to yield a posterior based on multiplicative normalising flows have to be derived from scratch in a non-trivial manner and second, result in a high computational complexity. In contrast, FALCON only requires the addition two simple loss terms to the existing cross entropy loss. We now clarify this in the manuscript. \\n\\n5. L_{adv} is the adversarial loss term which we now motivate and describe more extensively in section 2.2.2; for improved clarity we have now also expanded the formula along with a more detailed discussion around the definition.\"}", "{\"title\": \"We have added the ablation study and new baselines suggested by the reviewer and improved clarity.\", \"comment\": \"We thank the reviewer for constructive suggestions and have revised the manuscript accordingly.\\n\\nClarifications and consistency. We have followed the reviewer and corrected the linestyle in figure 4 and clarified explanations around L_S and L_{adv}. \\n\\nAdditional experiments. We now present an ablation study, illustrating that while the individual terms yield substantial improvements over the L2-dropout baseline, the combination of both loss terms results in a consistently better ECE compared to one term only. In addition, we include two more baselines, namely temperature scaling and an additional Bayesian neural network (SVI based on flipout), as suggested by the reviewer. In concurrent work ([1] referenced by reviewer #3), Ovadia et al. show that temperature scaling, while improving in-domain calibration, is not beneficial for calibration under domain-drift, which is also confirmed by our new experiments. In addition, we observe that SVI results in a decreased accuracy compared to other baselines and while calibration under domain drift improves in comparison to L2-dropout, it is substantially worse than FALCON and also EDL (Table S1 and Figure S3 appendix A.2). In total we now evalute 4 probabilistic baselines: Dropout as Bayesian approximation, Multiplicative Normalizing Flows, Evidential Deep Learning (predictions are modelled by placing a Dirichlet distribution on the class probabilities) and SVI based on flipout.\"}", "{\"title\": \"Point-by-point response to concerns regarding clarity (point 2)\", \"comment\": \"We thank the reviewer for suggestions on how clarity can be improved and have revised the manuscript accordingly.\\n\\n(a) \\\"high confidence\\\" refers to miscalibration in the sense that at high noise levels where accuracy approaches random levels (right panel), entropy should approach maximum entropy for calibrated predictions. In contrast, averaged over all predictions, entropy actually saturates at 0.5 (22% maximum entropy), indicating that wrong predictions are indeed often made with high confidence. EDIT: New Figure S8 in the appendix shows that the median confidence score is 75.8% at the highest noise level, also confirming this frequent overconfidence.\\n\\n(b) We now clarify that for wrong predictions our loss encourages a uniform distribution across all C classes and weaken our claim regarding the loss surface, putting it in context with EDL (reconstruction loss remains cross entropy for FALCON and entropy loss and adversarial loss do not directly affect accuracy). \\n\\n(c) We thank the reviewer for pointing us to the L2 norm in in the description of the adversarial loss. Unfortunately in the initial submission we had a typo in the formula that is now corrected: in the actual implementation the L2 norm is taken of the difference between confidence and acc(B_m) across all samples in the minibatch (rather than of the scalar \\u2013 we apologise for this mistake). This was chosen to penalise greater deviations of the confidence from accuracy. The reviewer is also correct regarding acc(B_m): accuracy is computed before each update and treated as constant. We have now expanded the corresponding section and explain this in more detail; we have also reformulated the loss function to correct the L2 issue and yield a more easily interpretable formulation. Finally, we have extended Algorithm 1 to better reflect the computation of all loss terms (Appendix A.3). \\n\\n(d) We have fixed the legend. FALCON is indeed also slightly underconfident for small perturbations and we now discuss and illustrate this in more detail. In Figures S4 and S5 in Appedix A.6 we now directly compare the confidence levels for EDL and FALCON for y-zoom, both via the sample used in Fig. 4 and at the dataset level via empirical CDFs (more commonly used version of confidence-threshold plots). We illustrate that the slight under-confidence of FALCON has little practical relevance, whereas the high levels of overconfidence (Fig. 1, 2, 4) displayed by all other methods can be highly problematic and misleading in practice: an underconfident prediction of 89% rather than well-calibrated 95% is unlikely to affect a decision making process, whereas a strong overconfidence of 80% rather than well-calibrated 10% can be detrimental in many real world application (Appendix A.6, Fig. S4, S5, A.7 Fig. S8). We report that FALCON is consistently less underconfident than EDL in terms of ECE and further illustrate this via empirical CDFs of confidence scores: While for FALCON 13.3% of predictions on the test set are made with a confidence of less than 90%, this increases to 30.5% for EDL, which is substantially higher than the 7.7% we observe for the well-calibrated Deep Ensembles (Fig. S4). In other words, only a minority of 5.6% of all predictions made by FALCON are made with a confidence of less than 90% and should have been made with a higher confidence. EDIT: New Figure S8 in the appendix illustrating the distribution of confidence scores also shows that EDL is substantially more underconfident than FALCON with a mean confidence score of 83.4% for EDL and 94.9% for FALCON.\"}", "{\"title\": \"Pont-by-point response to concerns regarding additional experiments (point 3)\", \"comment\": \"We have now extended our ablation study and report results on the relevance of both loss terms, demonstrating that the combination of both terms yields best results. In addition, we show that our method is robust to the choice of lambda_adv and lambda_s when varied over one order of magnitude (variation of micro-averaged ECE less than 0.04 when varying lambda_adv between 0.0005 and 0.03 and less than 0.06 when varying lambda_S between 10 and 100; Appendix A.4 - due to space constraints we could not move this to the main text).\"}", "{\"title\": \"Point-by-point response to concerns regarding ECE.\", \"comment\": \"We thank the reviewer for providing constructive comments and have attempted to address all issues in full. Below we provide a point-by-point response to all raised concerns. We are more than happy to provide further clarifications, should new questions arise from our response.\\n\\nUnknown domain drift and noise levels. We agree that in real-world applications the type of domain drift is not known/can be of various kinds - this is why we evaluate FALCON as well as various baselines on diverse types of domain drifts that were not seen during training, including different image perturbations but also character swaps in an NLP task; we argue that FALCON outperforming other methods on this diverse set of domain drifts is evidence that it is likely to also have a superior performance for other unknown types of domain drift. Since for practical applications neither the type of domain drift nor the amount of domain drift (noise level) is known in advance we also do not introduce a weighting of individual noise levels, but decided to use an unweighted average across all noise levels. \\n\\nCalibration for high levels of noise and truly OOD data. For very high levels of noise and in particular inputs that do not belong to any of the train classes, entropy can indeed be a complementary relevant metric and we now present additional evaluations demonstrating that FALCON also outperforms state-of-the-art methods in terms of entropy for high noise levels (Fig. S7 in Appendix A.7 shows FALCON outperforms baselines on MNIST, CIFAR and sequential MNIST; Fig. S8 shows the distribution of confidece scores in the OOD/high noise scenario and illustrates that FALCON has lower confidence scores than all baselines in the OOD case). However, we disagree that we wouldn't care about calibration for high noise levels: for very high noise levels where it is impossible to make a correct prediction, the accuracy is (close to) random and in this case perfect calibration means maximum entropy (and is therefore still meaningful and desirable). So we argue that it is actually because we care about calibration that we want to maximize entropy for high noise levels (at least in domain-drift scenarios where we assume that the distribution the inputs are drawn from changes, but the set of labels remains unchanged). The advantage of motivating maximum entropy via calibration is that we can use the same framework for the entire continuous spectrum of domain drift without the need of introducing an arbitrary threshold from which on we treat samples as OOD and maximize entropy. We now clarify that ECE is designed for such domain drift scenarios and discuss that in case of OOD samples with a different set of labels, entropy is a better metric, as suggested by the reviewer. \\n\\nRelation to ref [1]: Ovadia et al. evaluate - in an independent and concurrent submission - calibration under domain drift for a range of existing algorithms and datasets based on ECE, Brier score, and other metrics. The authors independently make the same argument as we do and argue that since the type of domain shift that will occur after training is unknown in practice, calibration should be quantified based on a range of different types of domain drift where the noise model can be controlled. They therefore generate shifted data for a range of perturbations and noise levels (16 distortions such as image blur at 5 noise levels) and compute ECE and Brier score across all noise levels and perturbations (section 4.2 in Ovadia et al.). It is worth noting that their findings confirm our results for baseline methods, in particular for deep ensembles and MC-dropout (they did not consider EDL and have not presented any new approach).\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper introduces a new loss function for training deep neural networks, which show good performance with respect to well-calibrated, trustworthy probabilities for samples after a domain shift. The authors conduct experiments with multiple datasets and multiple forms of perturbations, where the proposed method achieve superior performance.\", \"i_have_the_following_major_concerns_with_the_paper\": \"1. Presentation: The paper's presentation is very weak. It contains repetitive, long, convoluted statements and paragraphs all throughout making it difficult for the reader to understand anything. The first time I read this paper, I couldn't process what is happening. The introduction is more like realted work with less focus on what they are trying to pitch in the paper. However, the latter is the less severe concern.\\n\\n2. While I appreciate the use of reliability diagrams and the loss function inspired from it, I do not completely understand what are the entities in equations 1 and 2 or how they are used later. Please give them a name, and elaborate on them.\\n\\n3. While I was reading the paragraph before section 2.2 the first time, it seems to me that ECE is defined for those 10 pertubations specifically. I believe this is not the case; hence, the authors should make it more general and divert the specific details to experiments. \\n\\n4. I am not sure what the authors mean by, \\\"... while the loss surface remains largely unchanged\\\", in paragraph above section 2.2.2. I believe the authors have constructed a new loss function by adding a new term to it. That makes it difficult to understand the advantage the authors are talking about after this statement. Overall, I am not really convinced how in contrast to Bayesian Deep Learning, their approach can be EASILY applied to LSTM's and GRU's.\\n\\n5. What is L_{adv}? There is no equation number, no discussion around where this is defined. Again a presentation issue.\", \"minor\": \"Please clarify the reference for supplementary materials. For example, it is not clear that Table S1 is in supplementary material.\\n\\nOverall, this paper is good and has an interesting idea. The experiments are also extensive useful. However, I have reservations regarding the presentation of this paper at this moment. \\n\\n--- After Rebuttal ---\\n\\nI thank the authors for providing a detailed response to my questions and editing the paper. I am now more positive about the paper; however, I still feel the presentation of the paper could be further improved. At this point, I will not move to acceptance range and keep the same score.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper presents a method for calibrating neural networks on in- and out-of-distribution data using two additional loss terms: entropy-encouraging loss term which maximizes softmax probabilities for wrong classes and adversarial calibration loss term which pushes confidences to match accuracy on adversarial examples. The idea of using adversarial calibration training is interesting and promising, however, the clarity of the paper needs significant improvement and there are several issues which need to be addressed; for this reason, I recommend a weak reject for the current version.\\n\\n1. One of the contributions listed in the paper is that authors \\u201cillustrate the limitations of entropy as measure for trustworthy predictions and introduce a new metric to quantify technical trustworthiness based on the concept of calibration\\u201d; the section 2.1 discusses this in detail. The proposed metric is expected calibration error (ECE) averaged over different levels of noise perturbations applied to data during test time. This metric assumes that for a given dataset we know in advance what kind of noise perturbations can be implied to cause out-of-domain/domain-shift scenarios during test time. This assumption most likely doesn\\u2019t hold in real-world applications since domain shifts may be of various kinds. Moreover, calibration may not make sense at all for out-of-domain data if test inputs don\\u2019t belong to any of the train classes (as opposed to entropy of predictive distribution which still can be computed and is expected to be higher for such inputs). The metric is also dependent on the considered noise level range; the plot of ECE vs noise level can be illustrative and informative while the averaged value of ECE over all noise levels can be misleading (at least, we may want to have some discount factor to account for the fact that with noise level 0 we strongly care about calibration, while for very high noise levels calibration doesn\\u2019t give us much information since the useful patterns in data may be corrupted and it may be impossible even for humans to classify such objects, e.g. on Figure 1-top-left for noise level 100, the digit 6 is corrupted so much that we wouldn\\u2019t care about calibration for such inputs but just want to maximize the entropy).\\nThe paper also claims \\u201cRecent efforts in terms of evaluating predictive uncertainty have focused on entropy as measure for uncertainty-awareness for predictions under domain shift.\\u201d In previous work addressing uncertainty in domain shift [1], not only entropy but other various metrics are considered for out-of-domain detection (Brier score, thresholded confidences, etc), these metrics don\\u2019t depend on particular perturbations and are very informative.\\nThe introduced metric, ECE for different noise labels, makes sense only in the toy scenarios where we can control the noise level, however, it would still be better to either look at the plots or to use some discounted averaged ECE over different noise levels, but not equal average. So the significance of this contribution (introduction of new metric) is limited. \\n\\n2. The clarity of the paper could be significantly improved.\\n\\n(a) Figure 1: the top left plot suggests that for high noise level \\u201cwrong predictions are often made with high confidence\\u201d (so the entropy of this distribution is low) while the top right plot shows that on average predictive entropy gradually grows and is high for high noise levels, so the top right plot is probably not a representative example and it might be misleading to claim this is an \\u201coften\\u201d case.\\n\\n(b) Could you please clarify what you mean by \\u201cafter removing non-misleading evidence\\u201d in section 2.2.1? In the next sentence, the remaining probability is probably distribution uniformly across C-1, not C, classes. The predictive entropy loss term essentially maximizes the probabilities of wrong classes with a lower coefficient. How would you support the claim that the loss surface is unchanged?\\nThe loss term is \\u201cparameter-free\\u201d but still we need to tune a coefficient lambda_S for it.\\n\\n(c) The adversarial loss equation in section 2.2.2 is written as L2-norm of a scalar value, why is L2-norm needed? The acc(B_m) is not differentiable so it is probably just considered as constant in the paper? Please comment on that. \\n\\n(d) Section 3.2 and Figure 4. On Figure 4 Middle and Right plots have different colors than those listed on legend, please, fix this. On the middle plot, both FALCON and EDL have high ECE for noise levels <50 which indicates that they are both highly underconfident for in-distribution data and low noise levels, while other methods have close to 0 ECE on low noise levels. The authors only comment on EDL underconfidence: \\u201cit is worth noting that EDL has a substantially higher ECE for in-domain predictions, reflecting under-confident predictions on the test set\\u201d, and not on FALCON underconfidence. However, using the proposed score, ECE averaged over all noise levels (Table 2), it may look like the method is doing a good job, while underconfidence problem is revealed when looking at Figure 4. This is also an illustration of the concern about the proposed metric I raised in point 1.\\n\\n3. It would help to have an ablation study in the main text of the paper showing the significance of each loss term: in the appendix Figure S1, it is shown that adding adversarial loss to standard and entropy loss helps. Is the entropy loss needed at all? How does performance change if we only have standard and adversarial loss? How sensitive is performance to the choice of lambda_avd and lambda_s?\\n\\n\\n[1] Ovadia, Yaniv, et al. \\\"Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift.\\\" arXiv preprint arXiv:1906.02530 (2019).\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposed FALCON, a simple method to produce well-calibrated uncertainty estimation. The idea is to introduce two additional terms, one that directly encourage lower confidence for all negative classes of all data points, and another one that optimizes the ECE for adversarial samples. Experiments show that FALCON outperforms several state-of-the-art methods for calibrating neural network predictions.\\n\\nAlthough the first term, L_S, affects only negative predictions, it is still somewhat strange to uniformly operate on all data points. It would help to perform an ablation study to see how L_S affect the results.\\n\\nSome description in Section 2 could use more mathematical rigor (e.g., when describing L_{adv}).\\n\\nThe authors use EDL as a baseline, I was wondering why not use the more commonly used, and possibly more effective, temperature scaling (TS) method from Guo et al. 2017. Note that the EDL paper does not seem to explicitly compare EDL and TS. \\n\\nFigure 4 (middle and right) is confusing. The line style is not consistent in the figures.\\n\\nThere are a few places where the text is rather vague and confusing. For example, what do you mean by \\u2018non-misleading evidence\\u2019 when describing L_{adv}? It would also be better to provide more insight more L_S to help the readers out. For example, it would help to state that L_S operate only on negative predictions.\\n\\nMost baselines are rather simple non-probabilistic (non-Bayesian) methods. Besides, MNF, It would also be interesting to see how FALCON compare other probabilistic NN method such as natural parameter networks, where they also explicitly evaluated uncertainty estimation.\"}" ] }

[ "Chandra Bhagavatula", "Ronan Le Bras", "Chaitanya Malaviya", "Keisuke Sakaguchi", "Ari Holtzman", "Hannah Rashkin", "Doug Downey", "Wen-tau Yih", "Yejin Choi" ]

[ "Abductive Reasoning", "Commonsense Reasoning", "Natural Language Inference", "Natural Language Generation" ]

{ "note_id": [ "7I6ohZKYQl", "rkxSWudcor", "H1gU-PuqiS", "Sye4_Uu9or", "rylgzEOqsB", "ByltNYc0tr", "BklqO2_AKH", "BJlqUqW3tH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731582, 1573713916996, 1573713661631, 1573713515951, 1573712904455, 1571887409365, 1571880049579, 1571719761979 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1752/Authors" ], [ "ICLR.cc/2020/Conference/Paper1752/Authors" ], [ "ICLR.cc/2020/Conference/Paper1752/Authors" ], [ "ICLR.cc/2020/Conference/Paper1752/Authors" ], [ "ICLR.cc/2020/Conference/Paper1752/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1752/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1752/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper presents a dataset, created using a combination of existing resources, crowdsourcing, and model-based filtering, that aims to tests models' understanding of typical progressions of events in everyday situations. The dataset represents a challenge for a range of state of the art models for NLP and commonsense reasoning, and also can be used productively as a training task in transfer learning.\\n\\nAfter some discussion, reviewers came to a consensus that this represents an interesting contribution and a potentially valuable resource. There were some concerns\\u2014not fully resolved\\u2014about the implications of using model-based filtering during data creation, but these were not so serious as to invalidate the primary contributions of the paper.\\n\\nWhile the thematic fit with ICLR is a bit weak\\u2014the primary contribution of the paper appears to be a dataset and task definition, rather than anything specific to representation learning\\u2014there are relevant secondary contributions, and I think that this work will be practically of interest to a reasonable fraction of the ICLR audience.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Overall Comments (for all reviewers): Pre-training on alpha-NLI consistently improves performance on datasets that have small training sets\", \"comment\": \"We thank all reviewers for their overall positive comments about our work.\\n\\nBased on AnonReviewer2's suggestion, we performed experiments that show that pre-training on alpha-NLI consistently improves performance on related datasets that have relatively few training examples. \\n\\nIn particular, we compared the performance of 1) BERT fine-tuned on a given dataset and 2) BERT sequentially fine-tuned on alpha-NLI and then on the chosen dataset. \\n\\nDataset BERT-Ft(Dataset) BERT-Ft(alpha-NLI)\\u2192BERT-Ft(Dataset)\\n--------------------------------------------------------------------------------------------------------------------------------\\nWinoGrande [1] 65.8% 67.2%\\nWSC[2] 70.0% 74.0%\\nDPR[3] 72.5% 86.0%\\nHellaSwag[4] 46.7% 46.1%\\n\\nFor SNLI and HellaSwag, datasets with large amount of training data, there is no significant performance improvement upon pre-training on alpha-NLI. \\n\\n[1] Sakaguchi et al. WINOGRANDE: An Adversarial Winograd Schema Challenge at Scale\\n[2] Levesque et. al. The winograd schema challenge.\\n[3] Rahman and Ng. Resolving Complex Cases of Definite Pronouns: The Winograd Schema Challenge\\n[4] Zellers et. al. HellaSwag: Can a Machine Really Finish Your Sentence?\"}", "{\"title\": \"Thank you for the positive feedback!\", \"comment\": \"We appreciate AnonReviewer3 for encouraging comments about the importance of the proposed abductive inference and generation tasks and about the value of our proposed dataset. We address the main concerns individually below:\", \"adversarially_filtering_using_bert_and_gpt_gives_deep_learning_models_a_disadvantage\": \"While BERT originally achieved high performance on the originally collected dataset, several recent studies [1][2][3][4] have found the presence of annotation artifacts in crowdsourced data that inadvertently leak information about the target label. This subsequently leads to overestimation of the performance of AI systems on end tasks. Our adversarial filtering (AF) algorithm aims to address the problem of overestimation of performance. In spite of targeting GPT/BERT during AF, human performance on the AF resulting dataset is still high. The significant gap between human and BERT performance leaves scope for inventing new methods for abductive reasoning.\", \"ensemble_of_bert_models\": \"An ensemble of three BERT models achieves an accuracy of 68.9%, very close to a single model 68.6%.\", \"average_score_of_human\": \"The average score of human annotations is 89.4%. This is directly comparable with BERT-Ft [Fully Connected] model\\u2019s performance of 68.6% in Table 1.\\n\\n\\nRe. Ground Truth: \\nThe ground truth is assigned based on whether a hypothesis was collected during the plausible (Appendix A1 Task1) or implausible (Appendix A1 Task2) phase of the data collection procedure. To measure human performance, we had three annotators select the correct hypothesis and measured human performance as the accuracy of their majority-vote. Please let us know if this answers your question. If not, could you please clarify your question?\", \"generative_task_vs_classification\": \"We completely agree. While the generative task is more general and much more interesting, the challenge of evaluating generations is significant, particularly for this task. This is due to the fact that there could be multiple distinct plausible explanations for a given pair of hypothesis.\", \"consider_the_following_example\": \"\", \"o1\": \"Kelly and her friend wanted to take a train to the city.\", \"o2\": \"They had to wait for another one.\", \"plausible_explanations\": \"1. They read the timetable incorrectly and arrived at the station just after a train had left.\\n2. The train was full.\\n\\nBoth explanations are plausible, and explain the observations, but automated evaluation metrics are not reliable enough to capture this phenomenon based on their reliance on surface level similarities. To simultaneously make progress on the novel abductive reasoning task and due to the ease of evaluation, we additionally introduce a discriminative version of the task. Nonetheless, we agree that in its most general form, there could be any number of observations and models should be required to generate explanatory hypotheses in natural language (alpha-NLG task).\\n\\n\\n\\n[1] Gururangan et al. Annotation artifacts in natural language inference data.\\n[2] Poliak et al. Hypothesis only baselines in natural language inference.\\n[3] Tsuchiya et al. Performance impact caused by hidden bias of training data for recognizing textual entailment.\\n[4] Sakaguchi et al. WINOGRANDE: An Adversarial Winograd Schema Challenge at Scale\"}", "{\"title\": \"Pre-training on alpha-NLI consistently improves performance on datasets that have relatively small training sets\", \"comment\": \"We thank AnonReviewer2 for their appreciation of our work as an \\u201cimportant contribution\\u201d for commonsense reasoning and for their detailed review with several insightful suggestions. Based on these suggestions, we present new experimental results and address individual concerns below:\", \"clarifications_on_table_1_results\": \"\\u201cGPT AF\\u201d column in Table 1 shows the performance of various models on a dataset that is generated by using GPT as the adversary. Similarly, the \\u201cART\\u201d column uses BERT as the adversary. The aim of our adversarial filtering step is to arrive at a dataset that is hard for current state-of-the-art models. The table shows that using a strong model as the adversary results in a dataset that is hard for all other weaker models. Nonetheless, human performance remains high. \\nAt the end of our adversarial filtering algorithm, we have a large set of instances whose hypothesis choices are adversarially selected. This set of instances is randomly split into train/dev/test splits. While modifying only the train (or only the test) set is possible, the distribution of instances in the training and test splits will be distinct \\u2014 thus making the results hard to interpret. \\nWe will update the text to make this distinction between col 1 and col 2 of Table 1 clear.\", \"release_the_entire_set_of_hypotheses\": \"Yes, we will release all correct and incorrect hypotheses for the training instances, but the dev and test sets will only contain adversarially selected hypothesis pairs. Table 5 in the Appendix has the statistics of the ART dataset that we plan to make public with this work. \\n\\n\\n\\nTrain with multiple distractors, test with one distractor: \\nThank you for the suggestion and the pointer to the paper. We performed this experiment and found that training with multiple distractors results in a performance of 68.1%, while training with a single distractor achieves 68.6%. While the performance difference is not significant, interestingly, training with multiple distractors seems to be more stable to hyperparameter changes (e.g. batch size, no. of epochs, learning rate). We will include these findings in the final version of the paper.\", \"transfer_learning_to_other_tasks_or_language_model\": \"Thank you for another great suggestion. We have performed these experiments as well, and they show that pre-training on alpha-NLI consistently improves performance on datasets that have relatively few training examples. \\nIn particular, we compared the performance of 1) BERT fine-tuned on a given dataset and 2) BERT sequentially fine-tuned on alpha-NLI and then on the chosen dataset. For SNLI and HellaSwag, datasets with large amount of training data, there is no significant performance improvement due to pre-training on alpha-NLI.\\n\\n\\nDataset BERT-Ft(Dataset) BERT-Ft(alpha-NLI)\\u2192BERT-Ft(Dataset)\\n--------------------------------------------------------------------------------------------------------------------------------\\nWinoGrande [1] 65.8% 67.2%\\nWSC[2] 70.0% 74.0%\\nDPR[3] 72.5% 86.0%\\nHellaSwag[4] 46.7% 46.1%\", \"winograd_schema_challenge_and_nli_recast_as_abductive_reasoning\": \"While instances of WSC and NLI may require abductive reasoning, alpha-NLI is the first dataset to isolate this important, specific type of reasoning for future studies. Pre-training on alpha-NLI helps improve performance on WSC.\", \"comment_on_the_other_commonsense_categories\": \"We considered more categories (e.g., negation, physical, temporal commonsense) but the performance difference between humans and BERT-ft in those categories was not statistically significant. We will include these categories in the camera ready. Nonetheless, we\\u2019d like to emphasize that reliable and consistent categorization of instances into commonsense categories is a challenge. Often, solving an instance might require multiple different types of commonsense reasoning and the distinction of different types of commonsense is also nebulous.\", \"presentation_concerns\": \"Yes, AF meant adversarial filtering. \\nThe best performance of BERT-ft [Fully Connected] is 68.6%. \\nWe have updated the paper to show the plateauing effect more vividly (Please see Fig 6 in the updated paper). \\nWe have fixed the footnote in the updated PDF. \\nWe will fix the presentation issues in the final version. \\n\\n\\n[1] Sakaguchi et al. WINOGRANDE: An Adversarial Winograd Schema Challenge at Scale\\n[2] Levesque et. al. The winograd schema challenge.\\n[3] Rahman and Ng. Resolving Complex Cases of Definite Pronouns: The Winograd Schema Challenge\\n[4] Zellers et. al. HellaSwag: Can a Machine Really Finish Your Sentence?\"}", "{\"title\": \"Thank you for the positive feedback!\", \"comment\": \"We thank AnonReviewer1 for their positive comments about the interesting-ness of our proposed abductive reasoning tasks (inference and generation) and the associated benchmark dataset. We address specific concerns individually below:\", \"discussion_about_e_snli\": \"A key distinction between e-SNLI and Abductive-NLI is that the explanations in e-SNLI serve the purpose of justifying model decisions. In contrast, the goal of Abductive-NLI and Abductive-NLG is to select or generate explanatory hypotheses for given observations. Indeed, analogous to e-SNLI for SNLI, Abductive-NLI can be extended to \\u201ce-Abductive-NLI\\u201d by providing explanations that justify the selected hypothesis.\", \"consider_the_following_example_that_bert_fails_to_predict_correctly\": \"\", \"o1\": \"Chad loves Barry Bonds.\", \"h1\": \"Chad got to meet Barry Bonds online, chatting.\", \"h2\": \"Chad waited after a game and met Barry.\", \"o2\": \"Chad ensured that he took a picture to remember the event.\\n\\nThe e-Abductive-NLI task would require models to generate an explanation for selecting H2. For the above example, a possible explanation for selecting H2 could be: \\u201cPeople need to be physically co-located to take a picture with someone. Meeting online does not mean two people are physically co-located\\u201d.\\nWe think generating such justifications is a great next step and hope that our work will foster such interesting future research. \\n\\n\\n\\nRe. somewhat limited contribution: \\nWe appreciate the opportunity to briefly restate our contributions and to discuss its significance. \\nAbductive Commonsense Reasoning, a critical capability in human reasoning, is relatively less studied in NLP research. To support this line of research, our work introduces a dataset that focuses explicitly on this important reasoning capability. Furthermore, several recent works [1,2,3,4] have shown the presence of annotation artifacts in crowdsourced datasets -- which poses a significant challenge for dataset curation. Our work makes the following contributions: \\ni) proposes and formalizes two novel tasks of Abductive Inference and Abductive Generation, \\nii) presents a new dataset in support of these tasks collected through careful crowdsourcing design and an adversarial filtering algorithm, \\niii) establishes strong baselines on the task proving the difficulty of the tasks and\\niv) analyses the types of commonsense reasoning that current state of the art models fall short on.\\n\\n\\n\\nRe. limited form of Abductive Reasoning: \\nThe simplifying assumptions, mentioned in the paper, allow us to i) formulate the tasks concretely and ii) curate the dataset and evaluate models viably. We show that in spite of the assumptions, our dataset presents significant challenges for current models. We totally agree that in its most general form, there should be any number of observations and models should be required to generate explanatory hypotheses in natural language (as in the alpha-NLG task). We hope our work will lead to this future line of research.\\n\\n\\n\\nRe. the title: \\nThanks for the suggestion. We will update the title to reflect that this work is aimed at language-based abductive reasoning.\", \"table_7_vs_table2\": \"Thanks for catching that. We\\u2019ve updated the paper with the fix.\\n\\n\\n\\n[1] Gururangan et al. Annotation artifacts in natural language inference data.\\n[2] Poliak et al. Hypothesis only baselines in natural language inference.\\n[3] Tsuchiya e. al. Performance impact caused by hidden bias of training data for recognizing textual entailment.\\n[4] Sakaguchi et al. WINOGRANDE: An Adversarial Winograd Schema Challenge at Scale\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposes a new task/dataset for language-based abductive reasoning in narrative texts.\", \"pros\": [\"The proposed task is interesting and well motivated. The paper contributes a dataset (20,000 commonsense narratives and 200,000 explanatory hypotheses). The construction of the dataset was performed carefully (e.g., avoiding annotation artifacts).\", \"The paper established many reasonable baselines.\", \"The paper conducted detailed analysis, which invites more research on this task: despite the strong performance of many existing systems on NLI/RTE, there are larger gaps between the performance of these models and human performance on the proposed task. The experiments well support the conclusions made in the paper.\", \"The paper is well structured and easy to follow. It is well written.\", \"Cons/comments:\", \"While this is a new and interesting task, the contribution (as discussed above in \\u201cpros\\u201d above) is somewhat limited. I also suggest the paper discusses e-SNLI a bit more.\", \"The paper has a specific form of formulation for abductive reasoning, where there are exactly two observations and one proceeds the other; the explanation happens in between. I can see this helps collect and annotate data, but also limit the form of abductive reasoning and how models should be developed.\", \"Should the title of the paper specify the paper is about \\u201clanguage-based\\u201d abductive reasoning.\", \"A minor one: \\u201cTable 7 reports results on the \\u03b1NLI task.\\u201d Should it be \\u201cTable 2\\u201d?\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper introduces two new natural language tasks in the area of commonsense reasoning: natural language abductive inference and natural language abductive generation. The paper also introduces a new dataset, ART, to support training and evaluating models for the introduced tasks. The paper describes the new language abductive tasks, contrasting it to the related, and recently established, natural language inference (entailment) task. They go on to describe the construction of baseline models for these tasks. These models were primarily constructed to diagnose potential unwanted biases in the dataset (e.g., are the tasks partially solved by looking at parts of the input, do existing NLI models far well on the dataset, etc.), demonstrating a significant gap with respect to human performance.\\n\\nThe paper, and the dataset specifically, represent an important contribution to the area of natural language commonsense reasoning. It convincingly demonstrates that the proposed tasks, while highly related to natural language entailment, are not trivially addressed by existing state-of-the-art models. I expect that teaching models to perform well on this task can lead to improvements in other tasks, although empirical evidence of this hypothesis is currently absent from the paper.\\n\\nBelow are a set of more specific observations about the paper. Some of these comments aim to improve the presentation or content of the paper.\\n\\n1. In Section \\u201c5.1 - Pre-trained Language Models\\u201d and attendant Table 1describe results of different baselines on the ART inference task. The results in the table confused me for quite some time, I\\u2019d appreciate some clarifications. With respect to the differences with columns 1 (GPT AF) and 2 (ART, also BERT AF) I would like the comparison to be made more clear. As far as I understand it, there are 2 parts of the dataset that can be varied: (1) the train+dev sets and (2) the test set. Furthermore, it seems that it makes sense to vary each of these at a time, if we are to compare results with variants. For example: we can fix the test set, and vary how we generate training and dev examples. If a model does better with the same test set, we can assume the train+dev examples were better for the model (for whatever reasons, closer distribution to test, harder or more \\ninformative training examples, etc). We can also keep the train+set constant, and vary the test set. This allows us to evaluate which test set is harder with respect to the training examples. The caption of Table 1 implies that both columns are evaluations based on the \\u201cART\\u201d test set. If that is correct, then the train+dev set generated from the GPT adversarial examples is of better quality, generating a BERT-ft (fully connected) model that is 3% better. But the overall analysis seems to indicate that this is not what was done in the experiment. Rather, it seems that *both* the train+dev _and_ the test sets were modified concurrently. If that is the case, I would emphasize that the text needs to make this distinction clear. Furthermore, I would say that varying both train and test sets concurrently is sub-optimal, and makes it a bit harder to draw the conclusion that BERT adversarial filtering leads to a stronger overall dataset.\\n\\n2. Along the lines of the argument in (1), above, I would urge the authors to publish the *entire* set of generated hypotheses (plausible and implausible) instead of only releasing the adversarially filtered pairs. Our group\\u2019s experience with training inference models is that it is often beneficial to train using \\u201ceasy\\u201d examples, not only hard examples. I suspect the adversarially filtered set will focus on hard examples only. While this is fine to do in the test set, I think if the full set of annotated/generated hypotheses are released, model designers can experiment with combining pairs of hypothesis in different ways.\\n\\n3. Furthering the argument of (2): in certain few-shot classification tasks, one is typically asked to identify similarity between one test example and different class representatives. Experience shows that it is often beneficial to train the model on a larger number of distractor classes than what the model is eventually evaluated on (e.g., https://papers.nips.cc/paper/6996-prototypical-networks-for-few-shot-learning). In the alpha-NLI setting, have you experimented with training using multiple distractors, instead of only 1, during training (even if you end up evaluating over 2 hypotheses)?\\n\\n4. One potential argument for introducing a new natural language task is of transfer learning: learning to perform a complex natural language task should lead to better natural language models more generally, or, for some other related tasks. This paper does not really touch on this aspect of the work. But, potentially, one investigation that could be conducted is through a reversal of the paper\\u2019s existing NLI entailment experiment. The paper shows that NLI entailment models do not perform well on alpha-NLI. But it would be interesting to see if a model trained on alpha-NLI, and fine-tuned or multi-tasked on NLI entailment, does better on NLI entailment (i.e., is there transfer from alpha-NLI to entailment NLI?).\\n\\n5. Another option is to evaluate whether alpha-NLI helps with other commonsense tasks. One other example is Winograd Schema Challenge, which current systems also perform well below human performance. It also seems that the Winograd Schema Challenge questions are not too far from abductive inference.\\n\\n6. In the abstract of the paper, before the paper defines the abductive inference/generation task specifically, the claim that abductive reasoning has had much attention in research seemed awkward. Informally, most commonsense reasoning (including NLI entailment) could be cast as abductive reasoning.\\n\\n7. In at least one occasion, I found an acronym which was hard to find the definition for (\\u201cAF\\u201d used in Section \\u201c5.1 - Pre-trained Language Models\\u201d; I assumed it was \\u201cadversarial filtering\\u201d.)\\n\\n8. In Section \\u201c5.1 - Pre-trained Language Models\\u201d it seems that the text quotes an accuracy for BERT-ft (fully connected) of 68.9%, but Table 1 indicates 69.6%.\\n\\n9. In Section \\u201c5.1 - Learning Curve and Dataset Size\\u201d, there is a claim that the performance of the model plateaus at ~10,000 instances. This does not seem supported by Figure 5. There appears to be over 5-7% accuracy (absolute) improvements from 10k to 100k examples. Maybe the graph needs to be enhanced for legibility?\\n\\n10. It is great that the paper includes human numbers for both tasks, including all the metrics for generation.\\n\\n11. Period missing in footnote 8.\\n\\n12. The analysis section is interesting, it is useful to have in the paper. However, Table 3 is a bit disappointing in that ~26% of the sampled examples fit into one of the categories. It would be great if the authors could comment on the remaining ~74% of the sampled dataset.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary: the paper purposes a dataset of abductive language inference and generation. The dataset is generated by human, while the testing set is adversarially selected using BERT. The paper experiments the popular deep learning models on the dataset and observe shortcoming of deep learning on this task.\", \"comments\": \"overall, the problem on abductive inference and abductive generation in language in very interesting and important. This dataset seems valuable. And the paper is simple and well-written.\", \"concerns\": \"I find the claim on deep networks kind of irresponsible.\\n1. The dataset is adversarially filtered using BERT and GPT, which gives deep learning model a huge disadvantage. After all, the paper says BERT scores 88% before the dataset is attacked. \\n2. The human score of 91.4% is based on majority vote, which should be compared with an ensemble of deep learning prediction. To compare the author should use the average score of human.\\n3. The ground truth is selected by human.\\n\\nOn a high level, the main difficulty of abduction is to search in the exponentially large space of hypothesis. Formulating the abduction task as a (binary) classification problem is less interesting. The generative task is a better option.\", \"decision\": \"despite the seeming unfair comparison, this task is novel. I vote for weak accept.\"}" ] }

[ "Melih Elibol", "Lihua Lei", "Michael I. Jordan" ]

[ "optimization", "variance reduction", "machine learning", "deep neural networks" ]

{ "note_id": [ "RMDYTxVXKY", "Bylt-T73oH", "ByeDcCHjiS", "Hyexcsydor", "H1eEA5y_sr", "B1ldttyujr", "S1xl3O1uiB", "S1gBwSJRtB", "S1g8clOTFH", "rJx0y3HZYS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731552, 1573825793477, 1573768846701, 1573546888070, 1573546699994, 1573546367683, 1573546151916, 1571841372842, 1571811470288, 1571015653528 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1751/Authors" ], [ "ICLR.cc/2020/Conference/Paper1751/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1751/Authors" ], [ "ICLR.cc/2020/Conference/Paper1751/Authors" ], [ "ICLR.cc/2020/Conference/Paper1751/Authors" ], [ "ICLR.cc/2020/Conference/Paper1751/Authors" ], [ "ICLR.cc/2020/Conference/Paper1751/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1751/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1751/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"Congratulations on getting your paper accepted to ICLR. Please make sure to incorporate the reviewers' suggestions for the final version.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Summary of Updates to Second Revision\", \"comment\": \"Our most recent submission includes the following notable changes:\\n\\n1. We have added a natural language processing experiment in response to Reviewer #1. This is a generative LSTM language model which involved additional algorithmic design decisions to deal with intermediate states at different time scales. The model is trained on the Penn Treebank corpus. This is compared to an SGD baseline with constant learning rate.\\n\\n2. We have added a sparse matrix factorization experiment in response to Reviewer #2. This is a Bayesian Personalized Ranking model train on the 100k MovieLens dataset. We implement a learning rate schedule for our algorithm, which we consider a fair analogue to an exponential decay learning rate schedule for SGD. The SGD baseline for this experiment uses an exponential decay learning rate schedule.\\n\\n3. We have updated our related work, and added a subsection which more clearly defines our contributions.\\n\\nWe were planning to include an experiment to demonstrate an implementation of sparse back propagation for wall-clock time comparison in response to Reviewer #3, but we ran out of time.\\n\\nWe'd like to thank the reviewers, area chairs, and anyone else involved in the decision making process for their time and consideration.\"}", "{\"title\": \"Response\", \"comment\": [\"I do agree that this paper has some theoretical contributions related to sparsity for general nonconvex problems.\", \"Please add experiments on general non-convex example to show the advantages of your proposed methods with others (not just only with SpiderBoost -- I believe that in term of numerical experiments, SpiderBoost is not the best one).\", \"Please revise your related work properly.\", \"Please highlight your contributions in the introduction part.\", \"I will update my score accordingly.\"]}", "{\"title\": \"Reply to Reviewer #3\", \"comment\": [\"Thank you for your thoughtful feedback. Please find below our responses to your comments.\", \"queries/N and wallclock time: We are currently looking into adding additional experiments to address this question.\", \"Confusion about rtop operator: We apologize for the typos here and thank you for providing a more intuitive explanation. We updated the paper to address confusion about the definitions of $k$, $k_1$, and $k_2$. We also updated definitions and examples for corrections and enhanced clarity. We added a more intuitive explanation of what the rtop operator does before formally introducing the operator.\", \"Linearity of rtop Operator: By linearity we mean that rtop(x, y+y') = rtop(x, y) + rtop(x, y') for a fixed S and hence E[rtop(x, y+y')] = E[rtop(x, y)]+E[rtop(x,y')]. We added a sentence clarifying this.\", \"Theoretical Motivation of Hyperparameters: Based on Theorems 1 and 2, Our experiments are carried out by specifying the fraction $k/d$ of gradient coordinates we want used for variance reduction. We then set set $k_1 = k // 2$, where $//$ is integer division, and $k_2 = k - k_1$. In our experiments, we set $k/d = 0.1$. We set $B = c b$ for a small constant $c$ (in our experiments we use $c=10$). We then set $m = B/b$.\", \"Impact of non-uniform minibatch sampling: Non-uniform minibatch sampling has no negative effect on sparsity because it only changes the probability that each entry is sampled and not the sparsity pattern.\"]}", "{\"title\": \"Reply to Review #2\", \"comment\": \"Thank you for your feedback. Indeed, some implementations of SVRG-style variance reduction methods are shown to be ineffective for training deep neural networks (Defazio et al. 2018). We have reproduced these results and acknowledge these findings in our paper with the appropriate citations. The covariance between the gradient snapshot computed every $m$ iterations and the gradient computed in the inner loop decreases rapidly for very large deep learning models. Our algorithm reduces the cost of computing variance reduction terms, which reduces the number of inner loop iterations required to see the benefits of the variance reducing computations. Furthermore, by using sparsity, our algorithm naturally reduces the overall potential variance of these methods. This potential variance occurs when the aforementioned covariance approaches zero. Our algorithm is only beneficial when there is sparsity structure in the gradients of the objective, and we experimentally show that such sparsity structure exists in the gradients of some deep neural networks.\\n\\nWe view this work as taking a step toward making SVRG/SCSG-style variance reduction methods practical. These methods provide a unique way to estimate gradient sparsity by taking advantage of the full/large-batch gradient computed at the beginning of each outer-loop. In other words, the correction term in SVRG/SCSG-style algorithms can be used for two purposes at the same time without extra overhead: Variance reduction and estimation of gradient sparsity. By contrast, this cannot be done with SGD-style algorithms without introducing significant computational overhead.\\n\\nWhile this work can be viewed as addressing some of the observations made by Defazio et al., we plan to directly address them in future work. Lei et al. (2017) comment that the mechanism used by SVRG/SCSG-style variance reduction methods to accelerate gradient-based methods is qualitatively different than momentum (in SGD-momentum/Adam) and adaptive stepsizes (in AdaGrad/Adam). We believe that, if properly combined with existing state-of-the-art techniques, SVRG/SCSG-style variance reduction methods have the potential to further reduce stochastic variance.\\n\\n- Experiments: To concretely address your feedback about our experiments, we will add at least one additional non-convex problem which exhibits gradient sparsity.\\n\\n- Figures in Log-scale: Thank you for the suggestion. We updated our figures to log-scale. \\n\\n- Connection With SARAH: Thank you for pointing out our missing references. We added them into our references and discussed them in Section 2.\"}", "{\"title\": \"Reply to Review #1\", \"comment\": [\"Thank you for your thoughtful feedback. Please find below our responses to your comments.\", \"Typos in Section 2.1: We apologized for the typos. In the new draft, we updated definitions and examples for corrections and enhanced clarity.\", \"Section 2.2 Definition of Activation Functions: This is now simplified to use only $\\\\phi$ instead of introducing $g$.\", \"A1: what is $\\\\overline{x}_0$ in $\\\\Delta f$: $\\\\tilde{x}_{0}$ should be $x_{0}$. We correct it in the updated draft.\", \"We are currently looking into adding additional experiments to address your suggestion to broaden the variety of experiments.\"]}", "{\"title\": \"Thank You\", \"comment\": \"We thank all reviewers for their thoughtful comments. We also thank them for pointing out typos and potential issues on clarity. We will revise our paper twice before discussions end. The first revision, which we have already submitted, addresses all the issues noted in our response to each reviewer. Our second and final revision adds experiments based on the suggestions of each reviewer.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary: This paper introduces a sparse variant to SpiderBoost which reduces the complexity cost of updating gradient estimates by way of sparse updates. The authors prove that this variant incurs a negligible increase in worst case complexities as soon as certain assumptions are satisfied, and that when their algorithm captures sparsity correctly, they improve upon SpiderBoost's complexity.\\n\\nThis paper is clearly, and the experiments support the theoretical contributions. \\n\\nIn Figure 1, you report results as a function of gradient queries/N. Given Theorem 2, I assume that the graphs would look similar as a function of wall-clock time; can you confirm this?\", \"recommendation\": \"Accept.\", \"minor_comments_and_questions_for_the_author\": [\"I am slightly confused by the introduction of the rtop operator. Specifically,\", \"1) What is the relation between k1, k2, and k?\", \"2) You write that S is a random subset of size k. Should this be k2?\", \"3) In your first example, should we have rtop(x,y) = (0, 16, 0, 0, 1), since for \\\\ell = 2, y_\\\\ell = 4, d-k1 = 4, k2=1? Am I missing something?\", \"More generally, my understanding is that the rtop(x,y) operator randomly sparsifies y based on x, which essentially provides indication of where sparsity would be least harmful; when not sparsifying, rtop applies a rescaling that guarantees unbiased estimates. If this is correct, I would recommend making that intuition more clear early on in the paper, in order to improve upon the clarity of the paper.\", \"You state that rtop is linear in y; since rtop depends on the random variable S, is the claim that E[rtop(x, y+y')] = E[rtop(x, y)]+E[rtop(x,y')] (which follows from unbiasedness)?\", \"For your experiments, could you discuss how your choice of hyperparameters relates to the constraints in Theorem 1 and 2?\", \"I believe Table 1 would be more impactful if it also included the initial entropy ratios at the beginning of training, rather than reporting those values below.\", \"Other variance reduction techniques for minibatching focus on choosing the minibatches themselves with non-uniform sampling. Under such a sampling mechanism, do you foresee any complications to using SpiderBoost with Sparse Gradients, eg., decrease in overall sparsity?\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper aims at improving the computational cost of variance reduction methods while preserving their benefits regarding the fast provable convergence. The existing variance reduction based methods suffer from higher per-iteration gradient query complexity as compared to the vanilla mini-batch SGD, which limits their utility in many practical settings. This paper notices that, for many models, as the training progresses the gradient vectors start exhibiting structure in the sense that only a small number of coordinates have large magnitude. Based on this observation, the paper proposes a modified variance reduction method (by modifying the SpiderBoost method), where a 'memory vector' keeps track of the coordinates of the gradient vectors with large variance. Let $d$ be the size of the model parameter. During each iteration, one computes the gradient for $k_1$ coordinates with the highest variance (according to the memory vector) and an additional $k_2$ random coordinates.\\n\\nThe paper shows that in the worst case, the proposed method has the same gradient query complexity as the SpiderBoost variance reduction method. Assuming that the proposed method can track the sparsity of the gradient vector, the proposed method achieves a gradient query complexity which is $O(\\\\sqrt{(k_1 + k_2)/d})$ times that of the SpiderBoost method. The paper demonstrates the gradient query complexity improvement over the SpiderBoost method on MNIST and CIFAR-10 data set.\\n\\nThe paper presents novel results by utilizing the ideas from the field of communication-efficient distributed optimization. As far as the reviewer can tell, the results in the paper are correct. That said, there is quite a bit of room for improvement in terms of the writing of the paper. \\n\\nThe paper appears to have way too many typos. For example, \\n\\nIn Section 2.1:\\n- Why is $k$ introduced?\\n- $S$ denotes a random subset with size $k$ ---> $k_2$?\\n- drawn from the set ${\\\\ell : |y_{\\\\ell}| < |y_{(k)}|}$ ----> ${\\\\ell : |x_{\\\\ell}| < |x_{(k_2)}|}$?\\n- $rtop(x, y) = (0, 12, 0, 0, 1)$ --> $rtop(x, y) = (0, 16, 0, 0, 1)$\", \"in_lemma_1\": [\"while defining $top_{-k_1}(x, y)$, \\\".... if |x_{\\\\ell}| >= |x_{(k_1)}|\\\" ----> \\\".... if |x_{\\\\ell}| <= |x_{(k_1)}|\\\"?\", \"In Section 2.2:\", \"What are $g_0, g_1,..., g_{L-1}$? Shouldn't these be $\\\\phi_0, \\\\phi_1,..., \\\\phi_{L-1}$?\"], \"in_a1\": \"- right after (5), what is $\\\\tilde{x}_0$ in the definition of $\\\\Delta_f$?\\n\\nThe authors may also consider making the empirical evaluation more comprehensive by considering tasks from the NLP domain, e.g., language modeling. This would further help asses the utility of the proposed method.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary:\\nThe author(s) provide a method which combines some property of SCGS method and SpiderBoost. Theoretical results are provided and achieve the state-of-the-art complexity, which match the one of SpiderBoost. Numerical experiments show some advantage compared to SpiderBoost on some deep neural network architecture for some standard datasets MNIST, SVHN, and CIFAR-10.\", \"comments\": \"1) It is true that variance reduction methods achieve the state-of-the-art complexity theory for finding first order stationary point of general nonconvex optimization problems. However, it is well-known that variance reduction methods are not very efficient for training deep neural networks. All of the experiments in this paper are focusing on deep learning problems. If the author(s) would like to show good performance, I would suggest to compare the algorithms with the state-of-the-art algorithms in Deep Learning such as Adam, SGD-Momentum. Showing some improvement over SpiderBoost for deep learning problems would have low impact. \\n\\n2) I would suggest the author(s) to switch directions to focus on general nonconvex problems, that is, to find some different examples on general non-convex optimization problems rather than for deep learning problems. In other words, to find examples which show that your algorithm has more advantage than SGD-type algorithms, SVRG-type. \\n\\n3) I would also suggest the author(s) to plot all figures in log-scale in order to see in more detail performance. \\n\\n4) According to my knowledge, SpiderBoost is an alternative way of re-writing the SARAH algorithm [1, 2] with some small modification, that is a variant of SARAH. Therefore, the SARAH algorithm should be highly related to this paper and need to be discussed and mentioned more clearly. \\n\\n[1] Nguyen et al 2017a, \\u201cSARAH: A Novel Method for Machine Learning Problems Using Stochastic Recursive Gradient\\u201d.\\n[2] Nguyen et al 2017b, \\u201cStochastic Recursive Gradient Algorithm for Nonconvex Optimization\\u201d.\"}" ] }

[ "Jack Turner", "Elliot J. Crowley", "Michael O'Boyle", "Amos Storkey", "Gavin Gray" ]

[ "model compression", "architecture search", "efficiency", "budget", "convolutional neural networks" ]

{ "note_id": [ "tumHqHLIlx", "BJxriKiZiH", "Syghz8ibsr", "S1ljr-j-ir", "rkl7kPYxcS", "rkxUDhtpFB", "ByxSoEyLKr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731524, 1573136797315, 1573135891692, 1573134659465, 1572013786530, 1571818590320, 1571316892672 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1750/Authors" ], [ "ICLR.cc/2020/Conference/Paper1750/Authors" ], [ "ICLR.cc/2020/Conference/Paper1750/Authors" ], [ "ICLR.cc/2020/Conference/Paper1750/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1750/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1750/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"Two reviewers recommend acceptance. One reviewer is negative, however, does not provide reasons for rejection. The AC read the paper and agrees with the positive reviewers. in that the paper provides value for the community on an important topic of network compression.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"We would first like to thank the reviewer for their detailed comments and analysis of our work. We are glad they liked the idea, and we are particularly happy that they appreciated the simplicity of the method, as we believe that complicated methods can present a major obstacle to deployment.\", \"we_answer_the_queries_below\": [\"The sampling under budget constraints is done using rejection sampling: we generate architecture proposals at random and only save those that satisfy our parameter budget. Since we are only generating proposals there is no need to create or instantiate any networks, we are able to infer the parameter budget directly from the configuration, making this a very cheap operation. We will make this clear in the paper. This could of course be improved in future, perhaps using a constraint solver to generate possible networks, or learning a distribution over block types.\", \"The suggestion that student networks be allowed to vary in number of blocks is a very interesting one, and not one that we had considered. We worry that comparing Fisher potential across different depth networks may raise complications. However, one means to achieve this might be to introduce \\u201cIdentity\\u201d as a block choice, which would allow for depth shrinking/expansion.\", \"Regarding the use of Fisher information for general NAS: this also is an interesting suggestion, and something we plan to investigate in future work.\", \"Thank you for pointing out the related work by Chen et al. Their statement of budget constraints as a problem of Bayesian optimisation is very interesting. We will add this to our related work and provide some discussion.\", \"On the reduced correlation of Fisher potential with final performance after training with several minibatches of data: At the moment we can only speculate on this. At initialisation, with randomly sampled weights, the activations may be more conditionally independent than later in training, when the weights are being learnt. The use of the Fisher potential implicitly assumes conditional independence as a measure of the information a block has about the label.\", \"We can also confirm that each baseline network is trained with the same Attention Transfer mechanism as the BlockSwap ones, or with a \\u201cborn again\\u201d strategy [1] in the case of CondenseNet. We will amend the figure caption to make this clear.\", \"The budgets we tested with were chosen to be log-linearly spread based on the parameter count of the original model (WRN-40-2). We then also chose SNIP budgets to match.\", \"Regarding sparsity and non-zero parameter counts; we were specifically comparing with unstructured sparsity approaches, due to its relationship with SNIP. We will make this more explicit in the paper.\", \"We hope that this answers all of the questions the reviewer raised and again thank them for such constructive feedback.\", \"[1] Furlanello et al. Born Again Neural Networks, ICML 2018\"]}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We would like to thank the reviewer for their questions, and we are very happy to answer them. We hope this reiterates the important point of this paper: it shows how good mixed blocktypes can be found and extensively demonstrates their superior performance.\\n\\n1. Attention transfer is the term for the widely-used distillation technique we employ. We cite the original paper [1] in the introduction on the first use of the term so hopefully readers who are unfamiliar with the method can follow up on it. We provide a formal and informal definition in Section 3.2 for completeness. We will explicitly add that attention transfer is a distillation method in the introduction. \\n\\n2. The blocks chosen are a representative subset of those that have previously been demonstrated to provide substantial compressive advantage. The Standard (S) block is used in most residual networks and the remaining blocks are cheapened versions of S. The Bottleneck (B) block is very popular in compact network designs. Grouped convolutions (as seen in G) are prevalent in many mobile-sized networks (e.g. MobileNet, CondenseNet), and the BG block is simply a combination of B and G. These blocks are not highly-engineered, so we can demonstrate that the success of our method is due to their combination, rather than their raw representational capacity. The fact we can make gains from using this small selection of simple blocks is itself enlightening.\\n\\n3. Thank you for checking our numbers regarding the convolutions in the S block. In this instance, we are correct in the paper: if the number of input channels to a convolution is $N$, and the number of output channels is also $N$, then each filter of kernel size $k\\\\times k$ has depth N. There are N of these $N\\\\times k\\\\times k$ filters, each one producing one of the N output channels.\\n\\n4. $N_L$ can be read as the number of layers in the network. L is not itself an object, just a notation. We will change the notation and writing here to make things very clear to prevent others from being distracted by this.\\n\\n5. We mention in Section 3.2 that the attention transfer mapping f() is taken directly from [1]. It is critical that we use the common default attention transfer mapping so any demonstrated benefit is not arguably down to some selection of f(), but due to the BlockSwap procedure.\\n\\n6. Fisher potential is theoretically motivated: the Fisher potential is the trace of the Fisher Information matrix of the activations. In each block the Fisher potential is (locally) the total information the block contains about the class label (under a simplifying conditional independence assumption). During training it is this information about the class that drives learning, and the initial learning steps are key. Hence higher information values tend to result in higher efficiency block utilisation. It is interesting to note that at initialisation, with randomly sampled weights, the activations may (speculatively) be more conditionally independent than later in training, and this may explain the benefit of doing this computation on the first minibatch. In future we will explore this further. In addition to the theoretical motivation, we empirically show Fisher potential to be better than common, less well-founded alternatives in Table 1. \\n\\n7. We chose our training hyperparameters to closely mirror those detailed in the original WideResNet paper [2] for CIFAR to ensure the empirical results focus on the BlockSwap procedure. All the design hyperparameters for BlockSwap (e.g. number of samples taken) are justified in Section 4.\\n\\n8. The major novelty of this paper is a very efficient method that determines which of a large number of mixed blocktype networks---in which the blocks considered are cheaper versions of the original S block---is a powerful network. It is more important to cheapen the structure in some blocks of the network than others. This relative importance is found via search, rejecting structures with the wrong balance using the Fisher potential.\\n\\n9. In fact, the primary contribution of this paper is precisely that it introduces an algorithm that finds a good mixed blocktype. This is done through search across randomly generated blocks that satisfy a budget for a mixed block network that optimizes a very cheap Fisher potential calculation. Because the Fisher potential is cheap to calculate (it does not involve learning) and correlates very well with learnt network performance, a large number of random networks can be tested, and hence we can output very high performing networks. We provide substantial empirical evidence that our method for finding such networks outperforms randomly assembling them. We also provide analyses of exactly what makes a good architecture in Appendix B. \\n\\n[1] Zagoruyko and Komodakis, Paying More Attention to Attention: Improving the Performance of Convolutional Neural networks via Attention Transfer. ICLR, 2017\\n\\n[2] Zagoruyko and Komodakis, Wide residual networks, BMVC 2016\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We would like to thank the reviewer for their positive review. We are delighted that they found the results remarkable, and the experiments insightful and thorough.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper studies a fast algorithm for choosing networks with interleaved block types \\u2013 BlockSwap. It passes a single minibatch of training data through randomly initialized networks and gauging their Fisher potential.\\nThe teacher-student network is used here to learn compressed network on a budget. They conduct various experiments, including cifar-10, imagenet, and coco-detection. All the experiments show the advance of proposed model, which is quite remarkable.\\nOverall, this paper is well organized, and very well written. The insightful experiments thoroughly discuss and compare the proposed method.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"1) in the introduction, it's not clear what attention transfer means\\n 2) why did they specifically choose these four blocks? \\n 3) (3.1) substitute blocks $S$: ``each using $N$ lots of $N \\\\times k \\\\times k$ filters\\\" should it be ``$N$ number of $k \\\\times k$ filters\\\"? \\n 4) (3.2) ``consider a choice of layers $i = 1, 2, \\\\dots, N_L$\\\" what is $L$ and why does the number of layers depend on $L$? \\n 5) (3.2) Why is that specific $f$ chosen? \\n 6) (3.3) They could elaborate more on why they chose Fisher potential instead of other metrics for architecture selection. Currently, they only provided the intuition of the metric. Since the paper suggests the Fisher potential is a crucial part of their method, they could provide more theoretical justification about this choice. \\n 7) (4.0) How did they decide on the hyperparameters? \\n 8) In the introduction, they suggest that the major novelty of their method ``assigns non-uniform importance to each block by cheapening them to different extents\\\", but in their method they only randomly assembled the mixed-blocktype models. \\n 9) Section 4.1 suggests good mixed blocktype yields low error but they didn't address how good mixed blocktype can be found\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper introduces an approach to compressing a deep network so as to satisfy a given budget. In contrast to methods that replace all convolutional blocks with the same cheaper alternative, the authors argue that mixing different types of such cheap alternatives should be effective. To achieve this in a fast manner, they propose to randomly sample architectures with mixed blocktypes that satisfy the given budget and rank these architectures using Fisher information.\", \"originality\": [\"The approach is simple, but seems effective. Although the different components that make up the final method are known, they are put together so as to address a different task, in a way that I find interesting.\"], \"methodology\": [\"From the paper, it is not entirely clear how sampling under budget constraints is achieved. I imagine that one could do it in a naive way, by choosing the first block uniformly randomly, and then removing the candidates that do not fit the budget for the subsequent blocks. This, however, would seem to give a different importance to the early blocks than to the later ones, since the former would essentially not have any budget constraints. I would appreciate it is the authors could explain their strategy in more detail.\", \"The assumption that the number of blocks in the student is the same as in the teacher seems constraining. A better architecture might be obtained by having fewer, wider blocks, or more, thinner blocks. Do the authors see a potential way to address this limitation?\", \"In principle, it seems that the proposed Fisher Information-based strategy could be used for general NAS, not just for compact architectures. Have the authors investigated this direction?\"], \"related_work\": [\"It seems to me that the literature review on compression/pruning is a bit shallow. I acknowledge, however, that most works do not tackle the scenario where a budget is given. However, Chen et al., \\\"Constraint-aware Deep Neural Network Compression\\\", ECCV 2018, do, and it would be interesting to discuss and provide comparisons with this work.\"], \"experiments\": [\"I appreciate the ablation study, which answered several of my questions.\", \"In Table 1, it seems counterintuitive that the (negative) correlation becomes smaller as the number of mini-batches increases. Do the authors have an explanation for this?\", \"In Section 5, are the baseline compact networks all trained using the same Attention Transfer algorithm as for the BlockSwap ones?\", \"In Table 2, the budget values (P. (K)) seem fairly arbitrary? How were they obtained? They seem to match those of SNIP. Is this because the authors ran SNIP, and then set their budget accordingly?\", \"Below Fig. 3, the authors mention that they are generous with sparsity-based methods because they count the number of non-zero parameters. Note that several structured-sparsity compression methods have been proposed (Alvarez & Salzmann, NIPS 2016, 2017; Wen et al., NIPS 2016), which, by contrast with regular sparsity ones, would cancel out entire filters.\"], \"summary\": \"Overall, I like the simple idea proposed in this paper. I would however appreciate it if the authors could clarify the way they sample the architectures and address my questions about their experiments.\"}" ] }

[ "Xinshi Chen", "Yu Li", "Ramzan Umarov", "Xin Gao", "Le Song" ]

[ "RNA secondary structure prediction", "learning algorithm", "deep architecture design", "computational biology" ]

{ "note_id": [ "ts9yuuqh4", "B1g-0PdH3r", "B1lJZQthiB", "r1eofYOniB", "BkxrY85dsH", "rJge8iyuoB", "HJeLA51_iS", "HJgMRKJuiB", "S1eeyK1_iH", "SyxL1Okuor", "HJxLplGwor", "HJelWTWPsB", "H1gcAa_yoS", "r1e3VSpiqr", "Hyx-fHaiqS", "B1xm6dZqqS", "S1xh9zkV9H", "SJlmuE8k5S", "Bklpej_AKH" ], "note_type": [ "decision", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_comment", "official_comment", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731493, 1574434761088, 1573847798947, 1573845267084, 1573590653304, 1573546824232, 1573546701522, 1573546441769, 1573546200232, 1573545950460, 1573490877735, 1573489912357, 1572994513567, 1572750643704, 1572750601204, 1572636858894, 1572233875943, 1571935338912, 1571879668817 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1749/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "ICLR.cc/2020/Conference/Paper1749/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "~Thomas_Litfin1" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "ICLR.cc/2020/Conference/Paper1749/Authors" ], [ "~Liang_Huang1" ], [ "ICLR.cc/2020/Conference/Paper1749/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1749/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1749/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Talk)\", \"comment\": \"This paper proposes a RNA structure prediction algorithm based on an unrolled inference algorithm. The proposed approach overcomes limitations of previous methods, such as dynamic programming (which does not work for molecular configurations that do not factorize), or energy-based models (which require a minimization step, e.g. by using MCMC to traverse the energy landscape and find minima).\\n\\nReviewers agreed that the method presented here is novel on this application domain, has excellent empirical evaluation setup with strong numerical results, and has the potential to be of interest to the wider deep learning community. The AC shares these views and recommends an enthusiastic acceptance.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"RNA Secondary Structure Prediction by Learning Unrolled Algorithms\\n\\nThis paper proposes E2Efold, which is an RNA secondary structure prediction algorithm based on an unrolled algorithm. Previous methods rely on dynamic programming (which does not work for molecular configurations that do not factorize) or rely on energy-based models (which require a minimization step, e.g. by using MCMC to traverse the energy landscape and find minima). The former does not work for all molecules and the latter can be difficult to optimize. The method presented here is novel, shows strong SOTA performance, and would be of interest to the wider deep learning community.\\n\\nThe method is based on an unrolled algorithm, which is motivated by the inclusion of three inductive biases / constraints important underlying RNA folding. These constraints limit the wide RNA search space. The first component of the method is a \\u201cDeep Score Network\\u201d which uses a stack of Transformer encoders (with relative and exact positional embeddings) followed by 2D convolutional layers to output a L x L symmetric matrix describing the \\u201cscores\\u201d of base pairing. As these scores may not obey the rules of RNA folding, a second post-processing network is trained end-to-end together with the \\u201cDeep Score Network\\u201d to enforce constraints. This network starts with a transformation that symmetrizes the matrix and applies a constraint-enforcing mask. The problem is transformed into an unconstrained problem by using Lagrange multipliers; it is then solved using a proximal gradient. Finally, a recurrent cell is defined that implements this algorithm in a deep learning framework. This method is creative, could be applied to other tasks with constraints, and would be interesting to the wider deep learning community.\\n\\nIn addition to developing the deep score network and post-processing network, the authors also develop a differentiable F1 loss, so that the network can directly optimize for precision and recall on the task. The performance of this method significantly outperforms previous methods. There was a fruitful discussion on OpenReview regarding whether this was a result of overfitting on the task. Indeed, it is critical in deep learning applications to carefully construct train/test sets to avoid high performance by memorization alone. To address this, the authors train on RNAStralign and test on ArchiveII. As the original ArchiveII dataset contains subsequences of other RNA sequences, which can result in overfitting, the authors re-ran their experiment with that removed, and similar results were achieved. To support the hypothesis that ArchiveII and RNAStralign capture different distributions, they perform a permutation test on the unbiased empirical Maximum Mean Discrepancy estimator, finding that the distributions are different. I do wonder why they did not check if P(ArchiveII) = P(ArchiveII) as they do check if P(RNAStr_train) = P(RNAStr_train). On the specific task of pseudoknot prediction, the method also performs well (F1 is >0.23 over the baseline). On sequence length-weighted F1, the model does even better.\\n\\nThe paper is rich with ablations. The analysis of the number of unrolling iterations T helps support the use of an unrolled method and builds intuition for its importance - it would be useful to include this in the appendix of the paper. I also appreciated the visualizations, which are a good sanity check that the model correctly handles pseudoknots. The performance of the method is broken down by RNA family, which is also quite interesting -- the method outperforms LinearFold on all classes, besides 5S RNA, SRP, and Group I intron. Further analysis is required to better understand why the method is weaker on those datasets. Additionally, further work should explore training on one set of families and testing on a held-out set of families. This was pointed out by public comments on this paper. This is potentially a limitation of E2EFold (the authors do not seem to have tried this suggested experiment) and further exploration is required. Exploring this limitation (even if it is not overcome) would make this paper even more rich.\\n\\nThat said, I recommend acceptance of this paper due to the extensive experiments, polished writing, novel method, and strong results, which can inspire future research.\"}", "{\"title\": \"Thank you very much for your positive feedback!\", \"comment\": \"Thank you very much for your positive feedback! Yes, we agree with you and we will keep refining the manuscript to make the final version more comprehensive.\"}", "{\"title\": \"RE: Author response\", \"comment\": \"I have read the other reviews (and comments from Liang and Thomas). My view of the paper (\\\"Accept\\\") is unchanged after reading that discussion.\\n\\nI believe it would be helpful to include a bit of that discussion in the paper (especially about how energy functions are commonly learned from data as well, just in an \\\"offline\\\", non-differentiable step), and certainly to include the additional results in a table or appendix, etc. I completely agree that it is important to make clear that the training and testing sets are different.\\n\\nFurther, even if we accept the criticisms of Liang and Thomas on the performance evaluation (again, I believe the authors credibly addressed these criticisms), there is still nice methodological contribution in this paper. As AnonReviewer3 said: \\\"(8) is cool!\\\"\"}", "{\"title\": \"Summary of major revisions\", \"comment\": \"We would like to thank all the reviewers for their careful reading, detailed comments, and overall positive assessment. We have responded to every raised question and concern, and incorporated your constructive suggestions into the revised version. The major revisions are summarized:\\n\\n\\n***More Details***\\n\\nWe\\u2019ve improved the clarity of the paper by adding\\n- More related works in Section 2 and Appendix A.\\n- Appendix B: Derivation Of The Proximal Gradient Step\\n- Appendix C: Implementation And Training Details\\nWe will also release our code for others to reproduce all the experimental results.\\n\\n\\n***More Experimental Results***\\n\\nWe\\u2019ve included additional results of experiments suggested by both the reviewers and the public comments.\\n- Appendix D.2. Two-sample Hypothesis Testing. (For better understanding data distributions in RNAStralign and ArchiveII)\\n- Appendix D.3. Performance On Long Sequences: Weighted F1 Score\\n- Appendix D.4. ArchiveII Results After Domain Sequences Are Removed\\n- Appendix D.5. Per-family Performances\\n\\n\\n***Flow of the paper***\\n\\nAs suggested, we add a new section \\u201cRNA Secondary Structure Prediction Problem\\u201d which follows the Related Work Section to formally state the problem and concrete constraints earlier than the previous version.\\n\\n\\nThank all the reviewers for the constructive suggestions on paper refinement!\"}", "{\"title\": \"Response To Official Blind Review #3\", \"comment\": \"We would like to thank the reviewer for the overall positive comments, constructive suggestions on paper refinement and references to interesting related works!\\n\\n\\n***Q1. \\u201c...output constraints doesn't occur until late...\\u201d\\n\\nThank you for your suggestion on the flow of our paper! In the revised version, we state the RNA secondary structure prediction problem, including the concrete constraints, in a newly added section \\u201c3 RNA Secondary Structure Prediction Problem\\u201d, which follows the Related Work section. \\n\\n\\n***Q2. Related work in NLP\\n\\nThank you for referring us to these related works in NLP! We also noticed the relation to NLP as we mentioned at the end of the introduction section. This indeed motivates us to use transformers and also the trick mentioned in BERT to compute position information by a series of base functions. However, it is interesting to know about projective parsing vs. non-projective parsing which we didn\\u2019t notice before! We\\u2019ve added a paragraph in the related work section to discuss this aspect. Thank you for pointing it out, which can help us relate our work to a larger range of problems in ML.\\n\\n\\n***Q3. Related work on Graphical Models and ISTA\\n\\nThanks! We found the \\u201cdeep unfolding\\u201d work very related and cited it in the revised paper.\\nIn our first submission, we\\u2019ve cited the unrolled ISTA paper \\u201cTheoretical linear convergence of unfolded ISTA and its practical weights and thresholds.\\u201d\\n\\n\\n***Q4. Motivation for the setup in \\\"Test On ArchiveII Without Re-training\\\"\\n\\nOne can think of ArchiveII as a separate held-out test dataset. E2Efold is only learned from RNAStralign training set, but can directly generalize to ArchiveII, and obtain the best test results. \\n\\nIn fact, testing on the ArchiveII dataset is a more challenging test for generalization, because the data distribution in ArchiveII can have a larger difference with the RNAStralign training set. To see this, we performed additional statistical hypothesis tests using Maximum Mean Discrepancy (MMD) [1] and attached the results in Appendix D.2. \\n\\nMore specifically, we computed the empirical MMD to evaluate the differences between\\n(a) RNAStralign_train and RNAStralign_test, where the MMD is 0.0025 (*can not reject* null hypothesis of no difference with p-value 0.1)\\n(b) RNAStralign_train and ArchiveII, where the MMD is 0.0296 (*reject* null hypothesis of no difference with p-value < 0.001)\\nThese tests suggest that the difference between RNAStralign training set and ArchiveII is much larger.\\n\\nTherefore, the data distribution in ArchiveII is very different from the RNAStralign training set. A good performance on ArchiveII shows a significant generalization power of our model.\\n\\n\\n***Q5. the number of optimizer iterations \\n\\nFor the explanation on the choice/effect of the number of optimizer iterations, please kindly refer to our response to this common question posted above.\\n\\n\\n***Q6. Does it work to train with a fixed number of unrolled iters, but at test time run the optimizer until convergence?\\u201d\\n\\nIt\\u2019s very interesting that the reviewer asked this question! We were also curious about this before and tried it empirically. We trained the model on T=20 and used it for T=50 iterations during the test phase. However, the performance is not as good as keeping T=20 for the test.\", \"we_think_the_reason_could_be\": \"we choose the discounting factor $\\\\gamma=1$ in equation 9, which is also a default choice in some related works. It gives the output at each time step T=t an equal weight. With a smaller $\\\\gamma$, the outputs at later steps can gain more weights. In this case, it is possible that the trained network will output a progressively closer approximation of the ground truth and further generalize to a larger number of iterations. We would investigate this option in the future.\\n\\n\\n***Finally\\n\\nYes, we also like the differentiable F1 score! Imbalanced data (more negative samples than positive samples) is a common issue in many computational biology problems (e.g. [4,5]) and our proposed method is very simple and effective in this case.\\n\\n\\n[1] Gretton, Arthur, et al. \\\"A kernel two-sample test.\\\" JMLR (2012)\\n[4] Zhou, Jian, and Olga G. Troyanskaya. \\\"Predicting effects of noncoding variants with deep learning\\u2013based sequence model.\\\" Nature methods 12.10 (2015): 931.\\n[5] Armenteros, Jos\\u00e9 Juan Almagro, et al. \\\"SignalP 5.0 improves signal peptide predictions using deep neural networks.\\\" Nature biotechnology 37.4 (2019): 420.\\n\\n\\nPlease expect the revised paper posted soon.\"}", "{\"title\": \"Response To Official Blind Review #1 (part 2)\", \"comment\": \"Here we provided our responses to the second part of reviewer 1\\u2019s comments.\\n\\n***Overstates the improvements\\n\\nAccording to the reviewer\\u2019s suggestion, we\\u2019ve moderated our statement by simply saying \\u201cbetter than previous SOTA in terms of F1 scores\\u201d. \\nBtw, the performances on RNAStralign are estimated on a held-out test set, so we think the comparison on this test set is also valuable.\\n\\n\\n***performance per RNA category\\n\\nTo balance the performance among different families, during the training phase we conducted weighted sampling of the data based on their family size. With weighted sampling, the overall F1 score (with shifted) is 0.83, which is the same as when we did i.i.d. sampling. The per-family F1 scores are provided below. More numbers are included in Appendix D.5 in the revised paper.\", \"table_1\": \"RNAStralign (test set): per-family F1 score\\n |16S rRNA| tRNA | 5S RNA | SRP | tmRNA | Grp I | RNaseP | telomerase \\nfamily size | 11620 | 9385 | 6443 | 468 | 572 | 1502 | 434 | 37\\nE2Efold | 0.783 | 0.917 | 0.906 | 0.550 | 0.588 | 0.387 | 0.565 | 0.954 \\nLinearFold | 0.493 | 0.734 | 0.713 | 0.618 | 0.393 | 0.565 | 0.567 | 0.515\\nMfold | 0.362 | 0.662 | 0.356 | 0.350 | 0.290 | 0.483 | 0.562 | 0.403\\nRNAstructure| 0.464 | 0.709 | 0.578 | 0.579 | 0.400 | 0.566 | 0.589 | 0.512\\nRNAfold | 0.430 | 0.695 | 0.592 | 0.617 | 0.411 | 0.589 | 0.544 | 0.471 \\nCONTRAfold | 0.529 | 0.758 | 0.717 | 0.563 | 0.463 | 0.603 | 0.645 | 0.529\\n\\nE2Efold performs significantly better than other methods in 16S rRNA, tRNA, 5S RNA, tmRNA, and telomerase, and these results are from *a single model*. \\n\\nIn the future, we can view it as multi-task learning and further improve the performance by learning multiple models for different families and learning an additional classifier to predict which model to use for the input sequence.\\n\\n\\n***Q3. Derivation for the proximal gradient steps\\n\\nWe added include the derivation steps for Eq. 3-5 in Appendix B in the revised version.\\n\\n\\n***Q4. Why is there a need to introduce an l_1 penalty term to make A sparse?\\n\\nThis is mainly due to the *fixed* number of iterations T. We prefer using a comparatively *small T* (e.g. T=20) as we explained above. In this case, we use a sparse penalty to help the algorithm quickly incorporate the constraints to the output within T iterations, since it is our prior knowledge that the global constraints (e.g. each $x_i$ can only be paired with at most *one* base $x_j$) will give us a sparse output.\\n\\nFor example, within T=20 iterations, we compare the differences between adding and not adding the sparse penalty:\", \"table_2\": \"Constraints-check.\\n \\t\\t | $\\\\max_{A}c(A)$ | $\\\\text{mean}_{A}c(A)$\\nWith sparse penalty | 4 | 0.65\\nWithout sparse penalty | \\t 28 | 3.975 \\n \\nIn Table 1, $c(A):=\\\\text{sum}(\\\\text{relu}(A\\\\mathbf{1} - \\\\mathbf{1}))$ is measuring whether the constraints are satisfied. The smaller the better. The benefit of adding the penalty term is obvious, in the case when T is fixed and comparatively small. \\n\\n\\n***Q5. On which data is Table 6?\\n\\nRNAStralign. We are sorry for the lack of clarity. We\\u2019ve indicated the dataset in the caption of Table 6 in the revised version.\\n\\n\\n***grammar error/typo\\n\\nYes! In the equation above Equation 2, it should be $-\\\\rho||\\\\hat{A}||_1$ instead of plus! Thank you for pointing all the typos/errors out! We believe they are now fixed. \\n\\n\\nPlease expect the revised paper posted soon.\"}", "{\"title\": \"Response To Official Blind Review #1 (part 1)\", \"comment\": \"We thank reviewer 1 for careful reading and constructive suggestions for paper refinement! We separate our response to reviewer 1 into two parts.\\n\\nThe first part of reviewer 1\\u2019s comments is about missed important details. We are sorry for the lack of clarity and thank the reviewer for pointing them out! Now the details are included in the revised paper, and we also explained them below:\\n\\n\\n***Compare to existing approaches for unrolling optimization problems\\n\\nWe explained the novelty/difference of E2Efold compared to the existing approaches below. They are now included in the main text and appendix A.\\n\\nFirst, our view of incorporating constraints to reduce output space and to reduce sample complexity is novel. Previous works [Belanger et al., 2017; Pillutla et al., 2018; Ingraham et al., 2018] did not discuss these aspects. The most related work which also integrates constraints is OptNet [Amos & Kolter, 2017], but it\\u2019s very expensive and can not scale to the RNA problem. Therefore, our proposed approach is a simple and effective one.\\n\\nSecond, compared to [Andrychowicz et al., 2016, Chen et al., 2018; Shrivastava et al., 2019], our approach has a different purpose of using the unrolled algorithm. Their goal is to learn a better algorithm, so they commonly make the architecture *more flexible* than the original algorithm for the room of improvement. However, we aim at enforcing constraints. To ensure that constraints are nicely incorporated, we keep the original structure of the algorithm and only make the hyperparameters learnable.\\n\\nFinally, although all works consider end-to-end training, none of them can directly optimize the F1 score. We proposed a differentiable loss function to mimic the F1 score/precision/recall, which is effective and also very useful when negative samples are much fewer than positive samples (or the inverse).\\n\\n\\n***Details on the unrolling constant T\\n\\nFor the explanation on the choice/effect of unrolling constant T, please kindly refer to our response to this common question posted above.\\n\\n\\n***Details on the design of the training process\\n\\nNow all the training details are included in Appendix C of the revised paper. \\n\\nConnecting the score network and PPN, the whole network is very deep. Both the pre-training and the augmented loss are the tricks that we use to speed up and stabilize the training process. \\n- Pre-train: Since we expect $U_{\\\\theta}(i,j)$ to be higher if $(x_i,x_j)$ are truly paired, we use the true secondary structure to pre-train the score network to quickly get a fairly good U.\\n- Augmented loss: During joint training, we use (loss in Eq 9) + $c \\\\cdot$(logistic regression loss on $U_{\\\\theta}$) where we set $c=1$. The second term is estimated at the *intermediate output*. Although this term is optional, we think it can help stabilize the training of the *very deep* network and also make each gradient step more efficient.\\n\\nBesides, we\\u2019ve submitted a link to our code through a private comment to reviewers. We will get this code well-organized and released to the public for others to reproduce all the experimental results.\"}", "{\"title\": \"Response To Official Blind Review #2\", \"comment\": \"We thank the reviewer for the positive comments about the approach and the suggestions for paper refinement! We\\u2019ve included all the important details suggested by the reviewer in either the main paper or appendix. Experimental results are demonstrated to show how we choose the unrolling constant T.\\n\\n\\n***Q1. details of the configuration\\n\\nWe\\u2019ve added a section (Appendix C) to explain our configurations including hyperparameters, MLP details, and the \\u2018pairwise concatenation\\u2019 details. \\n\\nBesides, we\\u2019ve submitted a link to our code through a private comment to reviewers. We will get this code well-organized and released to the public for others to reproduce all the experimental results.\\n\\n\\n***Q2. \\u201cWhat unrolling constant T is used? \\u201d\\n\\nFor the explanation on the choice/effect of unrolling constant T, please kindly refer to our response to this common question posted above.\\n\\n\\n***Q3. details on how U_\\\\theta is trained by itself\\n\\nWe are sorry for the lack of clarity in our first submission. Now all the training details are included in the revised paper. Please kindly refer to Appendix C. Briefly speaking, the training process of $U_{\\\\theta}$ is the same as E2Efold, which consists of two steps. First, quickly pre-train $U_{\\\\theta}$ using logistic regression. Second, use the loss in eq 9 to jointly train $U_{\\\\theta}$ and $PP_{\\\\phi}$ end-2-end (when $U_{\\\\theta}$ is trained by itself we can simply set T=0 in the PP network).\\n\\n\\n***Q4. English grammar and writing style\\nWe appreciate the reviewer for the very close reading! We believe the typos/errors are now fixed.\\n\\n\\nPlease expect the revised paper posted soon.\"}", "{\"title\": \"The choice of the number of unrolling iterations T\", \"comment\": \"We explained the choice of the number of unrolling iterations T in this post, since all reviewers asked questions about this constant.\\n\\nThe performance that we reported in the paper is when T=20. The overall ideas are:\\n- to ensure that the constraints are mostly satisfied, T can not be too small. \\n- a very large T will make the neural network very deep and the training process more expensive and unstable.\", \"what_we_actually_did_is\": [\"We first pre-train the score network $U_{\\\\theta}$ without considering the post process network.\", \"Then we fix this trained $U_{\\\\theta}$ and try different T without further training (Table 1).\", \"After the best T is selected, the score network and the post-processing network are trained end-2-end.\"], \"table_1\": \"Constraints-check and validation F1 scores\\n | $\\\\max_{A}c(A)$ | $\\\\text{mean}_{A}c(A)$ | validation F1\\nT = 5 |\\t 160\\t | 28.53 | 0.772\\nT = 10 | 33 | 7.78 | 0.791\\nT = 20 | \\t 4\\t | 0.65 | 0.806 \\nT = 30 | 1\\t | 0.03 | 0.809\\nT = 50 | \\t 1\\t | 0.03 | 0.808\\n\\nIn Table 1,\\n- $c(A):=\\\\text{sum}(\\\\text{relu}(A\\\\mathbf{1} - \\\\mathbf{1}))$ is measuring whether the constraints are satisfied. The smaller the better.\\n- validation F1 scores for T=20-50 are similar. This motivates us to choose T=20 for more efficient training. \\n\\n\\nIn addition, we\\u2019ve also tried to train the model for T=50 end-to-end.\", \"table_2\": \"T=20 vs T=50\\n | time per gradient step | best training loss | best validation loss | validation F1\\nT = 20 | \\t 0.64s\\t |\\t 0.63\\t |\\t 0.64 | 0.88\\nT = 50 | \\t 0.98s\\t |\\t 0.65 | 0.66\\t | 0.86\", \"in_table_2\": [\"The training time is evaluated on the Titan Xp card with batch size = 8. (When the batch size is 16, it will be out of memory for the case T=50.)\", \"The final performances are indeed similar. It is not that sensitive to the choice of T.\", \"The performance for T=50 is a bit worse, which might be caused by the depth of the network, making the optimization harder.\"]}", "{\"title\": \"Setting 1 is a reasonable, read-world setting and the solution is not trivial\", \"comment\": \"We are happy that Thomas found our method novel, but we can not agree with some of his statements.\\n\\n***setting 1 is a reasonable and real-world setting***\\n\\n(1) We can not agree with the conclusion that: the comparison is unfair since E2Efold can incorporate information from seen sequences and others can not. The setting is fair as long as in practice we have access to data like that. Especially, the dataset of RNA structures is continually growing.\\n\\n(2) As we mentioned earlier, some methods we compared with use a fitted energy function. We note that these energy functions can be fitted from a much larger set of available datasets that are not used in E2Efold. In some sense, these methods may actually have an advantage over E2Efold in terms of datasets used for training. \\n\\n(3) We do not think traditional models are *designed for* setting 2 as Thomas stated. Instead, they are designed in that way and are commonly adopted because it is more interpretable and biologists care about interpretability. In contrast, deep learning models have better abilities to learn the common pattern from data and make more accurate predictions, while it is less interpretable. (However, DL is a hot topic and now there are many methods to explain the prediction of DL models.)\\n\\nIn conclusion, in terms of prediction accuracy, our comparison is fair and predicting structures for unseen sequences is the real-world setting.\\n\\n\\n***predicting structures for unseen sequences in setting 1 is not a trivial task***\\n\\nPlease refer to the results of *CDPfold* in Table3 and the result of *$U_{\\\\theta}$+PP* in Table 6. Both of them are deep learning methods. If it is a trivial task and neural networks only need to memorize and overfit the training data, there won't be such a performance gap between those DL models and E2Efold. They should also work extremely well as long as the neural network is not so small. Especially, *$U_{\\\\theta}$+PP* has exactly the same architecture as E2Efold, except that the gradient is not pushed through the post-process during the training phase.\\n\\nTherefore, designing a suitable deep learning model to better learn the pattern of the structures is not easy. We should not take its performance for granted.\"}", "{\"title\": \"Results of the methods suggested by Thomas\", \"comment\": \"We thank Thomas for referring us to more related methods. We\\u2019ve investigated and tried these methods (see more details below), although they are not commonly used as baselines for RNA folding.\\n\\n***Turbofold is extremely slow and performs particularly bad on long sequences***\\n\\nWe investigated the TurboFold package. It has serious flaws compared to our approach. TurboFold is extremely slow, requiring more than 15 minutes for only 5 sequences. This is true even for the parallel version, TurboFold-smp, for which we used 32 processors. Without multi-processing, the program won\\u2019t return any results for just 5 16S-rRNA inputs in a reasonable time. The running time of TurboFold seems to increase exponentially in the sequence length. Also, to run TurboFold for a set of sequences, the RNA class for each sequence must be known. Its performance, 0.75 F1, is worse than E2Efold. In addition, it performs particularly bad on *long* sequences, such as 16S rRNA (F1=0.5). E2Efold achieves 0.78 F1 in this class. \\n \\nThe binary version of TurboFold has a specific input format requirement. We spent a very long time making the program successfully run on the whole testing dataset. Here is our code for running TurboFold:\", \"https\": \"//drive.google.com/open?id=19d8nBYQx2qHEtEq-cMsmUkQ8bS9F-UEZ\\n\\n\\n[1] Fallmann, Joerg, et al. \\\"Recent advances in RNA folding.\\\" Journal of biotechnology (2017)\\n[2] Altschul, Stephen F., et al. \\\"Gapped BLAST and PSI-BLAST: a new generation of protein database search programs.\\\" Nucleic acids research (1997)\\n[3] Sievers, Fabian, and Desmond G. Higgins. \\\"Clustal Omega, accurate alignment of very large numbers of sequences.\\\" Multiple sequence alignment methods. 2014 \\n[4] Zhen Tan, Yinghan Fu, Gaurav Sharma, David H. Mathews. \\\"TurboFold II: RNA structural alignment and secondary structure prediction informed by multiple homologs.\\\" Nucleic Acids Research (2017)\"}", "{\"title\": \"Clarification of evaluation\", \"comment\": \"I am in full agreement with the feedback of Liang Huang. The integration of hard constraints in the deep learning framework is a good idea which is well suited to the RNA structure prediction problem. However, I believe that the evaluation requires further clarification.\\n\\nRNA structure is remarkably conserved within a family even below 100% sequence identity. This means that we can consider 2 relevant settings for evaluation:\\n1) The setting whereby there exists known structures from the same family.\\n2) The setting whereby predictions are made for previously unseen families.\\n\\nThe authors propose a method for setting 1 and compare to methods designed for setting 2. In addition, my understanding is that they evaluate competing methods considering only a single-sequence prediction. However, this may be an unfair comparison with E2Efold which is able to incorporate information from many family-specific homologs. Within the RNAstructure package, Turbofold may be used to generate family-specific consensus secondary structures from a collection of homologous sequences.\\n\\nThey also briefly discuss a few-shot learning framework for families with few solved structures. In fact, a specific realisation of this framework is long-established in structural bioinformatics and is known in the field as homology modelling. The most naive version of this method is to align RNA sequences from the same family based on nucleotide substitution scores and transfer the known secondary structure to the unknown sequence based on the alignment. There are also more sophisticated realisations of this idea (eg Infernal). A comparison with a homology modelling method is necessary to evaluate the performance in setting 1, however the conservation of secondary structure within RNA families likely leaves little room for improvement over the alignment-based methods.\\n\\nHowever, the framework proposed by the authors can also be evaluated in setting 2 by evaluating the model on families unseen during training. Strong performance in this setting would certainly be important in the context of structural bioinformatics.\"}", "{\"title\": \"Response: a further clarification (part 2)\", \"comment\": \"*** Probknot vs RNAstructure\\n\\n\\\"The authors seem to report Probknot as RNAstructure, but the latter (by default) can't predict pseudoknots.\\\" : this statement is not true! \\n\\nRNAstructure web-server has been constantly updated, and Probknot is now included in the RNAstructure web-server:\", \"https\": \"//rna.urmc.rochester.edu/RNAstructureWeb/\\nIt does predict pseudoknots as shown in visualizations and table 5. \\n\\n*** Data requirement for DL-based methods\\n\\nThe training data requirement cannot be avoided for DL models, so we do not deny the fact (in the discussion section) that E2Efold does not perform well on telomerase where there are only 37 data points. This problem needs to be resolved by future studies on few-shot learning or curriculum learning. Meanwhile, we expect that such training data will be accumulating in the future, too. \\n\\nFor the past decades, DP-based algorithms have dominated the RNA secondary structure prediction. We believe that it is worth trying to apply deep learning to this problem even when some RNA families do not contain enough data at this moment. Furthermore, our methods present a more sensible bias-variance trade-off (deep while incorporating problem structure in the architecture), and thus a better result is reasonable. \\n\\n*** Methodology contribution\\n\\nFinally, we believe that the methodology proposed in our paper is of broad interest to the audience of the ML community. Though our work is originally motivated by the challenges in the RNA folding problem, we believe our ideas of\\n(i) incorporating hard constraints to reduce the output space and thus making it more data-efficient; and\\n(ii) using an unrolled algorithm for solving constrained programming as a building block in a neural network\\nwill be inspiring for other structured prediction problems and potentially useful for architecture design in a wider range of DL problems. Therefore, we think that our work is well-suited for ICLR.\"}", "{\"title\": \"Response: a further clarification (part 1)\", \"comment\": \"We thank Liang Huang for your comments. We have carefully designed our experiments, and clarified further the settings below:\\n\\n\\n*** Generalization ability\\n\\nArchiveII is a separate held-out test dataset. E2Efold is only learned from RNAStralign training set, but can directly generalize to ArchiveII, and obtain the best test results. \\n\\nWe note that LinearFold is not a learning-based method, but one energy function used in LinearFold is a learning-based energy from another paper (i.e. CONTRAfold), which can potentially be fitted from a much larger number of available datasets. Thus the per-family error of LinearFold is not a cross-validation error in the strict learning sense. \\n\\n\\n*** No redundant sequences\\n\\nAs stated in Sec 5, \\\"After removing redundant sequences and structures...\\\", we've carefully removed redundant sequences and structures, which reduces the RNAStralign dataset size from 37149 to 30451. Also, when we test the learned model on ArchiveII, we've excluded sequences that are overlapped with the RNAStralign dataset.\\n\\nThe similarity between sequences is a characteristic of the dataset. Some level of similarity between sequences is also a basis for model generalization. \\n\\n\\n*** Accuracy dominated by short-sequence?\\n\\nFor long sequences, E2Efold still performs better than other methods. In fact, we've computed F1 scores *weighted by the length of sequences* (Table 1), such that the results are more dominated by longer sequences (we've conducted this experiment before the submission, but did not report them in the original submission due to page limit). For instance, for the RNAStralign dataset: \\n\\nTable 1. RNAStralign: F1 after a weighted average by sequence length.\\n |E2Efold|CDPfold|LinearFold| Mfold |RNAstructure|RNAfold|CONTRAfold\\noriginal | 0.821 | 0.614 | 0.609 | 0.420 | 0.550 | 0.540 | 0.633\\nweighted| 0.720 | 0.691 | 0.509 | 0.366 | 0.471 | 0.444 | 0.542\\nchange |-12.3% | +12.5% | -16.4% | -12.8% | -14.3% | -17.7% | -14.3% \\n\\n3rd row reports how much F1 score drops after reweighting. \\n\\n\\n*** Subsequences\\n\\nSince subsequences in ArchiveII are explicitly labeled, we filtered them out in ArchiveII and recomputed the F1 scores (Table 2) as Liang suggested. \\n\\nTable 2. ArchiveII: F1 after subsequences are filtered out.\\n |E2Efold|CDPfold|LinearFold| Mfold |RNAstructure|RNAfold|CONTRAfold\\noriginal| 0.704 | 0.597 | 0.647 | 0.421 | 0.613 | 0.615 | 0.662\\nfiltered | 0.723 | 0.605 | 0.645 | 0.419 | 0.611 | 0.615 | 0.659\\n\\nThe results do not change too much before or after filtering out subsequences.\\n\\nIn the original submission, we do not exclude subsequences by default because we want to follow RNA structure prediction benchmark. More specifically, \\n- as suggested in the latest review by [Mathews, 2019], \\\"The collection of benchmarking structures we collected, called ArchiveII, is available for download from our lab website at https://rna.urmc.rochester.edu/publications.html\\\", ArchiveII is considered the benchmarking dataset and by default both domain sequences and full sequences are included. \\n- the subsequences in ArchiveII are domains, whose secondary structure can be predicted independently since under most circumstances there is no cross-domain base-pairing.\\n\\n[Mathews, 2019] How to benchmark RNA secondary structure prediction accuracy, Methods, 2019\\n\\n\\n*** Runtime comparison\\n\\nIn practice, training needs to be done much less often than testing, and hence we focus on reporting the testing time. For instance, deep learning models for Imagenet need to be trained for a long time, but inference (testing) occurs more often, and lots of researches are done for improving inference time. \\n\\nFurthermore, it is undeniable that Dynamic Programming (DP) based algorithms are not easy to be parallelized or sped up by GPU due to their *sequential decision* nature. Thus, the runtime of DP-based algorithms has a significant dependency on the length n. \\n\\nIn contrast, the operations in E2Efold are mainly *matrix computations*, which can be easily sped up by GPU using modern programming languages such as pytorch and tensorflow.\\n\\nIndeed, E2Efold will have even more advantages if we only report the runtime on long sequences. For example, we run E2Efold and LinearFold for sequences of length from 1500 to 1800 (\\\"Long\\\" in the table):\\n |E2Efold (Python)|LinearFold (C)| \\n All | 0.40s | 0.43s |\\nLong| 0.41s | 1.25s |\\n\\nIn addition, other methods are even implemented in C, which is by nature faster than Python that we used for E2Efold.\"}", "{\"title\": \"interesting idea but seriously flawed in evaluation\", \"comment\": \"As the primary author of the LinearFold paper (Huang et al, 2019) which this submission compares to, I found this deep learning approach novel and interesting, being vastly different from the mainstream dynamic programming (DP)-based algorithms. However, the evaluation in this work is seriously flawed in many ways:\\n\\n1. Dataset: (a) mistakenly using subsequences as full sequences\", \"this_paper_uses_two_datasets_from_the_mathews_group\": \"RNAStralign and ArchiveII. However, the authors did not notice that many sequences in the 16S and 23S families are labeled as *sub domains* and thus sub sequences, not full sequences. For example, 16S is family of ~1,500 nts with length in the range of [950,1995] in ArchiveII, while in Table 1 of this paper, 16S's minimum length is only 73 and 54 for the two datasets. This obvious mistake shortens the average length of the datasets considerably.\\n\\n2. Dataset: (b) each family is extremely homogeneous in sequences and structures\\n\\nThis paper is mainly trained and tested on RNAStralign, but that dataset was made for homologous folding -- meaning within each family (such as tRNA), all sequences are very similar to each other, and they share almost the same structures. As an example, let's see the first 5 sequences in tRNA:\\n\\n1 72 tdbD00000002.ct\\nGGGCUCAUAGCUCAGCGGUAGAGUGCCUCCCUUGCAAGGAGGAUGCCCUGGGUUCGAAUCCCAGUGAGUCCA\\n(((((((..((((.......)))).(((((.......))))).....(((((.......)))))))))))).\\n2 72 tdbD00000003.ct\\nGGGCUCAUAGCUCAGCGGUAGAGUGCCUCCUUUGCAAGGAGGAUGCCCUGGGUUCGAAUCCCAGUGAGUCCA\\n(((((((..((((.......)))).(((((.......))))).....(((((.......)))))))))))).\\n3 72 tdbD00000004.ct\\nGGGCUCAUCGCUCAGCGGUAGAGUGCCUCCCUUGCAAGGAGGAUGCCCUGGGUUCGAAUCCCAGUGAGUCCA\\n(((((((..((((.......)))).(((((.......))))).....(((((.......)))))))))))).\\n4 72 tdbD00000005.ct\\nGGGCUCGUAGCUCAGCGGGAGAGCGCCGCCUUUGCGAGGCGGAGGCCGCGGGUUCAAAUCCCGCCGAGUCCA\\n(((((((..((((.......)))).(((((.......))))).....(((((.......)))))))))))).\\n5 72 tdbD00000006.ct\\nGGGCUCGUAGCUCAGCGGGAGAGCGCCGCCUUCGCGAGGCGGAGGCCGCGGGUUCAAAUCCCGCCGAGUCCA\\n(((((((..((((.......)))).(((((.......))))).....(((((.......)))))))))))).\\n\\nTherefore, training and testing on the same families is always too easy and kind of overfitting. The correct way to do evaluation is the \\\"cross-validation\\\" way: training on N-1 families, and testing on the held out one, and repeat for all families and report the average.\\n\\n3. Evaluation metrics\\n\\nThis paper only reports overall accuracy, but no per family accuracy. In addition, the overall accuracy is averaged over sequences, not families. By contrast, standard RNA folding evaluation should report per-family accuracies and while for each family, the accuracies are averaged over all sequences in this family, the overall accuracy should be averaged over all families (i.e., equal weight per family). Please refer to the LinearFold paper and other papers from the Mathews group which the authors cited (e.g., ProbKnot). In the authors' evaluation, the accuracy is dominated by short sequences (see Figure 6). They also admit prediction accuracy for some families are bad in Discussion: \\u201cFor telomerase class, which only contains 37 samples in the dataset, E2Efold performs worse than classic methods.\\u201d Thus, their claim in abstract that \\u201cit predicts significantly better structures compared to previous SOTA (29.7% improvement in some cases in F1 scores and even larger improvement for pseudoknotted structures)\\u201d is misleading.\\n\\n4. Model generalization:\\n\\nAs mentioned in 2, this paper does not apply cross-valid training and testing, instead they use RNAStralign-train (split from RNAStralign) as a training set, and RNAStralign-test and ArchiveII as testing sets. Moreover, they do not include all families in ArchiveII, but only the ones that have overlapping types (5S, 16S rRNA etc) in RNAStralign-train; i.e., they do not include 23S rRNA when testing on ArchiveII. Thus, their model\\u2019s generalization ability is suspicious and it may not work well on 23S rRNA. It is likely that their prediction examples perform well because of overfitting.\\n\\n5. Runtime analysis\\n\\nThis paper claims they are faster than LinearFold and RNAstructure for inference time (see table 4). But they don\\u2019t provide the runtime with sequence length, instead they only compared the total runtime on RNAstralign. It is likely that their model costs much more time on long sequences. Moreover, the runtime is even tested on different hardware, for theirs on GPU, while LinearFold and RNAstructure on CPU. Besides, they do not report training time, which could be very costly.\\n\\n6. (minor) RNAstructure vs Probknot\\n\\nThe authors seem to report Probknot as RNAstructure, but the latter (by default) can't predict pseudoknots.\\n\\nOverall, this paper contains interesting ideas, but their evaluation is so flawed that their results are just overfitting. Cross-validation is absolutely needed for this type of data set where sequences/structures are so similar. Therefore, this paper can not be published in the current form.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper introduces an end-to-end method to predict the secondary structure of RNA, by mapping the nucleotide sequence to a binary affinity matrix. The authors decompose this problem into two part: (i) predicting an affinity score between each base pair in the input sequence, using a combination of a transformer sequence encoder network and a convolutional decoder, and (ii) a post-processing step that ensures that structural local and global constraints are enforced. An innovation is to express this post-processing as an unrolled sequence of proximal gradient descent steps, which are fully differentiable, and allow the full combination of (i)+(ii) to be trained end-to-end. A thorough set of experiments validate the approach.\\n\\nOverall, the paper is well written and easy to follow. The approach of unrolling structural constraints as shown in the paper is interesting and applicable to much wider domains than secondary structure prediction. The proposed approach appears to provide a novel, convincing and non-obvious solution to RNA secondary structure prediction, and subject to suggestions below, would represent a valuable contribution to ICLR.\\n\\nThe principal area for improvement would be to include additional detail (perhaps in appendix) on the model hyperparameter configurations that were used in the experiments. Moreover, more details on the set of \\\\psi functions, and the MLP details for P_i (e.g. number of hidden units, activation function, the use of dropout, batch normalization, etc) should be given, as well as more information on the specifics how how the \\u201cpairwise concatenation\\u201d is carried out in the output layer. What unrolling constant T is used? Finally, in the ablation study (p. 8) details on how U_\\\\theta is trained by itself (without the post-processing step) should be given.\", \"detailed_comments\": [\"Overall, the whole paper should be thoroughly reviewed for English grammar and writing style; a subset of suggested changes follow.\", \"p. 1: structure a result ==> structure is a result\", \"p. 2: energy based methods ==> energy-based methods\", \"p. 2: energy function based approaches ==> energy function-based approaches\", \"p. 2: view point ==> viewpoint\", \"p. 2: E2Efold is flexible ==> E2Efold are flexible\", \"p. 2: nearly efficient ==> nearly efficiently\", \"p. 3: typically scale ==> typically scale as\", \"p. 3: few hundred. ==> few hundreds.\", \"p. 4: all binary matrix ==> all binary matrices\", \"p. 4: output space can help ==> output space could help\", \"p. 5: formulation are the ==> formulation are that the\", \"p. 6: eq. (7) should contain quantities indexed by $t$ in the RHS\", \"p. 8: pesudoknotted ==> pseudoknotted\", \"p. 9 ff: in the bibliography, all lowercase rna should be uppercase RNA. Use {RNA} in bibtex entries.\"]}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The authors proposed an end-to-end method (E2Efold) to predict RNA secondary structure. The method consists of a Deep Score Network and a Post-Process Network (PPN). The two networks are trained jointly. The score network is a deep learning model with transformer and convolution layers, and the post-process network is solving a constrained optimization problem with an T-step unrolled algorithm. Experimental results demonstrate that the proposed approach outperforms other RNA secondary structure estimation approaches.\\n\\nOverall I found the paper interesting. Although the writing can be improved and some important details are missing.\\n\\nMajor comments\\nAs the authors point out, several existing approaches for unrolling optimization problems have been proposed. It would be helpful to clarify the methodological novelty of the proposed algorithm compared to those.\\n\\n\\nTraining details and implementation details are missing; these hinder the reproducibility of the proposed approach. The author stated pre-training of the score network, how is the PPN and score network updated during the joint training? Does the model always converge? The authors vaguely mentioned add additional logistic regression loss to Eq9 for regularization. What is a typical number of T? How does varying T affect the performance, both in terms of training time (and convergence) and in terms of accuracy/F1?\\n\\nMinor comments\\nThe 29.7% improvement of F1 score overstates the improvements compared to non-learning approaches.. This performance was computed on the dataset (RNAStralign) on which E2Efold was trained. A fair comparison, as the authors also stated, is on the independent ArchiveII data. On this data, E2Efold has F1 score 0.686 versus 0.638 for CONTRAfold. The author should report performance improvement under this line.\\n\\n\\nIt would be helpful to report performance per RNA category, both for RNAstralign data and ArchiveII data, while the ArchiveII data should still remain independent. Different models may have their strengths and weaknesses on different RNA types.\\n\\n\\nIt is not obvious to me how the proximal gradient was derived to (3)-(5). It would be helpful if the authors show some details in the supplements.\\n\\n\\nWhy is there a need to introduce an l_1 penalty term to make A sparse?\\n\\n\\nOn which data is Table 6?\\n\\nTypos, etc.\\nThe references are not consistently formatted\\n\\u201cstructure a result\\u201d -> \\u201cstructure is a result\\u201d\\n\\u201ca few hundred.\\u201d -> \\u201ca few hundred base pairs.\\u201d\\n\\u201cobjective measure the\\u201d -> \\u201cobjective measures the\\u201d\\n\\u201csection 5\\u201d -> \\u201cSection 5\\u201d (in several places)\\nIn the equation above Equation 2, should it be -\\\\rho||\\\\hat{A}||_{1} instead of plus? Otherwise, the \\u201cmax\\u201d could be made arbitrarily large.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"*Summary*\\nThe authors perform RNA secondary prediction using deep learning. The outputs are subject to hard constraints on which nucleotides can be in contact with others. They unroll a sophisticated optimization algorithm for a relaxation of the task of finding the optimal contact map subject to these constraints. This work is in a long line of work demonstrating that end-to-end training of models that incorporate application-specific optimization routines as sub-modules is very useful. In particular, it outperforms an approach where the inputs to this optimization problem come from a network that was trained using a simple loss that ignores the fact that it will feed into this structured optimizer. The paper also considers an application domain that will be unfamiliar to many ICLR readers interested in deep structured prediction, and may serve as a call to arms for the community engaging with additional problems in this field. \\n\\n*Overall Assessment*\\nThe paper is well written, well executed, and part of a general research thread that ICLR readers care about. There are a number of technical details, such as the loss function in (8) that will be of general interest. I advocate for acceptance.\\n\\n*Comments*\\nThe actual specification of the output constraints doesn't occur until late in the paper. Before then, the discussion of them is very abstract. Given that the constraints are easy to describe, the exposition would be improved notably if you described the specific constraints earlier on. This would help me understand the problem domain better.\\n\\nFyi, the idea of nested structures vs. non-nested structures appears in NLP in terms of projective parsing vs. non-projective parsing. There may be some relevant reading for you to do there. Your specific work (minus the unrolled constraint enforcement) is similar to \\\"Dozat et al. 2017. Deep biaffine attention for neural dependency parsing.\\\"\\n\\nThe idea of backpropping through some constraint-enforcing process is reminiscent of backpropping through belief propagation. See, for example, Domke's \\\"Learning Graphical Model Parameters with Approximate Marginals Inference.\\\" Or Hershey et al. \\\"Deep Unfolding: Model-Based Inspiration of Novel Deep Architectures.\\\" You should also cite work using unrolled ISTA to learn sparse coding dictionaries. They have terms similar to (5). \\n\\nWhat exactly was your motivation for the setup in \\\"Test On ArchiveII Without Re-training?\\\"\\n\\nHow sensitive is performance to the number of optimizer iterations? Does it work to train with a fixed number of unrolled iters, but at test time run the optimizer until convergence?\\n\\n(8) is cool!\"}" ] }

[ "Ruishan Liu", "Akshay Balsubramani", "James Zou" ]

[ "correspondence", "datasets", "side information", "transport cost", "cost function", "available", "images", "multiple datasets", "important problem" ]

{ "note_id": [ "3QRlH-tVV_", "rJeKM2Z2oH", "H1e1pibnir", "rylf9sZ2ir", "BJx8MZsRFS", "SklG_8hatS", "r1lYrm46FS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731465, 1573817361133, 1573817270750, 1573817226306, 1571889422043, 1571829354210, 1571795776885 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1748/Authors" ], [ "ICLR.cc/2020/Conference/Paper1748/Authors" ], [ "ICLR.cc/2020/Conference/Paper1748/Authors" ], [ "ICLR.cc/2020/Conference/Paper1748/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1748/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1748/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The paper proposes an algorithm for learning a transport cost function that accurately captures how two datasets are related by leveraging side information such as a subset of correctly labeled points. The reviewers believe that this is an interesting and novel idea. There were several questions and comments, which the authors adequately addressed. I recommend that the paper be accepted.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response\", \"comment\": \"Thank you for your thoughtful review and suggestions.\\n\\nRegarding your motivation for dataset alignment, aligning datasets across different conditions is a very important problem in biology, especially in single cell analysis. This is highlighted in several recent high-profile papers, e.g. Butler et al Nature Biotechnology 2018 \\u201cIntegrating single-cell transcriptomic data across different conditions, technologies, and species\\u201d. In these experiments, there is typically insufficient data or strong batch effects within a single dataset, and how to align datasets in order to increase sample and remove batch effect is considered one of the most important open challenges in computational genomics [Haghverdi et al Nature Biotechnology 2018]. This motivates our work on data alignment.\\n\\nRegarding your specific suggestions.\\n\\n1. In single cell analysis, which is our motivating example, we often have annotations that certain cluster of cells in a dataset corresponds to one cell type (e.g. B cells and T cells) based on known marker genes (e.g. Tabula Muris Nature 2018). When we have datasets D1 and D2, we typically know which are the B cells in D1 and in D2 respectively. This naturally give rise to the subset matching information that we can leverage in our algorithm. We have added more discussion of this context to the text. \\n\\n2. The main advantage of our approach is that we learn a good cost function for the optimal transport, which could be quite different from the Euclidean cost. With the new cost function, we can then immediately compute transport maps between new data. This allows our approach to robustly generalize to new data. The baseline experiment that you suggested has the same performance as OT-baseline in the paper, as no side information about the test data is known to the algorithms. \\n\\n3. Thanks for this suggestion, we have added details about Algorithm 1 and how we backpropagate through the Sinkhorn-Knopp iterations in the revised text; see B.3 of the appendix.\\n\\n4. The complexity of Algorithm 1. Updating the weight by one forward and backward pass has complexity of O(n^2) up to logarithmic terms. OT-SI has the same complexity as the Procrustes-based OT methods which alternatively optimize over coupling matrix and linear transformation matrix (Page 4). In practice, taking the MNIST watering image alignment as an example, the running time for OT-SI and OT-Procrustes is 229.5s and 196.7s, respectively.\\n\\n5. The code is now in the folder.\"}", "{\"title\": \"Response\", \"comment\": \"Thank you for your careful review! Following your suggestion, we have added additional motivation for aligning datasets with subset correspondence. In particular, aligning datasets across different conditions is a very important problem in biology, especially in single cell analysis. This is highlighted in several recent high-profile papers, e.g. Butler et al Nature Biotechnology 2018 \\u201cIntegrating single-cell transcriptomic data across different conditions, technologies, and species\\u201d. In these experiments, there is typically insufficient data or strong batch effects within a single dataset, and how to align datasets in order to increase sample and remove batch effect is considered one of the most important open challenges in computational genomics [Haghverdi et al Nature Biotechnology 2018].\"}", "{\"title\": \"Response\", \"comment\": [\"Thank you for your review and your helpful suggestions!\", \"We have added a discussion of [1] and related works on semi-supervised metric learning to the text.\", \"This is an interesting idea to try to convert subset matches into random pairwise matches. It doesn\\u2019t work well in our experiments. This is likely because there are many ways to generate random pairs from matched subsets.\", \"For flips, OT-Procrustes performs better because it assumes a linear transformation between datasets and has no model mismatch, but OT-SI has a more complicated parametrization. If we increase the training data size from 200 to 1000 images per class, the accuracy of OT-SI increases from 57.5% to 80.0%, while the accuracy of OT-Procrustes remains as 83.5%.\", \"When the perturbation is not simply linear (e.g. the other three distribution types), OT-Procrustes can behave even worse than the baseline due to the lack of flexibility to handle nonlinear artifacts.\"]}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary:\\nThe paper presents a gradient-based method for learning the cost function for optimal transport, applied to dataset alignment. The algorithm is based on the Sinkhorn-Knopp iterative procedure for approximating the optimal transport plan. The current paper proves that the Sinkhorn transport plan with a parameterized cost function is infinitely differentiable. The cost function can thus be optimized by unrolling the Sinkhorn iteration for a fixed number of steps and then backpropagating through the iterative updates. The cost function is optimized via a loss function derived from side information, specifically subsets of the two datasets to be aligned that contain elements that should be matched in the optimal transport plan. Experiments on a variety of synthetic and real datasets show that the proposed method is often better than, and almost always competitive with, other modern approaches.\\n\\nI recommend that the paper be accepted. The paper presents a well-justified and seemingly novel approach to solving an important problem, and demonstrates empirically that it compares favorably to other approaches. The proposed technique seems to be a fairly small extension to existing work, but together with its analysis and experiments it is sufficiently novel.\\n\\nDetails / Questions:\\n* An interesting point of comparison would be methods for semi-supervised metric learning that are not necessarily tailored to the OT setting. E.g. [1] (picked fairly arbitrarily)\\n\\n* Is it possible to adapt algorithms that need pairwise matches to your subset matching setting by e.g. choosing pairs in the matching subsets at random and calling them matching pairs?\\n\\n* Do you have any intuitions for the pattern of results in Table 4? Why might flips be a difficult transforation for OT-SI to learn?\", \"trivia\": [\"Page 6 is blank when I print the paper. Not sure if anyone else has this problem.\"], \"references\": \"[1] Bilenko, M., Basu, S., & Mooney, R. J. (2004, July). Integrating constraints and metric learning in semi-supervised clustering. In Proceedings of the twenty-first international conference on Machine learning (p. 11). ACM.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": [\"# Summary\", \"This paper proposes a new way to learn the optimal transport (OT) cost between two datasets that can utilize subset correspondence information. The main idea is to use a neural network to define a transport cost between two examples instead of using the popular Euclidean distance and have a L2 regularizer that encourages the resulting transport plan to be small if two examples belong to the same subset. The results on synthetic datasets (2D, MNIST image) and real-world datasets (RNA sequence) show that the learned transport cost outperforms the Euclidean distance and Procrustes-based OT.\", \"# Originality\", \"This paper considers a new problem which is to learn OT given \\\"subset correspondence\\\" information. The prior work does not have a mechanism to utilize such information or requires more side information such as pair-wise matching.\", \"# Quality\", \"The idea of parameterizing the transport cost using a neural network is novel to my knowledge. The idea of regularizing the transport cost based on the subset correspondence information sounds sensible.\", \"The empirical result is quite comprehensive and convincing in that it includes from illustrative examples to real-world datasets (RNAs) and includes many relevant baselines.\", \"# Clarity\", \"The paper is well-organized and well-written.\", \"Since this paper considers a new problem (learning OT with subset correspondence data), it would be great to motivate and explain the problem better in the introduction. RNA-seq examples in the introduction were not easy to parse as I do not have any background on biology. It would be better if the paper presented more illustrative/intuitive examples ideally with a figure that describes the problem.\", \"\\\"an principled\\\" -> \\\"a principled\\\"\", \"# Significance\", \"This paper seems to present a neat and sensible idea for learning OT with neural networks given subset correspondence information. The problem and method are new, and the empirical results look comprehensive. One minor issue is that it is unclear how significant this new problem setup is (learning OT cost with subset correspondence). Although the paper presents real-world examples (RNAs), the paper could be stronger if it presented more real-world examples/results and emphasized that this is an important problem to solve.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The authors consider the alignment problem for multiple datasets with side information via entropic optimal transport (Sinkhorn). The authors formulate it as a transport cost learning in optimal transport framework with constraints giving by side information. Empirically, the authors illustrated the effectiveness of the proposed approach over state-of-the-art on several applications (e.g. single-cell RNA-seq, marriage-matching, MNIST with its perturbed one.\\n\\nThe paper show good results of the proposed method for dataset alignment on several applications. However, the motivation of the dataset alignment seems quite weak (e.g. it is unclear the advantage for jointly analysis when one merge datasets by alignment). The side information is not clearly easy to obtain in practice, since it simply enforces the constraints on the transportation polytope (i.e. how one knows that which subset of a given class C_A on dataset A is matched with a certain subset of another class C_B on dataset B, it seems easy to obtain the side information about matching over the whole class C_A with C_B, and it seems nontrivial to obtain the subset matching). Additionally, some important information about the proposed method is not in the main text but in the appendix which makes the paper hard to read without checking appendix constantly (especially the main proposed method). Overall, I lean on the negative side.\", \"below_are_some_of_my_concerns\": \"1) As stated above, I am curious how one can obtain the side information about the subset matching? For example, in Figure 1, it seems trivial to have a side information about the whole class matching and quite nontrivial to obtain a matching for subsets of any given class.\\n\\n2) The side information enforces the constraints on the feasible set of the transportation polytope. It seems to me that one can solve the variant of optimal transport (or Sinkhorn) with those extra constraints directly, it is unclear to me how this kind of side information can navigate the transport cost learning. Moreover, given a new pairs of input, the proposed method needs to optimize the transport cost again which make the benefits to learn to transport cost quite limited. Overall, a trivial baseline for the proposed method is to solve directly optimal transport with those extra constrains (from side information) directly (maybe with the Euclidean cost metric), but not the optimal transport without extra constraints as in the current setting. \\n\\n3) The authors may need to consider to reorganize the presentation. Much important information of the proposed method is not in the main text, but in the appendix which make the paper hard to follow without checking the appendix constantly. The derivation of gradient in line 6 of Algorithm 1 is not trivial. It is better if the authors give more detail information and discussion. I am quite confused about the idea to back propagate the \\\"gradient\\\" through Sinkhorn-Knopp iterations?\\n\\n4) It seems that the complexity of gradient update of Algorithm 1 is quite high. The authors should give a comparison about time consumption in the experiments (besides the accuracy matching)\\n\\n5) It seems that there is nothing in the code sharing folder?\"}" ] }

[ "Sung Whan Yoon", "Jun Seo", "Jaekyun Moon" ]

[ "few-shot learning", "meta-learning", "semi-supervised learning", "task-adaptive clustering", "task-adaptive projection space" ]

{ "note_id": [ "-5sy4JS1Q", "BJlxo1StjH", "r1gwDySYoS", "SyxxSkSYsr", "S1gq60EFjB", "BkgW8i5RKB", "Bkxb6KTpYH", "SJltUIRoYB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731436, 1573633944106, 1573633886563, 1573633848348, 1573633729871, 1571887945325, 1571834296788, 1571706449448 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1746/Authors" ], [ "ICLR.cc/2020/Conference/Paper1746/Authors" ], [ "ICLR.cc/2020/Conference/Paper1746/Authors" ], [ "ICLR.cc/2020/Conference/Paper1746/Authors" ], [ "ICLR.cc/2020/Conference/Paper1746/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1746/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1746/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes an approach to semi-supervised few-shot learning. In a discussion after the rebuttal phase, the reviewers were somewhat split on this paper, appreciating the advantages of the algorithm such as increased robustness to distractors and the ability to adapt with additional iterations, but were concerned that the contributions over Ren et al were not significant. Overall, the contributions of this paper don't quite warrant publication at ICLR.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Official Blind Review #2\", \"comment\": \"We agree that the proposed method is indeed a combination of TapNet and soft-clustering. However, we maintain that the proposed method is not an easily anticipated extension/combination of prior work. Note that the proposed iterative updates of projection space provide an increasing degree of task-conditioning, a newly introduced notion, which we believe is very significant. Ren et al. also attempt iterative updates of class prototypes, but there is no evidence of performance improvement with this process. In fact, our own evaluation indicates that the Ren et al. method does not provide performance gain with iterative clustering (see Figure 3 in our paper). Note that in Ren et al. (and also in the original Prototypical Networks) there is no explicit mechanism for task-specific conditioning, whereas in our method the projection space reflects task-conditioning (as it changes with a new set of tasks). In this sense, an iterative strengthening of projection space, as introduced in this paper, means a stronger task-conditioning.\\n \\nWe also would like to clarify that the performance gap between TAC and Soft-k-Means is not simply due to the performance advantage of the baseline TapNet. When soft-clustering is simply combined with TapNet (clustering in embedding space and a subsequent classification in projection space), 1-shot and 5-shot accuracies for miniImageNet are 52.82% and 68.76%, respectively. These numbers, when TAC is used, become 55.50% and 69.21%. For tieredImageNet, these numbers are 54.18% and 71.84% for the simple combination versus 58.48% and 72.05%, respectively, for the proposed TAC algorithm. In both datasets, while 5-shot performances show small improvements, 1-shot accuracy gains are substantial. Given the importance of fewer-shot results, we feel that the significance of the proposed method is high.\"}", "{\"title\": \"Response to Official Blind Review #`1 (2/2)\", \"comment\": \"D) We understand that Ren et al. also talks about repetitive clustering. But we wanted to emphasize that in their work multiple steps of clustering did not help (as confirmed in Figure 3 of our paper) whereas in our case they did. To us, what this means is that our method allows different levels of task-conditioning through a repetitive update of projection space. Since the formation of the projection space represents task-conditioning, we felt that our approach introduced a way to control the amount of task-conditioning.\\n\\nE) We now added a separate section on TapNet baseline with a better explanation in the revised version. This hopefully makes the contribution of this paper clearer. \\n\\nG) We have shortened the introduction as suggested.\"}", "{\"title\": \"Response to Official Blind Review #1 (1/2)\", \"comment\": \"A) This is an important question, and in answering it we first note that in any TapNet-based semi-supervised method, the projection space must always be updated before final classification is done if the refined class centroids are to be utilized. This is because in TapNet final classification is done by comparing the query with the stand-alone per-class reference vectors on the projection space (not with the prototypes themselves). Since the reference vectors do not change during meta-testing, the only way the refined prototypes could influence classification accuracy is through projection space update. This means that to employ the proposed method, there must be at least one projection space update right before final classification. Now having said that, we conducted an ablation study on what if the iterative clustering is done through repetitive soft-clustering on the initial projection space, with only one-time projection space update in the end before final classification (which is done using the per-class reference vectors now projected onto this new space). The resulting miniImageNet 1-shot and 5-shot accuracies without distractor are 55.61% and 69.05%, which are similar to the reported results conducting projection update with each new clustering. For the tieredImageNet, the 1-shot and 5-shot accuracies without distractor are 58.89% and 71.99%, which are also similar to the accuracies of TAC with projection update for each new clustering. We note that even with the one-time projection update, the projection space undergoes a significant change reflecting the refinements in cluster centroids. Our short answer to the question on whether the initially computed projection space already sufficiently separates the classes or not is that the final projection space (whether obtained through a series of updates or computed at once) provides a considerably better separation than the initial projection space, as seen from the reported results of accuracy versus iteration number (blue curves in Figures 3 and 8). We also note that we already provided an ablation study on doing iterative soft clustering in embedded space and then find a projection space at the end where the classification is done. The results in Figure 8 (red curves) show that this approach is significantly worse than our proposal. In summary, the projection is better be done in the beginning and then at least once more at the end of the iterative clustering process.\\n\\n \\nB) The reviewer is correct in that the additional centroid we introduce reflects even the samples belonging to valid classes. However, it is easy to see that said samples are actually closer to their correct centroids than this additional centroid, because the new centroid necessarily averages samples from all classes, including the distractors. In other words, an unlabeled sample belonging to some valid class will likely be viewed as a member of that cluster eventually, even though this sample influenced the computation of the additional centroid. In the work of Ren et al. the \\u201czero prototype\\u201d is added for handling distractor samples. Even the arbitrarily chosen zero-vector seems to work, although ours is slightly better. For 1-shot and 5-shot miniImageNet with distractor samples, when we initialize the prototype for distractor as a zero vector, the accuracies are 51.35% and 67.10%, while the presented 1-shot and 5-shot accuracies are slightly better at 51.56% and 67.75%, respectively.\\n \\n\\nC) The reviewer raises an interesting point. For miniImageNet, the suggested inference-only baseline accuracies are 53.19% for 1-shot and 69.53% for 5-shot, when the distractor is not considered. These compare to 55.50% for 1-shot and 69.21% for 5-shot in our reported results. Thus, for 1-shot, training consistently between meta-learning and meta-testing is somewhat better, while for 5-shot iterative clustering at meta-testing only gives a slightly better or similar result. On the other hand, for tieredImageNet, the inference-only baseline accuracies are 55.66% and 71.54% for 1-shot and 5-shot, when the distractor is not considered. These accuracies improve to 58.46% and 72.05% in our results. Thus, for the tieredImageNet case, training consistently between meta-learning and meta-testing seems to show definitely better results. With distractor samples, our TACdap consistently shows better results than the inference-only baselines on both miniImageNet and tieredImageNet experiments. We have included all these results for inference-only baselines in the uploaded version of paper (as TapNet + Semi-Supervised Inference in Tables 1 and 2).\"}", "{\"title\": \"Response to Official Blind Review #3\", \"comment\": \"Unfortunately, we must disagree with the reviewer in that our abstract and introduction very clearly state the goal, motivation and differentiating features of the proposed method. In particular, we make it clear that our method employs explicit task-conditioning in which unlabeled sample clustering for the current task takes place in a new projection space different from the embedding space. We also state that our method introduces a concept of controlling the degree of task-conditioning, the process in which the amount of task-conditioning varies with the number of updates for the projection space. We are not really giving the sketch of the algorithm but rather the key idea and the gist of the method. We do admit that there is a sentence in the abstract about soft labeling that could be viewed as a bit descriptive of the algorithm. Likewise, the second to the last paragraph of the introduction could be shortened. We have made these changes in our newly uploaded revision. Also, to help clarify the terminologies that have come from the work of TapNet, we have now provided a separate section on the background of TapNet. As for the impact of the proposed work, our message has been consistent in that iterative task-conditioning in utilizing unlabeled samples is a new notion that is significant and by doing so one could achieve state-of-the-art semi-supervised few-shot learning accuracies.\\n\\nWe do not agree that our proposed algorithm is simply adding soft labeling to TapNet. The proposed repetitive updates of projection space provide an increasing degree of task-conditioning, which is a new concept, as already emphasized above. This argument is strengthened by the fact that the performance of the proposed method exceeds well beyond the simple combination of soft labeling and TapNet, especially for fewer-shot experiments. For example, for 1-shot miniImageNet, the simple combination gives a 52.82% accuracy, which improves to 55.50% with the proposed TAC. For 1-shot tieredImageNet, the improvement is from 54.18% to 58.48%.\\n\\nClustering is a term not specific to our work but is rather a basic concept for semi-supervised learning. Classification space is also general; it should naturally mean a space where classification takes place. We have never used a term \\u201cper class network\\u201d; we used \\u201cper-class network output average\\u201d, which is a concept discussed in the well-known Prototypical Networks of Snell et al., 2017. As mentioned in our introduction, it is the per-class averages of the features that act as prototypes for representing given tasks.\\n \\nAs for the construction/description of M, the reviewer seems to be under the impression that (i) and (ii) are two separate objectives in finding M. This is not true; as stated in the paper, any M that satisfies epsilon_n x M = 0 for all n does contain columns that span the task-adaptive classification space, unless of course M is all zero. Perhaps the reviewer is worrying about the trivial case of M being all zero. While this should be obvious, we have made it explicitly clear in the revised paper that we are barring the all-zero M when finding M.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Paper Summary:\\n\\nThis paper extends TapNet, a few short learning approach, to the setting of semi-supervised few shot learning. For that, it borrows the idea of soft-labeling from Ren et al 2018, a semi-supervised few shot learning algorithm. Experiments over mini and tiered imagenet compare the approach with alternatives.\", \"review_summary\": \"The paper could be clearer. The main idea of the paper and its goal/motivation is not concisely given in the intro and abstract. The paper refers to a lot of terminology from TapNet without defining them which makes it difficult to understand. I am not an expert in the field and I have a hard time judging whether the empirical improvement are exceptional but I would advocate against accepting the paper based on clarity alone at this point. Moreover the contribution -- adding soft labeling to TapNet -- seems modest.\", \"detailed_review\": \"The abstract should not be a sketch of the proposed algorithm. It should highlight: what is the problem, what is the key idea of what you propose to address it (one sentence), why is it original and better than prior work, what is the foreseen impact of this work.\", \"the_terminology_specific_to_your_problem_and_family_of_model_need_to_be_defined\": \"what is the clustering space? which clustering? what is the classification space? what is a per class network? None of these concepts are defined.\\n\\nIn section 2.2, please rephrase \\\"Nulling of the errors is essentially finding a projection matrix M such that epsilonM = 0 for all n. The resulting M is a matrix whose columns span the task-adaptive classification space\\\" as \\\"We identify M such that (i) epsilonM = 0 for all n and (ii) M columns span the task-adaptive classification space\\\". The text make it sounds like your are finding the solution of (i) which happened to also verify (ii) which is not true. For instance, M=0 satistifies (i) but not (ii).\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary\\n========\\nThis paper tackles the more realistic variant of few-shot classification where a large portion of the available data is unlabeled, both at meta-training time in order to meta-learn a learner capable of fast adaptation, as well as meta-test time for adapting said learner to each individual held-out task. \\n\\nTheir approach is based on TapNet, a model which constructs a task-specific projection based on the support set of each episode, and then classifies each query according to its distance to each class prototype in the projected space. The projection is computed so that the class prototypes of the episode (averaged support examples) are well-aligned with a set of meta-learned references. Intuitively, those references learn to be far from each other, so that aligning the prototypes with them leads to a space where the episode\\u2019s classes are well separated, allowing for easier classification between them.\", \"they_then_extend_this_model_to_incorporate_unlabeled_examples_by_performing_task_projection_as_follows\": \"1) the projection is computed so as to align the initial prototypes (computed only using labeled examples) to the meta-learned references. 2) In that projected space, each unlabeled example is assigned a predicted label based on its proximity to the projected prototypes. 3) Then, back at the original space, those predicted labels of the unlabeled examples are used to refine the class prototypes (weighted average as in Ren et al). 4) The projected space so that the *refined* prototypes are best aligned with the meta-learned references. 5) Possibly repeat 2-4 (at meta-test time).\\n\\nExperimentally, they outperform recent approaches to semi-supervied few-shot classification on standard benchmarks, though not by far. Perhaps more interestingly, their performance degrades less than that of Ren et al as the distractor ratio increases, and they show that their method benefits from additional steps of task adaptation, whereas that of Ren et al reaches its performance limits after the first step of soft clustering.\\n\\nHigh-level comments\\n==================\\nA) Ablation: An interesting ablation would be: instead of going back and forth between the embedding space and the projected space, the task adaptation happens only in the initially-computed projection space (i.e. the one computed based on the labeled data only). This would amount to: computing the projection space, and then performing a few steps of soft clustering, similar to Ren et al. in that space. This would help determine how beneficial it is to re-compute that projection space according to the current \\u2018best guess\\u2019 of where the class prototypes lie at each iteration. The way I see it, it is an empirical question whether the initially computed projection space already sufficiently separates the classes or not. I assume this would also lead to a more computationally efficient solution?\\n\\nB) Handling distractors: In the case of distractors, they use an additional centroid (and a corresponding additional reference vector) for the purpose of \\u2018capturing\\u2019 the unlabeled examples that don\\u2019t belong to candidate classes. I find the initialization of this strange: this additional centroid is computed as the mean of all unlabeled examples, and the initial projection construction is influenced by a term that matches this centroid to a corresponding reference. This would mean, however, that even the unlabeled examples which do indeed belong to one of the candidate classes end up far from those classes in the projected space, in order to be close to the designated extra reference in that space. We know that this is not ideal, since we assume that some unlabeled examples do belong to the same classes as the labeled ones. Is there a way to quantify how severely this affects the quality of this initial projection? I would also be curious about the meta-learned location of the extra reference. Does it end up being roughly in the center of the references corresponding to the labeled classes?\\n\\nC) Inference-only baselines. Ren et al. experimented with inference-only baselines: meta-learning happens only using the labeled subset, and the proposed clustering approach only takes place at meta-test time. In this case this would amount to meta-training a standard TapNet and then performing the proposed refinement only in test episodes. This is interesting as it allows to understand the importance of learning an embedding end-to-end for being more appropriate for unlabeled example refinement. It is not obvious that this is required, so I would be curious to see these results too. (This differs from the reported TapNet baseline in that at meta-test time it would make use of the proposed semi-supervised refinement).\\n\\nClarity / quality of presentation:\\n============================\\nD) A lot of emphasis is placed on the ability of the proposed method to control the degree of task conditioning. I would like to emphasize that this is not something that previous methods lack. Ren et al.\\u2019s approach could also perform multiple steps of clustering for example. Whether or not this is beneficial is an empirical question, but I wouldn\\u2019t say that the proposed method does something special to \\u201ccontrol\\u201d how much adaptation is performed.\\n\\nE) It would have been useful to have a separate section or subsection that explains TapNet, since this is the model that the proposed method is built upon. Instead, the authors blend the explanation of TapNet in the same section as their method which makes it hard to understand the approach and to separate the contribution of this method from TapNet.\\n\\nG) The length of the paper is over the recommended pages, and I did feel that a few parts were too lengthy or repetitive (e.g. the last paragraph of the introduction described the model in detail. I felt that it would have been more appropriate to give the higher level intuition there and leave the detailed description for the next section).\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper presents a semi-supervised few-shot learning method that combines the TapNet architecture with the soft-clustering approach in the original semi-supervised few-shot learning method of Ren et al. 2018. Results are reported on mini-ImageNet and tiered-ImageNet, demonstrating superior results to Ren et al. 2018 and some more recent work.\\n\\nThough the results are strong, I'm personally leaning towards rejecting this submission. The main reason is that the contribution is an arguably simple combination of 2 methods (TapNet + Ren et al. 2018). Specifically, I notice that the difference between the performance of the TapNet and Prototypical Networks baselines (6%, 7.5% for 1 and 5 shot on miniImageNet, and 5% and 3% for 1 and 5 shot on tieredImageNet) is roughly also found in the difference between TAC and Soft- k-Means (Ren et al. 2018). This suggests that most of the benefits found in TAC can simply be explained by the adoption of the TapNet architecture. And I consider this proposed extension to semi-supervised few-shot learning of TapNet to be pretty straightforward.\\n\\nI don't really expect the authors can improve on the above point in a revision, so I doubt I'll be convinced to change my score. However, if the authors believe that I've missed a subtlety of their method that is less technically straightforward than my analysis suggests, of course I'd like to know.\"}" ] }

[ "Geoff Pleiss", "Tianyi Zhang", "Ethan R. Elenberg", "Kilian Q. Weinberger" ]

[ "Deep learning", "noisy data", "robust training" ]

{ "note_id": [ "yDQYOxjhoR", "BJerJ9yDoH", "HylRqYJvoH", "SyeJTO1vjB", "r1e5gNqptr", "S1eybeV2YH", "B1g0BwJdKS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731406, 1573480925307, 1573480854322, 1573480631396, 1571820530473, 1571729399091, 1571448646241 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1745/Authors" ], [ "ICLR.cc/2020/Conference/Paper1745/Authors" ], [ "ICLR.cc/2020/Conference/Paper1745/Authors" ], [ "ICLR.cc/2020/Conference/Paper1745/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1745/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1745/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper proposes a new, stable metric, called Area Under Loss curve (AUL) to recognize mislabeled samples in a dataset due to the different behavior of their loss function over time. The paper build on earlier observations (e.g. by Shen & Sanghavi) to propose this new metric as a concrete solution to the mislabeling problem.\\n\\nAlthough the reviewers remarked that this is an interesting approach for a relevant problems, they expressed several concerns regarding this paper. Two of them are whether the hardness of a sample would also result in high AUL scores, and another whether the results hold up under realistic mislabelings, rather than artificial label swapping / replacing. The authors did anecdotally suggest that neither of these effects has a major impact on the results. Still, I think a precise analysis of these effects would be critically important to have in the paper. Especially since there might be a complex interaction between the 'hardness' of samples and mislabelings (an MNIST 1 that looks like a 7 might be sooner mislabeled than a 1 that doesn't look like a 7). The authors show some examples of 'real' mislabeled sentences recognized by the model but it is still unclear whether downweighting these helped final test set performance in this case. \\n\\nBecause of these issues, I cannot recommend acceptance of the paper in its current state. However, based on the identified relevance of the problem tackled and the identified potential for significant impact I do think this could be a great paper in a next iteration.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response from authors\", \"comment\": \"W.R.T. additional benchmarks and datasets: We have since performed additional experiments on TinyImagenet. A ResNet-32 trained on 95000 training samples (5000 withheld for validation) receives 49.4% test error on the clean dataset. With 40% label noise this number drops to 65.5%. AUL reweighting achieves 55.7% error. We will update the paper with these results. Additionally, see the response to Reviewer #2 regarding real-world datasets.\\n\\nW.R.T. \\u201chard\\u201d clean examples: please see response to Reviewer #2.\\n\\nW.R.T. the MoG: we use the initial network to compute the AUL for all 45,000 training points. Then we fit the MoG to these 45,000 AULs.\\n\\nW.R.T the computation cost: thank you for noting this omission. It is roughly 1.5x as much as normal training. The initial AUL estimate requires training a network for 0.5x the usual training budget (right before the learning rate is dropped for the first time) Fitting the MoG is negligible compared to training the network.\"}", "{\"title\": \"Response from authors\", \"comment\": \"Thank you for your review. We hope that we are able to clarify some misunderstanding and address your comments.\\n\\nW.R.T. \\u201chard\\u201d clean examples: From our observations, it appears that \\u201chard\\u201d examples have higher AUL than most clean examples, but lower AUL than mislabeled examples. (Intuitively, this makes sense, because truly mislabeled examples will be the hardest for the network to generalize to.) Therefore, hard samples will be downweighted some, but not as much as noisy examples. We would note that curriculum learning (e.g. Bengio et al, ICML 2019; Hacohen and Weinshall, ICML 2019) similarly down-weights (rather than up-weights) difficult examples until the end of training. To some extent AUL-based downweighting will have a similar effect: the network will not reduce the loss of difficult examples until the high-weight easy examples have sufficiently low loss.\\n\\nW.R.T. other noise models: Please refer to our \\u201cpair flip\\u201d experiments in Figure 7 and Table 1. These experiments are exactly the noise model you propose (p probability of being correct, and (1-p) probability of being k). We note that our proposed method is adept at estimating and correcting for label error in this setting as well.\\n\\nW.R.T. presenting AUL distributions: we observed the bimodal behavior for both CIFAR10 and CIFAR100 on all experiments that we performed - regardless of noise model, noise percentage, or network architecture. We will update the paper with additional plots that demonstrate this bimodality.\\n\\nW.R.T. real-world datasets: To demonstrate our method\\u2019s applicability on real-world (non-image) datasets, we trained a model on a Machine Translation dataset (IWSLT14 German-to-English dataset). AUL determines that roughly 25 training samples have an incorrect target (the English sentence in the translation pair). Typically, these targets are not English and therefore are mislabeled. For example:\", \"source\": \"google glass : aroi mann 9 : mmm , aroi .\", \"target\": \"google glass : \\u0e2d\\u0e23 \\u0e48 \\u0e2d\\u0e22man 9 : mmm , \\u0e2d\\u0e23 \\u0e48 \\u0e2d\\u0e22 .\"}", "{\"title\": \"Response from authors\", \"comment\": \"Thank you for your review. We answer some of your questions below. (Additionally, thank you for catching typos and the NeurIPS 2019 citation error. We will correct these.)\\n\\nW.R.T. existing work: We will ensure that appropriate credit is given to Shen & Sanghavi in the revised version. We would note that our contribution goes beyond identifying the differences in training loss (also noted by e.g. Zhang et al., ICLR 2017), but introducing a stable metric (AUL) that is able to transform this observation into a robust method.\\n\\nW.R.T. normalizing by c: This is a common practice for several re-weighting methods (e.g. Ren et al., ICML 2018). We found that it improves the stability of training as it ensures that the effective learning rate is the same for all datasets regardless of label noise. We would note that in practice it would be incredibly rare for a minibatch to contain only bad examples. Even with 80% label noise, the probability of an all bad minibatch of size 256 is roughly 10^-25.\\n\\nW.R.T. Pearson\\u2019s vs Spearman\\u2019s: Spearman\\u2019s produces similar results (e.g. ~98% correlation for AUL across networks). We chose to measure the Pearson\\u2019s correlation because\\u2014-as noted in Figure 4\\u2014-the AUL scores from each network fall within very similar ranges. Therefore we expect correlation among the values and not just their rankings. We will revise the language in the Figure 5 caption which may have caused confusion.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a new method of detecting label noise through calculating the Area Under the Loss statistic, which is based on the learning trajectory.\\n\\nI believe this paper is NOT the first one to point out the different training behavior of clean and noisy samples in the label noise problem. While the author mentioned that Shen & Sanghavi used the training losses for selecting data, Shen & Sanghavi also observed the training loss for good and bad samples are different across time. Therefore, it would be good to give correct credit to previous work, and not over-emphasizing the contribution. \\n\\nPlease proofread the paper, and correct the typos. For example, on page 3, there are 2 typos in the last three lines in the paragraph \\u201crate of loss reduction indicates memorization need\\u201d. \\n\\nCan you explain why you normalize the loss of each batch by c ? If z_i s are all small within a batch, this loss function does not reweight the bad samples in the correct way.\\n\\nCan you explain why you use Pearson\\u2019s correlation instead of Spearman\\u2019s correlation to evaluate the consistency of the ranking? This sounds surprising to me. \\n\\nAlso, the cited paper \\u201cUnderstanding generalization of deep neural networks trained with noisy labels. \\u201c by Hu et al. is not in NeurIPS 2019. Please correct this citation error if possible.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"In this paper, the authors proposed methodologies to identify mislabeled training examples. Specifically, the authors proposed to examine the training loss of each example to identify examples with high training losses. To mitigate the randomness in the training loss, the authors proposed an AUL (area under the loss) metric, which is less noisy that directly examining the loss. The authors proposed to use Gaussian mixture to approximate the AUL distribution, and then estimate the probability of mislabeling. The authors proposed to downweight training examples with high likelihood of being mislabeled. The proposal is examined in numerical studies.\\n\\nWhile I agree the method has a good potential to be significant, in its current form, I have several concerns that prevent me from recommending acceptance.\\n1. How can we be sure that examples with large training loss are mislabeled examples, rather than examples that are inherently difficult to classify? For example, if the dataset consists of 99% bright images and 1% dark images, then it could be possible that dark image training examples have a higher loss. This question becomes more relevant when the model does not have sufficient complexity, or does not have the correct form of complexity.\\n2. If we cannot distinguish mislabeled examples with examples that are hard to classify, then it seems to me that a simple downweighting based on AUL could potential lead to the opposite outcome. Intuitively, we should upweight those difficult to classify examples to improve generalizability.\\n3. It is somewhat surprising to me that the AUL distribution is a nicely looking bimodal shape. Could it be due to the fact that the wrong labels are randomly generated with equal probabilities from other classes? What it we have a different error distribution, e.g., the label has p probability of being correct, and (1-p) probability of being k, where k is a constant?\\n4. Given the above concerns, I feel the numerical studies are insufficient. The authors may consider 1) including more noise distributions, 2) presenting the AUL distributions to check whether they are truly bimodal, and 3) trying more datasets and more model architectures.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"Pros:\\n1. This work suggests a surprisingly elegant approach to identify data with corrupted labels. The proposed AUL (area under the loss curve) makes intuitive sense, and exploiting the fact that noisy examples are hard to learn at the early stage of training, leading to higher AULs than clean examples.\\n2. The paper is well-written, with plenty of visualization to aid the understanding.\\n3. Experimental results suggest advantage of the proposal. Evaluation seem to suggest the using AULs is more robust against large corruptions compared with SOTA.\", \"cons\": \"1. The biggest concern is the generalizability of AULs to datasets that are more challenging than CIFAR10 and CIFAR100. This has been the concern not only for this work, but also for this general research area. For instance, on Tiny-ImageNet dataset (e.g., used in SELFIE: Refurbishing Unclean Samples for Robust Deep Learning (ICML19), which could btw be used in the benchmark), a clean dataset only gives ~55% test error, compared with about ~10% on CIFAR10 and ~30% on CIFAR100. The robustness of AULs needs to be tested on those harder problems as well. \\n2. AULs are based on the assumption of a clear separation of two types of examples -- clean and noisy. It ignores the difficulty of training examples. This raises the question of where to place hard-clean examples between those two modes.\\n3. It usually helps to include a real-world noisy datasets where its noisy corruption is unknown. One would be curious to know the noisy level estimate.\", \"questions\": \"1. It seems from the experiments that MoGs are fit per training example in order to compute weights for the cost?\\n2. The computational cost of AUL-based training is not mentioned. Is it 2x, 3x, considering there are distinct stages of training before and after AUL estimation?\"}" ] }

[ "Lichao Sun", "Yingbo Zhou", "Jia Li", "Richard Socher", "Philip S. Yu", "Caiming Xiong" ]

[ "sensitive data", "teacher", "student", "private deep learning", "teacher ensembles", "model", "information", "privacy", "paramount importance" ]

{ "note_id": [ "SOz567xbqI", "rylLJjxMoH", "B1gTzqlMjB", "Hkli6KxzoH", "SyeKgE2JqS", "H1eUlfF_tH", "HkxrciCwtB", "BJgZqAHoPS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "comment" ], "note_created": [ 1576798731376, 1573157598336, 1573157397202, 1573157314765, 1571959792557, 1571488238425, 1571445644889, 1569574537382 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1743/Authors" ], [ "ICLR.cc/2020/Conference/Paper1743/Authors" ], [ "ICLR.cc/2020/Conference/Paper1743/Authors" ], [ "ICLR.cc/2020/Conference/Paper1743/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1743/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1743/AnonReviewer3" ], [ "~Ricardo_Silva_Carvalho1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper presents a differentially private mechanism, called Noisy ArgMax, for privately aggregating predictions from several teacher models. There is a consensus in the discussion that the technique of adding a large constant to the largest vote breaks differential privacy. Given this technical flaw, the paper cannot be accepted.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Both PATE and our work are data-dependent approach with differential privacy guarantee\", \"comment\": \"Thank you for reviewing and comments. The following are the reply to some of your comments.\\n\\nOur work closely follows PATE uses the sensitive data D and partition it into n sub-dataset. Then individual teachers are trained on each data partition. After a student query sample x, then all teachers can return a voting result with the noise perturbation for privacy concerns. It is clear to see that due to the partition policy and uses the partition sub-dataset to train teachers, PATE is a data-dependent approach. Mathematically, PATE has a sensitive data D as input, and need a voting vector output V, the data transformer function f is f_{D, x}(D) -> V, where f is dependent on both sensitive data and student query, then it can transform the sensitive data D to voting vector V. In the whole process of PATE, for each student sample x, the data transformer function f_{D, x} is different, then PATE uses the composition theorem to estimate the total privacy loss of multiple mechanisms. \\n\\nWe follow each step as in PATE, and the difference here is that we use a different function f which adds a constant to the max. This can potentially make the sensitivity among different mechanisms different. However, for each mechanism (i.e. given corresponding query x and dataset D), the sensitivity is fixed and will not be changed. Just like PATE need to know the partition information to bound and estimate the sensitivity of each mechanism, we also need to estimate the sensitivity of each of our mechanisms based on the dataset and the output of the data transformer f_{D, x}. Finally, also use the composition theorem to estimate the total privacy loss of multiple mechanisms. As such, we think our proposed approach is technically correct.\\n\\nPlease let us know if you have any further questions and concerns. Thanks a lot again.\"}", "{\"title\": \"Individual Sensitivity Analysis\", \"comment\": \"Thank you for reviewing and comments. The following are the reply to some of your comments.\\n\\nIn fact, we mainly follow each step as in PATE. In PATE, for each student sample x, it will create an individual mechanism, and then multiple student data samples lead to multiple individual mechanisms. The difference here is that we propose to use each function f_{D, x} which adds a constant to the max, where D is the sensitive dataset. This can potentially make the sensitivity among different mechanisms different. Then, based on different sensitivities of different mechanisms, we add the noise based on each sensitivity of each individual function.\\n\\nSince both PATE and our work are data-dependent privacy analysis for each mechanism. In this case, what you mentioned \\\"If c>0, then one sample point can change the count by 1+c\\\" would not be for all mechanisms. Most mechanisms (with different sample points) can still hold the fact that at most change the count by 1 with adding a constant c, instead of c + 1. \\n\\nLastly, we estimate the total privacy budget by using the composition theorem. As such, we think our proposed approach is technically correct. Please let us know if you have any further questions and concerns. Thanks a lot again.\"}", "{\"title\": \"The final composition is also differential private.\", \"comment\": \"Thank you for reviewing and comments. The following are the reply to some of your comments.\\n\\nYes, our composition theorem is also differential private (DP). In more detail, our work closely follows PATE uses the sensitive data D and partition it into n sub-dataset. Then individual teachers are then trained on each data partition. After a student query sample x, all teachers can return a voting result with the noise perturbation for privacy concerns. Same as in PATE, for each query sample from the student, there is a particular mechanism, and we use the composition theorem to estimate the total privacy loss of multiple mechanisms. It still satisfies the DP requirements. As such, we think our proposed approach is technically correct. Please let us know if you have any further questions and concerns. Thanks a lot again.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"To improve the privacy-utility tradeoff, this manuscript proposes a voting mechanism used in a teacher-student model, where there is an ensemble of teachers, from which the student can get gradient information for utility improvement. The main idea of the proposed approach is to add a constant C to the maximum count collected from the ensemble, and then noise is furthermore added to the new counts. I can understand that by adding the large constant C, the identity of the maximum count could be preserved with high probability, leading to a better utility on the student side. However, equivalently, this could also be understood as that the noise is not added uniformly across all the counts, but instead a relatively smaller noise is added to the maximum count. Hence it is not clear to me whether the final composition will still be differentially private?\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper studies the teacher ensembles setting for differentially private learning. In this setting, each teacher holds part of the training set and trains a local model. The student uses unlabeled examples to query teacher model. Then the student trains a model from scratch using the examples labeled by teachers.\\n\\nIn order to make the labeling process differentially private, previous work uses noisy argmax mechanism. Each class of label is assigned with a count number. The student first queries the same example to multiple teachers. To guarantee differential privacy, the counts are perturbed by noise before releasing. Then, because of the post-processing property of differential privacy, the argmax operator on such noisy counts are still differentially private.\\n\\nThis paper proposes to add a constant c to the largest count before perturbing and releasing the counts. The authors argue this would improve the accuracy of the noisy argmax operator and yield the same privacy loss as previous approach. However, adding a constant c would increase the sensitivity and therefore degenerates the privacy guarantee. The added noise cannot guarantee the privacy if all others are the same as previous work. To see this clearer, for example, if c=0, then one sample point can at most change the count by 1. If c>0, then one sample point can change the count by 1+c. Because of this, the proposed method cannot guarantee the amount of differential privacy as the paper claimed.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper proposes an improvement on the PATE framework for achieving near-zero privacy cost, and showed privacy analyses and experimental evaluations of the proposed method.\\n\\nThe proposed method can be technically flawed. Adding a consent to the max will not guarantee privacy unless you account for the privacy cost for testing whether the distance of f(D) is larger than 2. This is because the distance of f(D) is data-dependent and revealing it violates privacy. Since the whole privacy analysis of PATE is based on the privacy guarantee of the Noisy ArgMax, the epsilon calculated here is voided.\"}", "{\"comment\": \"Hello,\\n\\nI believe that adding a big constant c depending on who has the largest count does not protect the information of who is the item with largest count, as going from D to D' one would first change the counts (affecting privacy) and only then add the constant. Basically, if on D a given item is the one with the largest count, you're adding the constant to make sure (or with very high prob) such item remains the largest count even after noise addition. That's basically just selecting the item with largest count. And even doing that for a gap > 2 between first/second is not DP.\\n\\nHaving \\\"largest count\\\" minus \\\"second largest count\\\" greater than 2 is not sufficient to satisfy DP, as this property itself is data dependent. The way to use this gap usually is to test privately that D is far from a dataset where the item with largest count is not the first anymore. This is normally done using Propose-Test-Release.\\nI suggest looking at the papers \\\"Differentially Private Feature Selection via Stability Arguments, and the Robustness of the Lasso\\\" and \\\"Model-Agnostic Private Learning\\\" to see more about this.\\n\\nSo in essence, the way the mechanism is working is revealing information about the item with largest count on D, so not really differentially private.\", \"title\": \"Not really differentially private\"}" ] }

[ "Geoff Pleiss", "Amauri Souza", "Joseph Kim", "Boyi Li", "Kilian Q. Weinberger" ]

[ "Out-of-distribution", "deep learning", "regression" ]

{ "note_id": [ "lQWvXUknDH", "BklMk3EojH", "HylFsPxioS", "S1xquvlooH", "S1ll8DxijB", "SyxnlDgojH", "BJxg_FwwqB", "B1xBgmDPcH", "ryl4xqCRKH", "SygZYijTFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731347, 1573764057788, 1573746592699, 1573746545935, 1573746503664, 1573746420126, 1572464999542, 1572463340646, 1571903980359, 1571826552745 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1742/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1742/Authors" ], [ "ICLR.cc/2020/Conference/Paper1742/Authors" ], [ "ICLR.cc/2020/Conference/Paper1742/Authors" ], [ "ICLR.cc/2020/Conference/Paper1742/Authors" ], [ "ICLR.cc/2020/Conference/Paper1742/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1742/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1742/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1742/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper investigates out-of-distribution detection for regression tasks.\\n\\nThe reviewers raised several concerns about novelty of the method relative to existing methods, motivation & theoretical justification and clarity of the presentation (in particular, the discussion around regression vs classification). \\n\\nI encourage the authors to revise the draft based on the reviewers\\u2019 feedback and resubmit to a different venue.\", \"title\": \"Paper Decision\"}", "{\"title\": \"d-dimensional features\", \"comment\": \"Thank you for answering my concerns.\\n\\nOf course information is most likely discarded when mapping from R^d->R, but this does not tell us much about the intrinsic dimensionality of the last layers output before this is happening. And this is the space in which OOD detection is carried out if I understand correctly.\\n\\nIt is possible to build invertible deep networks that preserve all information up until the final d-dimensional output and only then (when mapping from R^d->R), information is discarded. Wouldn't your assumptions break down in such a case?\"}", "{\"title\": \"Author response to review\", \"comment\": \"W.R.T. the generality of our method: We believe that our method is sufficiently general, as we find it successful on small CNNs (Rotated MNIST, Section 3) and larger ResNets/DenseNets (KY Houses and Adult Age Estimation, Section 6). We agree that it is possible to construct \\u201cmodels with very few final features where your method would not work.\\u201d Indeed, the fact that simple generative models are sufficient for OOD detection relies on the curse of dimensionality, as it is unlikely that in-distribution and OOD features occupy the same parts of a high-dimensional space. Nevertheless, most MLPs and CNNs that are commonly used in practice have 100-1000 dimensional activations, which we believe is more than sufficient for our method to work.\\n\\nW.R.T. application to classification problems: please see our response to Reviewer #1.\\n\\nW.R.T. table abbreviations: Thank you for catching that the caption is missing abbreviations; we will fix this. TNR @ TPR 95% stands for \\u201ctrue negative rate when using an OOD threshold that results in 95% true positive rate of detection.\\u201d AUROC stands for \\u201carea under receiver operating characteristic.\\u201d\\n\\nW.R.T. computing an OOD score: this score can be computed using the NLL of the GMM. We will clarify this point.\"}", "{\"title\": \"Author response to review\", \"comment\": \"W.R.T novelty of our method: we respectfully disagree that the originally our method is limited. When generative models have been applied to OOD detection, they are typically used to fit the input data distribution. In our work, we demonstrate that it is possible to achieve good OOD performance not by modelling input but instead by modelling hidden activations. We believe that this is an important contribution because input-based modelling methods cannot be applied to high-dimensional images (to the best of our knowledge), whereas our method empirically demonstrates great success on these problems.\\n\\nW.R.T. the \\u201cdiscarded information\\u201d from activations: We believe that there may have been some confusion with regards to where our method is applied. Our method ONLY looks at the activations of the last hidden layer (not the activations of other hidden layers). Information is always discarded from this layer, as these d-dimensional features are then mapped to a scalar prediction. Indeed, we specifically do not examine the activations of earlier layers because -- as you point out -- information may not be discarded and activation distributions are less likely to be intrinsically low dimensional.\\n\\nW.R.T. the work of Nalisnick et al. 2019: we apologize if we misstated the findings of this paper. We will revise the sentence in Section 2 that describes this work.\"}", "{\"title\": \"Author response to review\", \"comment\": \"Thank you for your review and for making us aware of the paper by Ren et al. We hope that we are able to clarify some misunderstandings and justify our decisions.\\n\\nW.R.T. claims in section 4: we believe that our language in this section may have caused some confusion. The purpose of this section was to empirically demonstrate that the final layer activations occupy a space with extremely low intrinsic dimensionality. The goal of Figure 3 and the accompanying text around it is to show that this low-dimensionality is correlated with (not necessarily caused by) increased depth and regulation. We will revise the language to clarify this.\", \"the_purpose_of_section_4_is_to_empirically_demonstrate_two_phenomena_that_influence_our_method\": \"1) the distribution of in-distribution activations is intrinsically low-dimensional---making it possible to fit with simple generative models---and 2) OOD activations do not lie within the same distribution. We will re-arrange this section to clarify these points.\\n\\nW.R.T. input-space generative models as baselines: while this is a natural baseline that could be used for the toy example, we don\\u2019t believe that it is possible to apply generative models directly to the KY Houses or Adult Age Estimation datasets, as the inputs are 224 x 224 images. To the best of our knowledge, input space OOD detection has not been applied to images that are larger than 32 x 32. \\n\\nW.R.T. classification OOD methods not generalizing to regression: we will improve the discussion of this in Section 2. As far as we are aware, the only classification method that generalizes is the baseline of Hendrycks and Gimpel (ICLR 2017) which flags inputs as OOD if the softmax confidence falls below a threshold. This has a natural analog to heteroscedastic regression (the var method in our experiments). However, most other classification methods make use of the entire softmax distribution. For example, the ODIN method (Liang et al., ICLR 2018) applies temperature scaling to the output logits.\\n\\nW.R.T. experiments on classification OOD: our method can indeed be applied to classification OOD. For example, with a DenseNet-40 trained on CIFAR-10, our method can be used to detect OOD inputs from Tiny-Imagenet with an AUROC of 97% - which is comparable to many classification-specific methods. Regression was the primary focus of our work because, as we explain, there are not as many existing OOD-detection methods for this domain. Nevertheless, we can include additional classification results in the supplementary material.\\n\\nW.R.T. fine-tuning ImageNet networks: we fine-tune the network and the top layer, as this is common practice for datasets with images that are larger than 224 x 224.\"}", "{\"title\": \"Author response to review\", \"comment\": \"W.R.T. novelty: Thank you for making us aware of Aigrain et al.\\u2019s work. However, we strongly disagree that our proposed method is \\u201cessentially the same thing.\\u201d Aigrain et al.\\u2019s method requires 1) generating adversarial inputs and artificial OOD inputs and 2) trains a neural network (a supervised method) on the logits to identify OOD inputs. Our method on the other hand 1) does not require any OOD inputs at training time (real or artificial) and 2) models the distribution of in-distribution features with a simple Gaussian Mixture Model. We believe that this makes our method substantially different and a novel contribution.\\n\\nW.R.T. the performance of heteroskedastic regression and ensembles: you note that these baseline methods \\u201cwork very well in several cases.\\u201d Our experiments indeed confirm this: the \\u201cvar\\u201d (heteroskedastic) model and the ensemble both achieve an AUROC of >90% on most of our experiments, making these methods strong baselines and not \\u201cpoor\\u201d performers. At the same time, we do not find it surprising that our proposed method outperforms them. The predicted variance from a heteroskedastic model (which we did in fact parameterize logarithmically) is a reduction of the feature space down to a single dimension. Likewise, an ensemble reduces predictions down to a handful of dimensions (one for each model in the ensemble). These reductions would be prone to the curse of dimensionality argument that we outline in Section 3. Our proposed method on the other hand uses all >1000 dimensions of the feature space to determine if inputs are OOD.\\n\\nW.R.T the datasets: We chose to focus on image datasets which are the most common datasets for OOD and domain adaptation datasets. This is likely because it is easy to generate images which are \\u201cout of distribution\\u201d (i.e. use images from an unrelated dataset), whereas it would be more difficult to generate OOD data for e.g. tabular datasets.\\n\\nWe introduced new datasets because we were not aware of any image datasets used for regression. We view this as a contribution in its own right, as future regression-OOD research can utilize these datasets/baselines for evaluation.\\n\\nW.R.T. error bars: we will update our results table with multiple runs of our experiments.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper proposes to detect inputs that are from a slightly shifted distribution (eg images of houses in CA instead of KY) or from a very shifted distribution (eg imagenet images) by fitting a density model to the last layer h(x) of an MLP trained with squared error, and using the likelhiood score p(h(x)) as a metric. This is a reasonable idea. However, it is not novel eg Aigrain'19 did essentially the same thing for classification models. (The difference between classification and regression is a trivial change to the loss function, and does not change the fundamental idea.)\\n\\nIn addition to lack of novelty, the experimental methodology is very weak. First, the toy 2d example is too trivial to be informative, since there is essentiallly no overlap between the two distributions of features, p(x) and q(x) - even a density model on input space could detect this. More importantly, the results on the two image datasets are suspect. First, it seems that using the predictive variance sigma(x) as the reliability metric (the \\\"var\\\" method) - which is totally standard approach known as 'heteroskedastic regression'. - works very well in several cases. I suspect when it fails it is due to implementation problems (eg trying to predict sigma instead of log(sigma)). Also ensembles are known to be very robust to distribtution shift (see eg Ovadia'19), so I am surprised at their poor performance. Another problem is that the datasets used are not standard, so it is impossible to compare to other papers. Finally, no error bars are reported, so it is hard to know if any of the results are statistically significant. \\n\\n\\n\\nJ. Aigrain and M. Detyniecki, \\u201cDetecting Adversarial Examples and Other Misclassifications in Neural Networks by Introspection,\\u201d in ICML Workshop on Uncertainty and Robustness in Deep Learning, 2019 [Online]. Available: http://arxiv.org/abs/1905.09186\\n\\n\\nY. Ovadia et al., \\u201cCan You Trust Your Model\\u2019s Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift,\\u201d arXiv [stat.ML], 06-Jun-2019 [Online]. Available: http://arxiv.org/abs/1906.02530\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"===Summary===\\n\\nThe authors propose to perform out of distribution detection for regression models by fitting a generative model in the feature space of the regression model. An input example is deemed to be out-of-distribution if it has low likelihood under this generative model.\\n\\n===Overall Assessment===\\n\\nI recommend that the paper is rejected. There are a number of aspects that need to be improved. You should fix these and resubmit to a future conference. \\nThe paper focuses on the difference between regression and classification tasks and claims that the paper's method addresses an unmet need for OOD for regression. However, both the proposed method and the analysis justifying it are generic enough to be applied to both regression and classification.\\nThe paper handles technical claims far too casually in sec 4 and does not provide sufficient justification that the claims are true.\\nThere are natural baselines, such as using a generative model on the raw input space, that are ignored.\\n\\n===Comments===\\n\\nRemark 1 feels to me like it was added for the sake of having more math in the paper, not because it is crucial to the paper's argument. \\n\\nYou remark at various places that existing methods don't naturally generalize from classification to regression. However, you never fully explain why. Also, your proposed method can be applied out-of-the box to classification problems. Your analysis in sec 4 trivially applies to binary classification tasks, and could be naturally extended to multi-class classification where w is not a vector but a num_classes x num_features matrix. \\n\\nThe parallel should be between classification and heteroskedastic regression, since there you have a distribution per example.\\n\\nThe logic in \\\"In-distribution features are intrinsically low dimensional\\\" is insufficient\\n\\nThe connection between section 4 and your proposed method is not particularly precise. You also have lots of technical claims in 4 that are unsubstantiated. For example, you write \\\"this new network will likely have less discarded information than the shallower network\\\". What does 'likely' mean? In what sense are you making an actual technical statement? Each of the subsections in sec 4 has similar issues.\\n\\\"The CNNs are pre-trained on ImageNet (Denget al., 2009) and the last layer is replaced with a linear layer that produces a single output.\\\"\\nWhy did you do this? Did you fine tune or just retrain the top layer?\\n\\n\\\"For these two baselines, the variance of the forward passes is used as a metric for detecting OOD inputs\\\" \\n Can you explain why these are reasonable baselines for OOD? Why no baseline that fits a generative model in input space?\\n\\nYou should cite Ren et al. \\\"Likelihood Ratios for Out-of-Distribution Detection\\\"\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This work is tackling the problem of doing out-of-distribution detection in regression. The core idea is to fit a simple generative model to hidden activations and use thresholded likelihood to decide whether an input is in or out of distribution. Using a simple generative model for this is motivated by the intrinsic low-dimensionality of hidden activations.\\n\\nMy issue with the narrative of this paper is two-fold. \\n\\nFirst, the intrinsic low-dimensionality of the hidden activation distribution seems entirely problem-dependent. One can train networks with orthonormal layers and therefor flat spectra that work fine and there are also invertible discriminative models that do not discard any information from layer to layer, albeit being very similar to standard ResNets. It might be that this heuristic works well for the problems discussed in the paper, but I don't see any reason for this to hold in general.\\n\\nSecond, Nalisnick et al. don't show that deep generative models are often overconfident (as stated in the paper here). What they show is that likelihood can be higher for datasets with lower entropy than the one the deep generative model has been trained on. They also show that \\\"graying\\\" inputs increase the likelihood and they can even show this analytically for flow models with data-independent determinant of their Jacobian. The GMM approach proposed here is thus a special case of what Nalisnick et al. described and should have the same problem. Reducing the variance of the features should increase the likelihood.\\n\\nAll in all, the presented method relies too much on heuristics that seem specific to the analyzed architectures and datasets and there is no theoretical reason for this to hold in general as far as I can see.\\n\\nThe originality of the paper is also somewhat limited, as generative models for OOD detection are well-studied. Thresholded likelihoods are problematic in many settings, so I am not convinced this approach will work well across the board.\\n\\n----\\n\\nEric Nalisnick, Akihiro Matsukawa, Yee Whye Teh, Dilan Gorur, and Balaji Lakshminarayanan. Do deep generative models know what they don\\u2019t know? In ICLR, 2019.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors present the empirical observation that regression models trained using weight decay tend to produce high-level features that lie on low-dimensional subspaces. They then use this observation to propose an algorithm for detecting out of distribution data by fitting a simple GMM to the features produced during training.\\n\\nThe main contribution of the paper is a way of calculating the \\\"OOD scores\\\" that are reported in the different figures. However, after reading the paper 3 times I'm still unable to find a definition of these scores. I'm guessing it's negative log-likelihood under the fitted GMM, but I'm unsure.\\n\\nThe observation (features on low-dimensional manifolds) is interesting, and the proposed algorithm is potentially useful. However the experiments are quite limited, with only 2 training datasets. The analysis of why/when the algorithm is expected to work and how this depends on the model is shallow. The two main results are for models pre-trained on Imagenet: presumably this has a huge impact on the features that end up being used for OOD detection, but this is not discussed in the paper at all.\", \"questions_to_the_authors\": [\"Why focus only on regression? It seems to me that your analysis for the features lying on a low-dimensional subspace should also apply to classification models.\", \"How do your results depend on the specific models that are used? I can come up with models with very few final features where your method would not work. Do you have any analysis or guidance?\", \"What do all abbreviations in the tables mean? Please make the captions more informative so the reader does not need to search in the main text.\"]}" ] }

[ "Sina Baharlouei", "Maher Nouiehed", "Ahmad Beirami", "Meisam Razaviyayn" ]

[ "fairness", "algorithms", "particular", "critical", "systems", "human lives", "error", "systematic discrimination", "individuals" ]

{ "note_id": [ "ss-RU_lVAM", "rJxe__mcsS", "r1eBV_Q5sS", "SylrUPXqsH", "HJlAucY_qB", "r1gYZ0Tg5B", "HyxTwiqAKH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731319, 1573693543913, 1573693484851, 1573693261103, 1572538997537, 1572032000674, 1571887973184 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1741/Authors" ], [ "ICLR.cc/2020/Conference/Paper1741/Authors" ], [ "ICLR.cc/2020/Conference/Paper1741/Authors" ], [ "ICLR.cc/2020/Conference/Paper1741/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1741/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1741/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The paper addresses the problem of fair representation learning. The authors propose to use R\\u00e9nyi correlation as a measure of (in)dependence between the predictor and the sensitive attribute and developed a general training framework to impose fairness with theoretical properties. The empirical evaluations have been performed using standard benchmarks for fairness methods and the SOTA baselines -- all this supports the main claims of this work's contributions.\\nAll the reviewers and AC agree that this work has made a valuable contribution and recommend acceptance. Congratulations to the authors!\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": [\"We thank the reviewer for the positive assessment of our paper. We have updated several sections of the paper as follows:\", \"A new section conclusion is added to summarize the contribution of the paper and show future direction.\", \"Supplementary section D is updated to give a more comprehensive description of the datasets used in experiments.\", \"A new supplementary section F is added in which we train a 2-layer neural network on the adult dataset.\"]}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"First, we thank the reviewer for his/her valuable assessment of our paper. Below, we address the main concerns of the reviewer and mention the revisions we have made on the paper.\\n\\n[Fair word embeddings]\\nWe appreciate the reviewer for bringing this point up. The notions of fairness and the proposed optimization problems covered in this paper are based on the availability of explicit labels for the sensitive attribute. Hence, it is not straightforward to directly apply this framework to promoting fairness in the training of word embeddings unless there is an explicit attribute such as gender associated with every word that is to be protected. This remains as a natural direction for future exploration, and we have added a line in the conclusion section about it.\\n \\n[More complex models]\\nWe have added experiments on more complex models and problems to capture the effectiveness of the proposed Renyi fair inference. We trained a 2-layer neural network model on the adult dataset (32000 training points), where the batch-size = 128, and the number of epochs = 50. As discussed in the supplementary section F, imposing fairness, in this case, increases the training time from 12 minutes to 15 minutes. This shows the effectiveness of the framework when used alongside more complex models in practice. \\n \\n[Impact of sensitive features on performance and fairness] We have updated supplementary section D to address this comment. For instance, for the adult dataset, the logistic regression model (without imposing fairness) is highly biased against women (p% = 31%). If we include gender during the training phase, at the end of the day, the absolute value of the corresponding weight will be bigger than the corresponding weights of 10 features out of 14. Note that even if we exclude gender (as the sensitive attribute) during training, this will not make much of a difference (see section D). On the other hand, we see Renyi fair inference consistently achieves high fairness regardless of whether or not the sensitive attribute is included.\\nTo demonstrate this, we compare four different cases: Fair training (or not) + sensitive attribute included in training (or not), reported here.\\n \\nTest error and p% when gender included, no fairness regularizer: 85.29%, 32.04%\\nTest error and p% when gender not included, no fairness regularizer: 85.15%, 32.02%\\nTest error and p% when gender included, adding Renyi regularizer: 82.96%, 99.13%\\nTest error and p% when gender not included, adding Renyi regularizer: 82.96%, 99.13%\\n \\n \\n[Adding the gradient of the sensitive feature] We penalize the objective function if the Renyi correlation between the prediction vector and the sensitive attribute(s) vector is large. It is not clear to us how adding the norm of the gradient of the objective function with respect to the sensitive attribute can help making the model fair and provide guarantees as we do in this paper. Please let us know if we are misunderstanding your comment.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We thank the reviewer for constructive feedback. Below, we respond to the questions and explain the changes that have been made to the paper based on your comments.\\n\\n[More detailed description in the numerical experiments] We have expanded the legends of the figures and the description of the experiments. While we expanded Appendix D, we decided not to move it to the main body due to space limitations. \\n\\n\\n[Fig. legends] We have expanded the legends for all figures so that they are now all self-explanatory. \\n\\n[Baselines in Fig. 1 (c)] We added a paragraph explaining the differences between the classifiers used as benchmarks in Fig 1 (c) (second paragraph of Section 6) and the reasons why Renyi fair classification outperforms them.\\n\\n[More complex models]\\nWe have added experiments on more complex models and problems to capture the effectiveness of the proposed Renyi fair inference. We trained a 2-layer neural network model on the adult dataset (32000 training points), where the batch-size = 128, and the number of epochs = 50. As it is discussed in supplementary section F, imposing fairness in this case, increases the training time from 12 minutes to 15 minutes. This showcases the effectiveness of the framework when used alongside more complex models in practice.\\n\\n\\n[Scalability of Renyi Classifier] To understand the computational overhead introduced to the algorithm for imposing fairness, let us compare the Renyi fair classification approach to Non-fair-GD method. Non-fair-GD method is the gradient descent method applied to the empirical risk minimization problem with no regularization or constraint for imposing fairness. Let us start by comparing Algorithm 2 with its non-fair-GD counterpart. First of all, notice that Algorithm 2 requires an extra step (Step 3 in Algorithm 2) in each iteration compared to non-fair-GD. The computational complexity of step 3 is O(n) and is the same as computing the gradient in step 4. Thus the per-iteration computational complexity of Algorithm 1 is order-wise the same as its non-fair-GD counterpart. Moreover, as discussed in the revised paper after Theorem 4.2, the number of iterations required for convergence of Algorithm 2 matches the iteration complexity lower-bound. Therefore, the number of iterations required for convergence of Algorithm 2 is order-wise the same as its non-fair-GD counterpart in non-convex smooth cases. This observation shows that overall, we only lose in constants (if at all losing in terms of computational complexity). \\n\\nNow, let us compare Algorithm 1 with its non-fair-GD counterpart. In this case, step 3 in Algorithm 1 requires finding the second singular vector of the matrix Q. The dimension of this matrix depends on the number of classes in the classification and the number of categories in the sensitive attribute. Therefore, the dimension of matrix Q does not scale with the number of data points or the dimension of the model; and therefore in many cases remains small. For example, as a case study, for a problem that has 10 target classes, and 4 possible values for the sensitive attributes (white-man, white-woman, black-man, black woman as an example) finding the second largest eigenvector of a 10*4 matrix is not computationally expensive. Notice that for such a matrix the first singular vector is known (see v_1 definition after equation (6)). Thus one can subtract the component corresponding to the largest singular vector and use fast methods such as power method to find the second singular vector. This shows that practically the per-iteration computational complexity of Algorithm 1 is no more than its non-fair-GD counterpart. However, the total number of iterations (see Theorem 4.1) is larger than the total number of required iterations for convergence in the non-fair-GD counterpart. This can be viewed as the computational cost of imposing fairness in this case. \\n\\n\\n[Conclusion section] We have now written a conclusion section summarizing our contributions and findings.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"This work shows an interesting approach to introduce fairness to machine learning. The introduction and presenation of the approach are well written. The introduction provides a valuable overview and characterization of existing work and motivates the approach proposed in this work. Sections 2 through 5, together with the supplementary material provide a comprehensible description and derivation of the approaches to classification and clustering, together with the corresponding algorithms.\\nHowever, experimental section leaves open a number of questions. First of all, for readers not familiar with the task used here, a more detailed description would be desirable, including moving Supplementary Section D into the main text. In Suppl. Sec. D also a succinct definition of the classification task should be provided. What exactly are the classiffication tasks you solve for the Bank and Adult dataset, i.e. what is the set (and number) of classes in both cases?\\nFigs. 1 (a) and (b): you should provide a legend defining the red and blue curves.\\nFig 1 (c): without knowing the other approaches cited here, some comment on the differences in the classifier models used would be helpful to evaluate the differences in test error. Effectively, the reader would like to get an indication, what the differences in test error are induced by: difference in the regulatory fairness terms, or also differences in the classifier models.\\nOverall, the three datasets used to test the approach seem to be fairly limited, as are the corresponding classifier models used. It would be interesting to get an indication of how the approach would perform on/scale to much larger tasks with (much) deeper classification models?\\nCompletely missing is a final overall conclusion section.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a method to use renyi correlation to improve fairness of ML models by reducing the dependence of outputs wrt sensitive inputs such as gender. Previous models either use a linear dependence measure or use more complex optimization objectives; this paper improves upon them. They first build a min-max objective where the goodness-of-fit and fairness are jointly optimized. This non-convex objective is difficult to optimize and the authors propose a reformulation of the renyi correlation for discrete random variable case where it reduces to finding second largest eigenvalue. Based on this, they reformulate the objective which can be optimized more efficiently. They show the performance of their model for supervised and unsupervised learning problems on 4 different dataset by comparing to standard correlations such as Pearson.\\n\\nOverall the paper is clearly written and I liked the idea of using renyi correlation which also has a nice theoretical formulation allowing to be optimized more efficiently. But, the experimental results are a bit weak since the only model they experiment with is logistic regression. Given that their main motivation is to capture non-linear dependencies, I think some results with neural networks is necessary. Since their main focus is on discrete case, the authors can show if training a word embedding model with renyi regularization helps improve fairness of word embeddings.\", \"i_have_several_question_regarding_the_paper\": \"The datasets that authors use have a predefined feature space. Can you show if the sensitive feature is really important for high accuracy? Can we get the same performance without the sensitive features?\\nSince the model is trained with gradient descent, how would a more simple baseline where the gradient of the sensitive feature is penalized work?\\nHow would this algorithm generalize to larger problems such as language models since Q_{theta} is regenerated at every iteration?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a novel approach to fair inference/learning by using the Renyi correlation in place of the standard correlation measures (e.g. Pearson correlation) as a measure of \\\"unfairness\\\" expressed as a form of dependence between an outcome and a sensitive variable (which includes a number of standard group fairness definitions including demographic parity and equality of odds). This is motivated by the limitation of standard notions of correlation that they only capture linear correlations.\\nThe proposed approach is a formulation using the Renyi correlation as a regularization term, which leads to a min-max formulation due to the definition of Renyi correlation as the maximum correlation between any functionals of the variables in question.\\nThe paper then describes in detail how the computation can be done by making use of the fact that Renyi correlation can be computed using singular values of the relevant matrix, and provides computation time analyses for a couple of different assumptions. \\nThe authors also apply the analogous idea to (k means) clustering to derive what they call the Renyi fair clustering. \\nIn the experimental evaluation section, they evaluate the performance of the proposed methods with respect to the accuracy fairness trade-off, using 3 publicly available real world data sets and with the 2 standard definition of group fairness. The experimental results show quite convincingly that the trade of is improved by Renyi correlation as compared to the use of standard correlations. They also evaluate the Renyi fair clustering method using a fairness metric they propose for evaluating the uniformity of clustering results.\\nOverall the paper presents a crisp new idea, develops sound theoretical analysis and algorithms, and validate their superiority with satisfactory experiments.\"}" ] }

[ "Xin WANG", "SiuMing Yiu" ]

[ "generative classifiers", "selective classification", "classification with rejection" ]

{ "note_id": [ "iym2I_Bw3U", "Wis-BzcTr", "HklRHtZ_jH", "SJlnF6GvsS", "BJgFY6lvoB", "SylJmNTn5H", "rkxXYzgb9r", "Syeg3q4AtH" ], "note_type": [ "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576904914104, 1576798731289, 1573554501724, 1573494147736, 1573485953265, 1572815894617, 1572041338881, 1571863207952 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Paper1740/Authors" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1740/Authors" ], [ "ICLR.cc/2020/Conference/Paper1740/Authors" ], [ "ICLR.cc/2020/Conference/Paper1740/Authors" ], [ "ICLR.cc/2020/Conference/Paper1740/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1740/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1740/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Thank you for your detailed and helpful meta-review\", \"comment\": \"Thank you for your detailed and helpful meta-review. We are so grateful for this, and will improve this draft accordingly.\\n\\nThe only sad thing is that we can only see these points in meta-review, and fail to receive any further responses in the rebuttal phase. So that we have no chance to address them in the rebuttal, and only to see this paper rejected.\"}", "{\"decision\": \"Reject\", \"comment\": \"This paper combines a well-known, recently proposed unsupervised representation learning technique technique with a class-conditional negative log likelihood and a squared hinge loss on the class-wise conditional likelihoods, and proposes to use the resulting conditional density model for generative classification. The empirical work appears to validate the claim that their method leads to good out of distribution detection, and better performance using a rejection option. The adversarial defense results are less clear. Reporting raw logits is a strange choice, and difficult to interpret; the table is also difficult to read, and this method of reporting makes it difficult to compare against existing methods.\\n\\nThe reviewers generally remarked on presentation issues. R1 asked about the contribution of various loss terms, a matter I feel is underexplored in this work, and the authors mainly replied with a qualitative description of loss behaviour in the joint system, which I don't believe was the question. R1 also asked about the choice of thresholds and the issues of fairness of comparison regarding model capacity, neither of which seemed adequately addressed. R3 remarked on the clarity being lacking, and also that \\\"Generative modeling of representations is novel, afaik.\\\" (It is not; see, for example, the VQ-VAE line of work where PixelCNN priors are fit on top of representations, and layer-wise pre-training works of the mid 2000s, where generative models were frequently fit on greedily trained feature representations, sometimes in conjunction with a joint generative model of class labels). R2's review was very brief, and with a self-reported low confidence, but their concerns were addressed in a subsequent update.\\n\\nThere are three weaknesses which are my grounds for recommending rejection. First, this paper does a poor job of situating itself in the wider body of literature on classification with rejection, which dates to at least the 1970s (see Bartlett & Wengkamp, 2006 and the references therein). Second, the empirical work makes little comparison to other methods in the literature; baselines on clean data are self-generated, and the paper compares to no other adversarial defense proposals. In a minor drawback, ImageNet results are also missing; given that one of the purported advantages of the method is scalability, a large scale benchmark would have strengthened this claim. Third, no ablation study is undertaken that might give us insight into the role of each term of the loss. Given that this is a straightforward combination of well-understood techniques, a fully empirical paper ought to deliver more insight into the combination than this manuscript has.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Responses\", \"comment\": \"Thank you for your comments.\", \"about_manifold_assumption\": \"Manifold assumption is a very basic assumption \\\\emph{implicitly} made by many ideas behind machine learning (See [1] section 5.11.3 Page 161-164 of Ian Goodfellow's Deep Learning book). In particular, one important citation of this paper, [2] make manifold assumption very clear in its section 2.\", \"the_key_difference_of_our_paper_is_that\": \"(1) we enforce the manifold on the representations learned from data; (2) the form of the manifold is explicitly constrained (i.e. supervised constraints of SDIM) to perform generative classification. This also constitutes the foundation of our rejection decision function. Illegal inputs (out-of-distribution samples or adversarial examples) will be rejected if their class conditionals are smaller than the corresponding thresholds.\", \"about_experiments\": \"The logits in Table 4 and Table 6 are simply class conditionals. Table 4 and 6 are provided as typical cases that how the adversarial examples are rejected (detected) by the chosen thresholds (listed are 1st percentile thresholds). In table 4, inspecting the logits, i.e. class conditionals of adversarial examples. If we simply output the label corresponding to the maximum (like what typical discriminative classifiers do), we make a wrong prediction. But we use the rejection decision function, we will successfully reject (detect) them (highest logits smaller than thresholds). \\n\\nI'd like to provide some clarification about the whole experiments.\", \"in_short\": \"(1) show that SDIMs are able to achieve same-level accuracies as discriminative counterparts (previous generative classifiers based on full generative models report very poor accuracies (<80%) on CIFAR10)\\n\\n(2) effectively reject illegal inputs (OoD samples or adversarial examples) based on off-manifold conjecture, i.e. reject samples if their class conditionals are smaller than thresholds.\\n\\nTypos are corrected in the updated revision, please check.\\n\\nWe would like to address your further concerns about this paper if you have any.\\n\\n\\n[1] Deep Learning, Ian Goodfellow and Yoshua Bengio and Aaron Courville.\\n\\n[2] Li, Yingzhen et al. Are generative classifiers more robust to adversarial attacks? ICML 2019\"}", "{\"title\": \"Responses\", \"comment\": \"Thank you for your comments.\\n\\n1. Your suggestion of a structure map was exactly what we wanted to do, but failed due to the 8-page submission limit. We've added a structure map of SDIM framework in the revised paper. Please check. \\n\\n2. Table 2 is the classification accuracies on the clean test datasets. Since the thresholds we choose are 1or2 percentiles of training samples' class conditionals, a small portion of \\\"legal\\\" (clean) test samples may be rejected. So we report the rejection rates of \\\"legal\\\" samples, as well as the classification accuracies on the left test samples. We can see, if we choose a higher threshold, i.e. 2-th percentile, more \\\"legal\\\" requests will be rejected by SDIM without a prediction label; but SDIM becomes more confident and accurate on the left test sets.\\n\\n3. The choice of the threshold is actually very simple. We choose one threshold for each class only based on the training sets. Say for class $c$, the class conditionals of all training samples is a set $S_c = \\\\{y_i\\\\}_{i=1}^{N_c}$. 1-th percentile is simply the (0.01 * $N_c$) th smallest value of set $S_c$, so on and so forth. Note that this also means 1 percent of the training samples will be rejected without outputing prediction label by our rejection decision function.\\n\\n4. In our experiments, we use MI lower-bound specified by Jensen-Shannon (JS) divergence, our MI losses are very close to, also bounded by, the maximum of JS divergence $\\\\log 2$ . The likelihood margin loss $\\\\mathcal{J}_{\\\\text{LM}}$ usually reduce to zero. With all these two bounded, the NLL is minimized, but is unbounded. \\n\\n5. I understand that even a trivial CNN model can easily get >99% accuracy on MNIST. In our experiments, for clarity and consistency, we use a popular definition of residual networks that can be universally evaluated on CIFAR10, SVHN, FashionMNIST, MNIST. For MNIST and FashionMNIST, we simply choose the simplest network of the definition($n=1$ for each residual stage), and no special attention is paid. It would also not surprising that achieving the same-level results with much smaller network. \\n\\n6. I understand your concern about the fairness of comparisons. But for here, discussing about the network size is actually meaningless, and also not applicable: (1) Poor performance. On CIFAR10, [1] reports <50% accuracy, and [2] is 54%. (2) Networks sizes of [1] [2] not available. [1] simply mentioned the poor result (<50%) on CIFAR10 in their preliminary experiments (the network structure not available in the open-sourced code); their generative classifier in the paper is trained on the features of VGG16, which is not comparable with SDIM. [2] also simply mentioned the poor result on CIFAR10, and their models are completely evaluated on MNIST.\\n\\n[1] Li, Yingzhen et al. Are generative classifiers more robust to adversarial attacks? ICML 2019\\n\\n[2] Lukas Schott, et al. Towards the first adversarially robust neural network model on mnist. arXiv preprint arXiv:1805.09190, 2018\"}", "{\"title\": \"Responses\", \"comment\": \"Thank you for your comments.\\n\\nWe've added more results of OOD detection (histograms of in-out samples' assigned class conditionals) in Section C of Supplementary part. Please Check.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper proposes a scalable approach to train generative classifiers using information maximizing representation learning, with the motivation that generative classifiers could be more robust to adversarial attacks than discriminative classifiers. An off-the-shelf mutual information maximizer (MINE, DIM) is used to learn low-dimensional representations of images. Then, class-conditioned generative models of the representations are learned avoiding full generative modeling of the images. An additional loss is used to train the generative classifier which maximizes likelihood margins. Finally, percentile-based thresholds of the class log-probabilities is proposed to be used to reject classification for out-of-manifold inputs.\\n\\nThe paper cites existing literature which indicates that generative classifiers might be more robust to adversarial attacks, and uses recently-proposed representation learning techniques to scale up learning generative generative classifiers. The motivation of the proposed technique is clear, and the problem itself is relevant. Similar prior work use generative modeling at the pixel-level. Generative modeling of representations is novel, afaik.\\n\\nThe technique is evaluated on out-of-distribution sample detection (FashionMnist->MNIST and Cifar10->SVHN) and adversarial attacks. FGSM, PGD, CW-L2 attack, deepfool.\\n\\nThe experiments section is not very clearly written. Some of the evaluation itself is nonstandard in which only the first example of digit 0 of MNIST is used. The paper needs to have a clearer explanation and interpretation of the results. It\\u2019s not clear what the logits in table 4 and table 6 are, and what is being shown by the comparison.\", \"summary\": \"The authors propose a new technique for training generative classifiers with the aim to improve robustness to adversarial attacks and confidence on out-of-distribution samples. The method is well-motivated and explained, but the experiment section is not very clearly written and I\\u2019m not confident whether the technique represents an advancement in the state-of-the-art or not.\", \"misc_comments\": \"\\u201cBy adopting a generative approach p(x, y) = p(y)p(x|y), we assume that the data follows the manifold assumption: the (high-dimensional) data lies on low-dimensional manifolds corresponding to their class labels.\\u201d\\n\\nThis sentence seems to be making a stronger claim than is needed, which is possibly incorrect. Assuming a generative approach doesn\\u2019t require assuming the manifold assumption.\", \"nit\": \"\\u201cstate-of-the-arts\\u201d\", \"typo\": \"\\u201cThe original image is the fist sample of class 0\\u201d\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper studies classification problems via a reject option. A reject option could be useful in prediction problems to handle Out-of-distribution examples. The classification procedure studied in this paper builds on three components 1. An auto-encoder that obtains a latent low-dimensional representation of the data point 2. A generative model that models the class-conditional probability model and 3. a margin based loss function that learns a classifier that provides a large probability mass to the class-conditional distribution corresponding to the correct class. The final decision procedure is to reject an input if the best class conditional probability is small and to use the class corresponding to the best class conditional probability otherwise.\\n\\nOn the whole I like the paper and think that the problem tackles an important problem. I have a few comments\\n1. I would like to see what is the log-likelihood assigned by the proposed procedure on OOD samples and would like to see a comparison of the log-likelihood assigned by other procedures.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposed a supervised method for robust classification. By utilizing mutual information constrain on the encoder, supervised probabilistic constraint on the class conditional probability, and introducing a margin to the maximum likelihood, the proposed method can manage a high classification accuracy and detect the out of distribution data.\\n\\nThe motivation is good, the writing is OK. The structure of the paper needs to be refined. The experiments are not very strong. Several concerns are listed:\\n\\n1.\\tIt will be better if the author can show a structure map for Encoder, classification, class condition embedding, and mutual information evaluation networks, to clarify their relationships.\\n2.\\tIn Table 2, what is the accuracy for the rejection (outlier detection)?\\n3.\\tThe choose of the threshold for rejection is tricky. It will be better if the author can provide some rules for choosing the threshold which can be generalized to a new dataset.\\n4.\\tThe overall loss functions have three components, what is the contribution of different components to the final performance experimentally?\\n5.\\tThe architecture for encoder is large. Is it proper for Mnist which is a relatively simple dataset?\\n6.\\tThe comparison talked in the paragraph \\u201cIs Fully Generative Model Necessary for Generative Classification\\u201d, are these accuracies obtained from a comparable network size? It only makes sense if they are obtained using a comparable parameter size.\"}" ] }

[ "Moti Kadosh", "Yael Moses", "Ariel Shamir" ]

[ "Geo Localization", "Deep Learning", "Computer Vision", "Camera Localization" ]

{ "note_id": [ "ky2WhDr-Xz", "S1gCdj5T9H", "BJxZfI20KB", "SklaSE3uKB" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731257, 1572871030175, 1571894793468, 1571501125036 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1739/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1739/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1739/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The submission studies the problem of geolocalizing a city based on geometric information encoded in so called \\\"lean\\\" images. The reviewers were unanimous in their opinion that the submission does not meet the threshold for publication at ICLR. Concerns included quality of writing, novelty with respect to existing literature (in particular see Review #2), and limited validation on one geographic area. No rebuttal was provided.\", \"title\": \"Paper Decision\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper evaluates a geo-localization task based on \\\"lean\\\" images only, obtained by projection of 3d models without texture information. Multiple levels of granularity of the lean images (edges/edges+faces/edges+faces+depth) are compared for the learning, both in a \\\"memorization\\\" setting and in a \\\"generalization to unseen poses\\\" setting. Moreover, the behavior of the learning when labels are shuffled is evaluated.\\n\\nThe paper is motivated by identifying lean images with \\\"pure geometrical\\\" representation, and argues that the study is a demonstration that the network can learn some spatial map. I do not share this analysis: the edges, shapes and colors of the lean images still create some characteristic shapes that do not have to be linked with the 3D map in order to be detected; it is not clear why the detection from lean images is more geometrical in essence w.r.t. textured images. The motivation of the paper is therefore not compelling, and the study would have more relevance as an ablation study of a method using both geometry and texture. Can you give more compelling arguments for the problem setup, rather than a mere thought experiment?\\n\\nOther than this particular setup of geo-localizing based on lean images only, the novelty of the paper and the approach seem limited and does not offer new insights in geo-interpolation methods, geometry-aware neural networks, or memorization. Therefore, I believe the paper does not meet the standards of ICLR.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"## Contribution of the paper:\\n1. The paper argues that \\\"lean images\\\" (geometric structure of a 2D scene) contains sufficient information for solving the memorization of geolocation tasks.\\n2. The paper provide an approach to evaluate the argument.\\n3. The paper provide experimental validation that the proposed method works fairly on a certain Berlin dataset.\\n\\n## Feedback\\nThe paper is interesting but it does not meet the standard of this conference. There are some key reasons:\\n1. The writing is poor. The paper is hard to follow. There are also missing references. The importance of geometry in geolocation has been studied in many previous works.\\n2. The paper is empirical so it has to provide strong empirical evidence to support the claim. The experiment is only conducted in Berlin data whose scope is fairly small. Without large scale experiments, it is hard to convince readers that CNN is able to \\\"memorize\\\" all the locations using just geometric features. Without comparing to RGB baseline or other existing approaches, it is hard to convince readers that the proposed approach works fairly well. It also seems the input images are just 2D projections so without experiments using real 2D images, it is hard to say the approach is effective.\\n\\n## Improvements that could be done\\n1. Please make sure everything in the tables are clearly explained. The metrics should be explained somewhere in the caption. It is hard to find the meaning of some notations in the table such as \\\"37K images\\\", \\\"Area 400x400\\\". Please clearly state the unit.\\n2. Please include some baseline results into the tables so that readers know how much improvments achieved by your approach.\\n3. Please reorganize the paper to make it more clear. \\n4. Please provide more literature review and discuss the difference of your work to existing ones.\\n\\n\\n## Questions\\n1. How large is the Berlin data?\\n2. How would your method work when there is only a street image? How would you method scale when 3D model is not always available.\\n3. How would you method work when there is some places without much structure? Rural area could be an example.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper evaluates the performances of Deep Learning for geo-localization tasks in a world of images without textures (\\\"lean images\\\"). More exactly, the lean images are images rendered from a 3D model of a city. made of the depth and/or the buildings' edges and/or the buildings' faces. For the purpose of the evaluation, no real images are used, only lean images both at training and test times. Many lean images are generated for training a network to predict the camera pose (2 translation parameters, two angles), as a classification problem or a regression problem.\\n\\nThe motivation for the study is to understand the use of geometric aspects by a deep network to solve the task. However, as the paper presents mostly the results of the quantitative evaluation, it is not clear what we really learn about the mechanisms of the deep network. I think the empirical results can be interesting for people working on geo-localization, but probably not for the audience of ICLR.\\n\\nI suspect the authors have in mind for the future to learn to predict a lean image from a real image. I would find this aspect very interesting. It would also be interesting to evaluate the performance on the geo-localization task when noisy images are used as input.\"}" ] }

[ "Holger Schwenk", "Vishrav Chaudhary", "Shuo Sun", "Hongyu Gong", "Francisco Guzmán" ]

[ "multilinguality", "bitext mining", "neural MT", "Wikipedia", "low-resource languages", "joint sentence representation" ]

{ "note_id": [ "zNhs71lX3", "HygOmHZm2H", "rJlWFce3sB", "SJgJ9AOsiH", "ByemxAuijr", "H1e-3TuojH", "SklmxWok9S", "B1eiO93RYB", "ryecO-JoYS" ], "note_type": [ "decision", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731227, 1574274335844, 1573812856660, 1573781126720, 1573780971476, 1573780905299, 1571954922958, 1571895923425, 1571643761975 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1738/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1738/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1738/Authors" ], [ "ICLR.cc/2020/Conference/Paper1738/Authors" ], [ "ICLR.cc/2020/Conference/Paper1738/Authors" ], [ "ICLR.cc/2020/Conference/Paper1738/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1738/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1738/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors present an approach to large scale bitext extraction from Wikipedia. This builds heavily on previous work, with the novelty being somewhat minor efficient approximate K-nearest neighbor search and language agnostic parameters such as cutoffs. These techniques have not been validated on other data sets and it is unclear how well they generalise. The major contribution of the paper is the corpus created, consisting of 85 languages, 1620 language pairs and 135M parallel sentences, of which most do not include English. This corpus is very valuable and already in use in the field, but IMO ICLR is not the right venue for this kind of publication. There were four reviews, all broadly in agreement, and some discussion with the authors.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper presents a multi-lingual multi-way pseudo-parallel text corpus automatically extracted from Wikipedia.\\n\\nThe authors use a variety of pre-existing techniques applied at large scale with substantial engineering effort to extract a large number of sentence pairs in 1620 language pairs from 85 languages.\\n\\nIn the proposed method 1) raw sentences are extracted from a Wikipedia dump, 2) LASER sentence embeddings and language IDs are computed for each sentence, 3) for each language pair candidate sentence pairs are extracted using a FAISS approximate K-nearest neighbor index on the cosine distance between sentence embeddings, 4) sentence similarity scores are computed between the candidate pairs using the \\\"max margin\\\" criterion of Artetxe & Schwenk, 2018 and finally 5) sentence pairs are selected according to a language-pair-agnostic threshold on the similarity scores. \\n\\nThe extraction method is symmetric w.r.t. language directions for each language pair.\\n\\nStructural metadata of Wikipedia, such as cross-lingual document alignments, is deliberately not exploited (some discussion is provided but I would have preferred an empirical comparison of local vs global extraction). \\n\\nThe similarity threshold is determined by evaluating training corpora extracted at different thresholds on a machine translation task on De->En, De->Fr, Cs->De and Cs->Fr translation directions, evaluated on WMT newstest2014, and manually selecting the threshold based on BLEU scores. The paper also reports that combining the automatically extracted corpora with Europarl results in strong BLEU improvements over training only on Europarl. BLEU scores on TED test sets obtained using only the automatically extracted corpus are also reported. The corpus has been released.\\n\\nOverall the methodology presented in the paper is strong and the corpus is likely going to become a valuable tool to build machine translation systems and other multi-lingual applications. However, I am concerned that ICLR 2020 may not be the appropriate venue for this paper, as in my understanding dataset release papers are not explicitly solicited in the Call for Papers https://iclr.cc/Conferences/2020/CallForPapers . The corpus generation method is based on existing techniques, and to the extent that the engineering effort is innovative, it might not necessarily transfer well to data sources other than Wikipedia, thus limiting its broad scientific value. Therefore I suggest to submit the paper to a different venue.\"}", "{\"title\": \"Response to authors\", \"comment\": \"Regarding point 1, to maintain anonymity it was enough to create, for example, a temporary account on GitHub. However, I am aware of the problem of anonymity and my consideration was about a check that could be performed by the reviewers, not about the tests presented in the article.\\n\\nRegarding point 2, section 4 explains how the data corpus has been prepared and talks about threshold optimization. The sentence I mentioned about the comparison of different mining parameters at this point seems misleading in my opinion because there is only one parameter that is optimized (comparisons are implicit in optimization). Reading about the comparison of parameters I expected instead of a discussion about parameters and results obtained from different systems. There is instead a comparison between the application on different languages. Therefore, I thank you for the clarification that allowed me to better frame the article.\\n\\nHaving said that, I would like to thank the authors for taking my comments well into consideration and for replying punctually to all of them, clarifying some of my doubts. I think I can happily increase my score.\"}", "{\"title\": \"We would like to thank the reviewer for his thorough review.\", \"comment\": \"\"}", "{\"title\": \"We would like to thank the reviewer for his work. In the following, we will comment on the remarks of the reviewer:\", \"comment\": \"1) \\u201cHowever, without access to the data and, more importantly, extensive testing of it, it is difficult to say how and how much it would help the advancement of the field\\u201d\\nWe would like to point out that we have already open sourced the data. This is mentioned in the abstract and the conclusion of the paper (the download URL is not given to guarantee anonymity). In addition, we have trained more than 1800 neural machine translation systems covering 45 languages and report BLEU scores for all of them on the TED corpus. 23 systems achieve BLEU scores over 30. We believe that this qualifies for \\u201cextensive testing\\u201d, given that in most of the papers in NLP, evaluation is limited to a handful of tasks or language pairs.\\n\\n2) \\u201cEverything is described at a high level without going into detail, e.g. no real comparison of the mining parameters.\\u201d\\nWe would like to point out that our algorithms are described in detail in section 4 which spans over 2 pages. In particular, our mining approach has only one parameter, the mining threshold. We provide an extensive study of the impact of this parameter in section 4.2 for four different language pairs (see Figure 1).\\n\\n3) \\u201cReal content is too little to be published at ICLR, informs only on the presence of a new dataset\\u201d\", \"we_would_like_to_argue_that_this_paper_makes_a_substantial_contribution_for_several_reasons\": [\"this is the first approach which can be applied to many language pairs without the need to adapt it to the specific language pair. In fact, the corpus is particularly useful for research in low resource translation, which is an active research area.\", \"the mined corpus is unique in its genre, with respect to the number of languages covered, including many low-resource languages, and the fact that we consider all possible language pairs (instead of the usual English/foreign pair).\"]}", "{\"title\": \"We would like to thank the reviewer for his work. Please find below our comments.\", \"comment\": \"We understood that the major concern of the reviewer is that this work is only an incremental extension of previous work. There is indeed a large body of research on bitext mining, as described in the related work section, which was recognized by the reviewer as very up-to-date.\\n\\nWe would like to point out that, to the best of our knowledge, our work is the first one to systematically mine for parallel data in Wikipedia, with one unified approach. None of the preceding approaches could be applied to such a large number of languages (we handle 85 languages). Almost all of the existing approaches focus on alignment with English only, while we provide alignments for all 1620 language pairs (Table 4 of the paper).\\n\\nFinally, we would like to mention that we make freely available all the mined parallel sentences to foster research on multilingual models.\\n\\nWe will change the text as requested, in particular the definition of the acronyms.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This ICLR submission deals with an strategy for the automatic extraction of parallel sentences from Wikipedia articles in 85 languages, based on multilingual sentence embeddings.\\nThe review is delivered with the caveat that I am not an expert in this particulat field.\\nThe paper is well written and structured, being within the scope of the conference. \\nThe literature review is very up to date and overall relevant to provide an appropriate context to the investigation.\\nI reckon this is a very interesting piece of work, but also that it draws too heavily on previous work from which the study is just an incremntal extension.\", \"minor_issues\": \"All acronyms in the text should be defined the first time they appear in the text.\", \"1st_sentence_of_section_2\": \"typo on \\u201ccomparable coprora\\u201d.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper creates a large dataset for machine translation, called WikiMatrix, that contains 135M parallel sentences in 1620 language pairs from Wikipedia. The paired sentences from different languages are mined based on the sentence embeddings. Training NMT systems based on the mined dataset, and comparing with those trained based on existing dataset, the authors claim that the quality of the dataset is good. The effect of thresholding values of similarity scores for selecting parallel sentences is studied. Since the data is huge, dimension reduction and data compression techniques are used for efficient mining. The study is the first one that systematically mine for parallel sentences of Wikipedia for a large number of languages. The results are solid and the dataset is valuable for research in multilinguality.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper presents WikiMatrix, an approach to automatically extract parallel sentences from the free text content of Wikipedia. The paper considers 1620 languages and the final dataset contains 135M parallel sentences. The language pairs are general and therefore the data does not require the use of English as a common language between two other languages.\\n\\nTo evaluate the quality of the extracted pairs, a neural machine translation system has been trained on them and tested on the TED dataset, obtaining good results in terms of BLEU score.\\n\\nThe article provides information on the system used to extract parallel sentences and opens up different directions for future investigations.\\n\\nThe dataset seems, from the given details, useful. However, without access to the data and, more importantly, extensive testing of it, it is difficult to say how and how much it would help the advancement of the field. For the moment it seems to be good. However, I am not really sure that this paper could be of interest to a wide audience, except for those involved in machine translation.\\n\\nIn general, the article describes everything at a high level, without going into the real details.\\nAn example of this is on page 6, section 4.2, where the article says that its purpose is to compare different mining parameters, but I do not see any real comparison. Some words are spent for the mining threshold, but there is no real comparison, while other possible parameters are not considered at all.\\n\\nFor this reason, I would tend to give a low score, which does not mean that the dataset is not good. It means that the real content of the paper seems to me to be too little to be published at ICLR, since the paper only informs about the presence of this new dataset, saying that it contains a large number of sentences and seems to allow good translations based on the results of a preliminary test.\", \"typos\": [\"on page 9 \\\"Aragonse\\\"\", \"on page 9, end penultimate line, the word \\\"for\\\" is repeated.\"]}" ] }

[ "S Chandra Mouli", "Leonardo Teixeira", "Jennifer Neville", "Bruno Ribeiro" ]

[ "Lifetime Clustering", "Deep Learning", "Survival Distributions", "Kuiper two-sample test" ]

{ "note_id": [ "F0tWpnIGyD", "rkeJx_dhjr", "rklkla9osS", "SJlq3yvEjH", "rJg3hAU4sH", "SJlhmaL4jr", "Sygn43LNir", "Hkeo3cU4iS", "S1lAJysb9r", "H1gI2HOCYB", "B1l4mKqqFS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731197, 1573844967061, 1573788902545, 1573314482227, 1573314228024, 1573313828220, 1573313587572, 1573313203009, 1572085477794, 1571878317809, 1571625243547 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1737/Authors" ], [ "ICLR.cc/2020/Conference/Paper1737/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1737/Authors" ], [ "ICLR.cc/2020/Conference/Paper1737/Authors" ], [ "ICLR.cc/2020/Conference/Paper1737/Authors" ], [ "ICLR.cc/2020/Conference/Paper1737/Authors" ], [ "ICLR.cc/2020/Conference/Paper1737/Authors" ], [ "ICLR.cc/2020/Conference/Paper1737/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1737/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1737/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors propose a clustering algorithm for users in a system based on their lifetime distribution. The reviewers acknowledge the novelty of the proposed clustering algorithm, but one concern left unresolved is how the results of the analysis can be of use in the real world examples used.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for updating your score\", \"comment\": \"We have fixed the minor error in Table 1, thanks. \\n\\nWe have stated the assumption about summary statistics in Section 3.1. Additionally, we have added sentences in both the Friendster and MIMIC experiments to remind the reader. \\nWe have added a sentence (last line of page 2) stating that all the users of a cluster share the same lifetime distributions (also stated in the Contributions paragraph). \\n\\nQ. I do not agree with the author that the minimum value in Equation (4) is a better choice than the summations. Consider that there are a pair of $\\\\hat{S}_i$ and $\\\\hat{S}_j$ are very similar to each other, while the differences between other pairs are moderate ... large. \\nYes, in the example provided, only $\\\\Delta(\\\\hat{S}_i$, $\\\\hat{S}_j)$ will be optimized for the current iteration and possibly several subsequent iterations of the optimization. However, as the optimization progresses, $\\\\Delta(\\\\hat{S}_i^\\\\text{new}$, $\\\\hat{S}_j^\\\\text{new})$ will become greater than the divergences between other pairs such that the minimum value in Equation (4) changes. That is, there will be $\\\\hat{S}_k$ and $\\\\hat{S}_l$ such that $\\\\Delta(\\\\hat{S}_k$, $\\\\hat{S}_l) < \\\\Delta(\\\\hat{S}_i^\\\\text{new}$, $\\\\hat{S}_j^\\\\text{new})$, and hence for that iteration, the model optimizes $\\\\hat{S}_k$, $\\\\hat{S}_l$. Continuing in this fashion, we can achieve higher divergence between every pair of clusters. \\nWe have performed experiments maximizing sum of divergences in the past, and have confirmed that it does not work (please see Figure 12 in the supplementary material). Here we provide an example against maximizing sum of divergences. Consider $K=4$ balanced clusters with lifetime distributions $\\\\hat{S}_1 = \\\\hat{S}_2 \\\\neq \\\\hat{S}_3 = \\\\hat{S}_4$. Note that such a clustering is not desirable since there are virtually only 2 clusters found (as $\\\\hat{S}_1$ and $\\\\hat{S}_2$ are the same, etc). But for such an assignment, sum is relatively high $\\\\sum_{i\\\\neq j} \\\\Delta(\\\\hat{S}_i, \\\\hat{S}_j) = 4 \\\\cdot \\\\Delta(\\\\hat{S}_1, \\\\hat{S}_3)$. The sum could be further optimized by simply increasing divergence between $\\\\hat{S}_1$ and $\\\\hat{S}_3$ while keeping all other clusters equal to one of $\\\\hat{S}_1$ or $\\\\hat{S}_3$.\\nOn the other hand, since $\\\\min_{i\\\\neq j} \\\\Delta(\\\\hat{S}_i, \\\\hat{S}_j) = 0$, our model will maximize the divergence between $\\\\hat{S}_1$ and $\\\\hat{S}_2$ first. \\n\\nWe have added section A.5 and Figure 12 in the supplementary material discussing the issues with using sum instead of minimum. We have also added a footnote below Definition 3 (clustering problem) pointing interested readers to section A.5.\"}", "{\"title\": \"Thank you for your response\", \"comment\": \"I appreciate you added Table 1 in the revisions. It helps me understand the definition such as $\\\\chi$ is the time elapsed and $S$ is a CCDF. Right now I am able to understand the model proposed in this paper. However, in table 1, I do not think $\\\\widehat{S}$ is a \\\"discrete vector\\\" because each component in this vector has a continuous value between 0 and 1.\\n\\nThe proposed method assumes that the membership of each subject is a function of the summary statistics over $M_i^{(u)}$ and $Y_i^{(u)}$; while all members in each cluster share the same life distribution. I suggest the authors state this assumption explicitly. \\n\\nI do not agree with the author that the minimum value in Equation (4) is a better choice than the summations. Consider that there are a pair of $\\\\hat{S}_i$ and $\\\\hat{S}_j$ are very similar to each other, while the differences between other pairs are moderate. If we choose to optimize for the minimum value in Equation (4), the model will update $\\\\hat{S}_i$ and $\\\\hat{S}_j$ only; while all clusters other $i$ and $j$ are ignored in the optimization, despite that the divergences between some other pairs of clusters might still be increased. When we optimize for the summation, all pairs of clusters, rather than the most similar pair, will be involved simultaneously in the optimization. I believe that summation is a better choice for the objective function, especially when the number of clusters is large. \\n\\nI have updated my review rating because my concerns are addressed.\"}", "{\"title\": \"Response to Review #3 (Part 2/2)\", \"comment\": \"Q4. \\\"In practice, how is K selected? based on the Friendster results for K=2,3,4,5 they all seem to produce distinct clusters, so which one should be used?\\\"\\nSimilar to standard clustering algorithms, the choice of $K$ is left to the practitioner; however, the validation performance of the clusters with respect to the Logrank scores can act as a useful guide. We have added section A.4 (pg 17-18) and Figures 10 and 11 in the supplementary material discussing this heuristic to select the optimal number of clusters.\\nFigure 10 shows the Logrank scores on a held-out validation set for clusters obtained from DeepCLife-KuiperUB on the Friendster dataset with $K=2\\\\ldots 7$. We observe that $K=4$ gives the best mean Logrank score and $K=5$ is close behind. However, for $K>5$, the Logrank scores drop drastically indicating that there are no more clusters of interest with distinct lifetime distributions from the ones already found.\\nPlease see Figure 11 for a similar discussion on MIMIC-III results based on Logrank scores for $K=2\\\\ldots 8$.\\n\\nQ5. \\\"This raises a question: how are the clusters, ... informing the use case? In the MIMIC III experiment, ... finds two clusters without strikingly different survival functions (relative to Friendster). Can one really consider $S_2$ as the high-risk group? Can one get higher risk clusters with larger K?\\\"\\nThe clusters found by DeepCLife-KuiperUB in MIMIC-III dataset with $K=2$ are significantly different as evaluated by the Logrank test (score $=205.25$, p-value $\\\\approx$ 0). Visual comparison of the clusters' lifetime ditributions across datasets can be misleading, especially since the timescale is very different (50 months in Friendster vs 30 days in MIMIC-III). In prior works (e.g., [2]), typically any two clusters with Logrank p-value less than 0.05 have been classified as high-risk/low-risk. \\nWe have added Figure 9 in the supplementary material showing the empirical lifetime distributions of the clusters found by DeepCLife-KuiperUB in the MIMIC-III dataset for $K=2...6$. We observe that there are no considerably higher risk clusters for larger values of $K$ than the one already obtained with $K=2$. However, we do obtain a very-low-risk cluster for $K=3,4,6$. \\nWe have provided a brief qualitative analysis of the clusters found by DeepCLife-KuiperUB in the Friendster dataset (please see last paragraph of Friendster experiment). We show that the clusters align with our intuition, i.e., a subject in the high-risk cluster has on average far fewer friends (1.56 vs 7.76) and sends fewer comments (1.07 vs 5.06) in the initial $\\\\tau=5$ months.\\n\\nQ6. \\\"Minor: $A_i^{(u)}$ is missing from the definition in the training data. In Section 3.2, if $\\\\xi^{(u)}$ is shared, why the superscript indicating that is subject specific?\\\" \\nWe describe the termination signals in the same paragraph describing the training data, but separate from other variables because the availability of termination signals depends on the domain (from the paper: \\\"The training data may or may not contain the termination signals, $A^{(u)}_i$ ... \\\").\\nWe have replaced $\\\\xi^{(u)}$ with $W_2$ to avoid confusion.\", \"references\": \"[2] Bair, Eric, and Robert Tibshirani. \\\"Semi-supervised methods to predict patient survival from gene expression data.\\\" PLoS biology 2.4 (2004): e108.\"}", "{\"title\": \"Response to Review #3 (Part 1/2)\", \"comment\": \"Thank you for your positive comments and feedback.\\n\\nQ1. \\\"It is not entirely clear how to handle the case when after the model is trained, ... what happens with $\\\\beta^{(u)}(W_2)$?\\\"\\n$\\\\beta^{(u)}$ are only computed and used during training. For a new test subject, the model outputs cluster assignments $\\\\alpha^{(u)}$ using only the covariates $X^{(u)}$ and the subject's events till time $\\\\tau$ as described in Equation (2). We do not need to calculate $\\\\beta^{(u)}$, $H^{(u)}$, $\\\\chi^{(u)}$ for a test subject; indeed we will not be able to calculate them since they are all tied to $t_m$. $H^{(u)}$ and $\\\\chi^{(u)}$ are only used during training to obtain a probabilistic proxy for true lifetime of $u$ via $\\\\beta^{(u)}$. \\nWe have added a paragraph in Section 2 to clarify this distinction between training and test subjects.\\n\\nQ2. \\\"When the termination signals are not available (friendster data) ... set artificially when training and evaluation, ... what is the point of doing this? ... what is the point of calculating C-Index and Integrated Brier score? I understand the motivation to compare to existing methods, it is really not clear how meaningful they are at evaluating performance.\\\"\\nThe paragraph titled \\\"Termination signals for evaluation and baselines\\\" (page 7 in the paper) clarifies this. The termination signals are artificially set only for the baselines (since they cannot handle unobservability), and during evaluation (since *all* the performance metrics including Logrank score require termination signals). Our approach is trained without any termination signals; they are learnt using the smooth \\\"timeout\\\" window $\\\\beta^{(u)}$ as part of the optimization. This unfairly helps the competing methods since they are trained and evaluated using termination\\nsignals, whereas our approach is not trained with these termination signals. \\n\\nQ3. \\\"In general, using C-Index and Brier scores ... may be misleading (or non-applicable) ... horizon. Here, $\\\\beta^{(u)}(W_2)$ is a function of $\\\\chi^{(u)}$ which is known ... and not dependent on covariates as in survival analysis. Further, SSC-Bair and SSC-Gaynor are clustering models, how are survival estimates obtained for these?\\\"\\nWe used all metrics available in the literature, Logrank, C-index and Brier score to further validate the discriminative power of the clusters. This is a standard quantitative evaluation procedure used in unsupervised learning, common, for instance, in learning unsupervised image representations (e.g., [1]). \\nThe answer to (1) above should clarify that $\\\\beta^{(u)}(W_2)$ and $\\\\chi^{(u)}$ (both tied to $t_m$) cannot be computed for test subjects. \\nThe clusters from all the methods (SSC-Bair, SSC-Gaynor, DeepHit+GMM, DeepCLife) are evaluated the same way. Given the cluster assignments $\\\\kappa(u') \\\\in \\\\{1, \\\\ldots, K\\\\}$ and the termination signals (possibly using a ``timeout'' window as described in answer to (2) above) for all the users $u'$ in the test data, we can obtain the empirical lifetime distribution of all the clusters $\\\\hat{S}_k, \\\\: \\\\forall k \\\\in \\\\{1,\\\\ldots,K\\\\}$ using the Kaplan-Meier estimates (over the test data alone). Then, the empirical lifetime distribution of a user $u'$ is given by that of her assigned cluster, i.e., $\\\\hat{S}^{(u')} := \\\\hat{S}_{\\\\kappa(u')}$ (shared for all users in the cluster). Integrated Brier score can be computed as usual using $\\\\hat{S}^{(u')}$ (Equation 14). C-index is computed by using the expected lifetime obtained from the lifetime distribution $\\\\hat{S}^{(u')}$ as the predicted lifetime of $u'$. \\nWe have updated the Metrics subsection in the paper to include the above discussion.\", \"references\": \"[1] Bengio, Yoshua. \\\"Deep learning of representations for unsupervised and transfer learning.\\\" Proceedings of ICML workshop on unsupervised and transfer learning. 2012.\"}", "{\"title\": \"Response to Review #2 (Part 2/2)\", \"comment\": \"Q6. \\\"How does the proposed neural network take inputs of different lengths?\\\"\\nAs described in Section 3.1: \\\"In our experiments, we compute summary statistics over the observed events $\\\\{M^{(u)}_i, Y^{(u)}_i\\\\}_i$ in order to make it compatible with the feedforward architecture.\\\" In Friendster experiments for example, we compute number of comments sent/received, number of people interacted with, mean inter-event times, etc. Note that our model is not restricted to a feedforward architecture, a recurrent architecture could be used instead. Our primary contribution is the loss function and a framework for lifetime clustering.\\n\\nQ7. \\\"How the empirical distribution $\\\\hat{S}_i$ in Equation (4) is computed is also not clear to me ... is a valid probabilistic distribution? Since $\\\\hat{S}_i$ is high dimensional, it looks very challenging to estimate the joint distribution.\\\"\\n$\\\\hat{S}_k$ is simply a shorthand for the vector $\\\\hat{S}_k(W_1, W_2; D)$; both denoting one-dimensional empirical lifetime distribution $P[T_k > t]$ of a cluster $k$. We use the notation $\\\\hat{S}_k(W_1, W_2; D)$ to emphasize that it is a function of the model parameters $W_1$ and $W_2$; however, $\\\\hat{S}_k$ is not a joint distribution of these parameters. We believe Table 1 (notations) will clarify this to the reader.\\n\\nPlease see Figures (2b), (2c) and (3) for example lifetime distributions. Since we consider discrete times $t=0,1,2..$, $\\\\hat{S}_k$ is a vector such that $\\\\hat{S}_k[j] = P[T_k > j]$, i.e., the probability that the lifetime of a random subject in cluster $k$ is greater than $j$ time units. \\n\\n$\\\\hat{S}_k$ is obtained using modified Kaplan-Meier estimates. As described in Section 3.3: \\\"Kaplan-Meier estimates are a maximum likelihood estimate of the lifetime distribution of a set of subjects assuming (a) hard memberships and (b) the presence of termination signals. We modify the estimates to account for partial memberships and probability of termination instead\\\". By construction (Proposition 1), $\\\\hat{S}_k$ is a valid CCDF. \\n\\nQ8. \\\"The overall objective function ... In Equation (4), why should we compute the minimum values across all possible pairs of clusters rather than the summation of all pairs?\\\"\\nYes, the final objective function is given by Equation (4) which matches our lifetime clustering goal defined in Equation (1). We want *every* pair of clusters to have different survival distribution from each other. By maximizing the minimum divergence (say $\\\\delta$), we make sure that every pair of clusters are at least $\\\\delta$ apart. On the other hand, if we maximized the sum of divergences, then the model could maximize the divergence between one pair of clusters unboundedly while keeping all other of clusters very similar to one another; note that in this case, the sum of divergences will still be maximized, whereas the minimum divergence will not.\\n\\nQ9. \\\"Equation (4) does not contain a component that maximizes a likelihood function ... It looks like the model will converge to a trivial solution that $\\\\beta$ is a constant such ... divergence between distributions.\\\"\\nThe empirical lifetime distribution $\\\\hat{S}_i$ of each of the clusters is obtained by maximizing the likelihood (please see Proposition 1 or Kaplan-Meier estimates). Additionally, $\\\\beta^{(u)}$ are constrained by the non-decreasing function $\\\\beta^{(u)} = 1 - e^{-W_2 \\\\cdot \\\\chi^{(u)}}$, where $\\\\chi^{(u)}$ is the period of inactivity of user $u$. The non-decreasing function provides a structure such that the model cannot arbitrarily assign 0 and 1 to $\\\\beta^{(u)}$. For example, if a user $u_1$ with 10 months of inactivity is assigned $\\\\beta^{(u_1)} = 1$ then all users $u'$ with inactivity period greater than 10 months will be assigned $\\\\beta^{(u')} = 1$ as well.\\n\\n\\nWe hope that we have addressed all of your concerns and convinced you that the paper does not contain technical flaws. Please let us know if you have additional concerns.\"}", "{\"title\": \"Response to Review #2 (Part 1/2)\", \"comment\": \"Thank you for the detailed review. We would like to clarify that the paper has absolutely no technical flaws, as we detail next.\\n\\nQ1. \\\"The problem setting of this paper is not clear to me. I am not sure which variables are observed and which are not.\\\"\\nTo help the reader, we have moved the Table of notations from the Supplementary Material to the main paper (now Table 1).\\nSince the clusters are unobserved, all variables with subscript $k$ are unobserved. In Definition 1, we use variables with subscript $k$ to indicate a random variable indexed by a random subject of cluster $k$. The definition describes only the underlying (hidden) generative process for the observed training data. \\n\\nConsider only the Friendster social network where $u$ denotes a user in the system. Then we consider the following variables:\\n- $X^{(u)}$ (observed) is a vector of user covariates (e.g., age, gender).\\n- $Y^{(u)}_i$ is the inter-event time between activity events (e.g., sending comments) and $M^{(u)}_i$ is a vector of event covariates.\\n - For training users, the activity events are observed for a long time (within $[0, t_m]$) to compute observed lifetime $H^{(u)}$ and period of inactivity $\\\\chi^{(u)}$ (both of these variables are tied to $t_m$). \\n - For test users, the activity events are observed only till time $\\\\tau = 5$ months from their joining. We wish to cluster the test users using their covariates and the initial activity events. \\n We have added a paragraph in the paper to clarify this distinction between training and test users.\\n\\n- $\\\\theta^{(u)}$ (observed) is the joining time.\\n- $A^{(u)}_i$ is the termination signal (observed in healthcare datasets, unobserved in social networks).\\n- $T_k$ (unobserved) is the lifetime of a random subject in cluster $k$. \\n- $S_k$ (unobserved) is the lifetime distribution (CCDF) $P[T_k > t]$ of cluster $k$.\\n\\n\\nQ2. \\\"In the paper, do we consider only one type or multiple types of events?\\\"\\nWe consider different types of activity events (e.g., login, send/receive comments). Terminal event (e.g., quitting social media) is one that ends all activity events. Terminal event can have different causes; but lacking additional information about these causes, we do not model them separately in this paper.\\n\\nQ3. \\\"Is $M_{k, i}$ a vector that represents the attributes or properties of an event?\\\"\\nYes, $M_{k, i}$ is a random vector representing the attributes of the $i$-th activity event of a random user in cluster $k$. $M^{(u)}_i$ is the same for a particular user $u$.\\n\\nQ4. \\\"In the Friendster experiment, the data of 5 months from joining is used for clustering ... termination windows are ... 10 months. Therefore, it is clear that the observed data will not contain the termination signals ... I do not believe the training of the model is possible without observing termination signals.\\\"\\nUnobservability of termination signals is a property of the domain/dataset (healthcare or social networks) and not due to the choice of timeout window. For example, in Friendster, we do not know if or when the users deactivate their account; we only have information regarding their activities (i.e., that they have been inactive for say 15 months). In contrast, time of death (if occurred) is available in healthcare datasets.\\n\\nA \\\"timeout\\\" window of 10 months is an artificial way of providing these termination signals: if a user is inactive for greater than 10 months, then consider the user \\\"terminated\\\" (left the social network). Note that we use artificial termination signals only for baselines and to compute performance metrics, and not for training our model. We train our model by jointly learning a smooth \\\"timeout\\\" window in the form of $\\\\beta^{(u)}$. Whereas we see only $\\\\tau=5$ months of activities of test subjects, we observe training subjects for longer times (e.g., $[0, t_m]$ is 6 years long in Friendster experiments), thus facilitating model training.\\n\\nQ5. \\\"How the non-decreasing function $\\\\xi^{(u)}$ is defined in Section 3.2? Is it a function of the observed data for each subject?\\\"\\nAs described in Section 3.2, we use $\\\\beta^{(u)} := 1 - e^{-\\\\xi^{(u)} \\\\chi^{(u)}}$ with a shared scalar rate parameter $\\\\xi^{(u)} = W_2$. This makes sure that $\\\\beta^{(u)}$ is a non-decreasing function of inactivity; longer the inactivity, higher the probability of termination. To avoid confusion, we have replaced $\\\\xi^{(u)}$ with $W_2$ in the paper.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"Thank you for your positive comments and feedback.\\n\\nQ1. \\\"Potential impact is an outstanding issue here because the proposed method is very special-purpose. Are there clusters that a service provider can use to improve their service? One might need to obtain a large number of clusters, via conditioning perhaps, before finding some really useful ones.\\\"\\n\\nLifetime clustering is a very important task, with impactful applications in multiple fields. \\n\\n1. There are numerous biomedical applications for this approach for example, clustering gene profiles to identify cancer subtypes [1,2,3], identifying subgroups of patients with different clinical characteristics [4]. \\nConcurrent to our work, Chen et al. [1] (October 2019) propose a supervised deep learning approach to identify cancer subtypes with different lifetime distributions. However, their approach requires at least a few ground truth lifetime clusters for supervision, whereas our approach does not require such supervision.\\n\\n2. Similarly, identifying user subgroups in a social network can be a very useful tool. For example, a service provider might wish to obtain lifetime clusters of users based on their usage pattern alone (e.g., a high-risk cluster might only include people that use features X, Y and Z of the service). Identifying such groups can provide insights about the best/worst workflows in the service (different from predicting the lifetime of individual users). Note that the service provider need not obtain a large number of clusters to find these groups; she simply needs to remove all other covariates (like age, number of friends, etc.) whose analysis is not required.\\n\\nMoreover, of independent interest, the Kuiper loss provided by our paper (through our cheap-to-compute bound) is shown to be superior to MMD in our task. MMD finds a wide range of applications in machine learning. One could dedicate an entire paper just to the Kuiper bound; in our paper, it is just one of the contributions (albeit important).\\n\\nQ2. \\\"Were number of friends and comments included in the analysis?\\\"\\nYes, the number of friends and comments were used as covariates for lifetime clustering (along with 58 other covariates). Matching our intuition, number of friends and comments were correlated with the cluster assignments, thus qualitatively validating the clusters.\", \"references\": \"[1] Chen, Runpu, et al. \\\"Deep learning approach to identifying cancer subtypes using high-dimensional genomic data.\\\" Bioinformatics (2019). \\n[2] Jiang, Limin, et al. \\\"Discovering cancer subtypes via an accurate fusion strategy on multiple profile data.\\\" Frontiers in genetics 10 (2019): 20.\\n[3] Guo, Yang, Xuequn Shang, and Zhanhuai Li. \\\"Identification of cancer subtypes by integrating multiple types of transcriptomics data with deep learning in breast cancer.\\\" Neurocomputing 324 (2019): 20-30.\\n[4] Hastie, Barbara A., et al. \\\"Cluster analysis of multiple experimental pain modalities.\\\" Pain 116.3 (2005): 227-237.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The authors propose an approach to cluster subjects into K clusters according to their underlying, often unknown, life distribution. The loss function to the neural-network-based approach is based on a tight upper bound to the two-sample Kuiper test. Experiments are conducted on artificial and two real life datasets.\\n\\nI enjoyed reading the paper, the authors propose an approach to address an important yet relatively under-explored problem in survival analysis.\\n\\nIt is not entirely clear how to handle the case when after the model is trained, H^(u) for a new subject u is larger than t_max when the model is trained. In such case, \\\\Chi^(u) will be negative, thus what happens with \\\\beta^(u)(W_2)?\\n\\nWhen the termination signals are not available (friendster data) termination signals are set artificially when training and evaluation, thus, is the model learning artificially set termination signals and thus artificial survival functions? what is the point of doing this? Moreover, if the termination signals are artificial, what is the point of calculating C-Index and Integrated Brier score? I understand the motivation to compare to existing methods, it is really not clear how meaningful they are at evaluating performance.\\n\\nIn general, using C-Index and Brier scores in the context of the presented application may be misleading (or non-applicable) because these metrics are usually applied to survival analysis scenarios where one seeks to estimate likelihood of survival or time to event (over a usually infinite time horizon). Here, \\\\beta^(u)(W_2) is a function of \\\\Chi^(u) which is known (also not comparable to survival analysis), artificially tied to a pre-specified time-horizon t_max, and not dependent on covariates as in survival analysis. Further, SSC-Bair and SSC-Gaynor are clustering models, how are survival estimates obtained for these?\\n\\nIn practice, how is K selected? based on the friendster results for K=2,3,4,5 they all seem to produce distinct clusters, so which one should be used? This raises a question: how are the clusters, their members or the number of clusters informing the use case? In the MIMIC III experiment, where events are observed, the model finds two clusters without strikingly different survival functions (relative to friendster). Can one really consider S_2 as the high-risk group? Can one get higher risk clusters with larger K?\", \"minor\": [\"A_i^(u) is missing from the definition in the training data.\", \"In Section 3.2, if \\\\xi^(u) is shared, why the superscript indicating that is subject specific?\"]}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a method to cluster subjects based on the latent lifetime distribution. The proposed model clusters by maximizing the empirical divergence between different clusters.\\n\\nThe problem setting of this paper is not clear to me. I am not sure which variables are observed and which are not. For example, in the Friendster experiment, the data of 5 months from joining is used for clustering. However, the termination windows are chosen to be 10 months. Therefore, it is clear that the observed data will not contain the termination signals, and I do not believe the training of the model is possible, without observing any termination signals. In the paper, do we consider only one type or multiple types of events? Is $M_{k, i}$ a vector that represents the attributes or properties of an event?\\n\\nSome details of the model are not clear to me. In Equation (2), the input of the neural network differs in length across different subject $u$, because the number of observed events for each subject is different. \\nHow does the proposed neural network take inputs of different lengths?\\nHow the non-decreasing function $\\\\xi^{(u)}$ is defined in Section 3.2? Is it a function of the observed data for each subject? \\n\\nHow the empirical distribution $\\\\hat{S}_i$ in Equation (4) is computed is also not clear to me. How $\\\\hat{S}_i$ is a vector? Is it constructed by concatenating $\\\\hat{S}_k(W_1, W_2; D)[t] $ with different $t$? How to normalize $\\\\hat{S}_i$ such that it is a valid probabilistic distribution? Since $\\\\hat{S}_i$ is high dimensional, it looks very challenging to estimate the joint distribution.\\n\\nThe overall objective function is given by Equation (4), is it correct? In Equation (4), why should we compute the minimum values across all possible pairs of clusters rather than the summation of all pairs? If Equation (4) is the overall objective function, then it looks like the model does not contain a component that maximizes a likelihood function. How is it guaranteed that the model will fit the data? It looks like the model will converge to a trivial solution that $\\\\beta$ is a constant such that $\\\\beta = 1$ for one cluster and $\\\\beta = 0$ for another cluster, if the likelihood function is not involved. This will give a maximum divergence between distributions. \\n\\nIn summary, it seems that numerous technical details are missing and the paper might contain technical flaws. I do not suggest the acceptance of this paper.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposes a deep learning method for clustering subjects (patients, users of social networks) into clusters of different lifetime (survival time) distributions. Inputs for a subject include covariates (age, gender, etc) and an event sequence. A key difficulty in the problem is that the lifetime of a subject is often unobserved. The proposed method is to learn a termination function that is non-decreasing with time, which essentially treats prolonged inactivity as termination in a probabilistic way. Clustering is done in a discriminative manner where the objective function considers cluster sizes besides between-cluster differences in lifetime distribution. (The inclusion of cluster size information in the objective function avoids degenerate clustering results.)\\nThe paper is strong in technical contents. The objective function is well-motivated and placed on solid theoretical foundation. Empirically, the proposed method performed significantly better than baselines.\\nThe weakness of the paper is that the utility of the work has not been demonstrated. For example, the Friendster dataset (1.1 million social network users) is partitioned into 2-5 clusters. However, are there clusters that a service provider can use to improve their service? It is doubtful whether such \\u201cuseful\\u201d clusters can be found using the proposed algorithm. One might need to obtain a fairly large number of clusters, via conditioning perhaps, before finding some really useful ones. \\n\\u2022\\tThe paper is strong in technical contents. The objective function is well-motivated and placed on solid theoretical foundation.\\n\\u2022\\tEmpirically, the proposed method performed significantly better than baselines. \\n\\u2022\\tThe utility of the work has not been demonstrated. Potential impact is an outstanding issue here because the proposed method is very special-purpose.\", \"question_to_the_authors\": \"It is observed that the numbers of friends and comments are strongly correlated with the clustering. Were those two covariates included in the analysis? If not, why?\"}" ] }

[ "Hong Liu", "Mingsheng Long", "Jianmin Wang", "Michael I. Jordan" ]

[ "Transfer Learning", "Fine-tuning", "Deep Neural Networks" ]

{ "note_id": [ "OXspbi9VZF", "BygS4-hBoS", "S1e6J-3rjr", "r1gpWRjBoH", "H1lX2ToBiH", "BJgi1poSor", "r1eGgJ6h5r", "H1gSNMbatr", "Bkx40G0oFr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731167, 1573400876869, 1573400805410, 1573400068534, 1573399979319, 1573399779417, 1572814569595, 1571783213392, 1571705548333 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1736/Authors" ], [ "ICLR.cc/2020/Conference/Paper1736/Authors" ], [ "ICLR.cc/2020/Conference/Paper1736/Authors" ], [ "ICLR.cc/2020/Conference/Paper1736/Authors" ], [ "ICLR.cc/2020/Conference/Paper1736/Authors" ], [ "ICLR.cc/2020/Conference/Paper1736/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1736/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1736/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper studies the transfer of representations learned by deep neural networks across various datasets and tasks when the network is pre-trained on some dataset and subsequently fine-tuned on the target dataset. The authors theoretically analyse two-layer fully connected networks and provide an extensive empirical evaluation arguing that the loss landscape of appropriately pre-trained networks is easier to optimise (improved Lipschitzness).\\nUnderstanding the transferability of representations is an important problem and the reviewers appreciated some aspects of the extensive empirical evaluation and the initial theoretical investigation. However, we feel that the manuscript needs a major revision and that there is not enough empirical evidence to support the stated conclusions. As a result, I will recommend rejecting this paper in the current form. \\nNevertheless, as the problem is extremely important I encourage the authors to improve the clarity and provide more convincing arguments towards the stated conclusions by addressing the issues raised during the discussion phase.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to AnonReviewer #2 (Part 1)\", \"comment\": \"We really appreciate your insightful comments. All reviewers unanimously believe that this work is dedicated to an interesting and important open problem: understanding the transferability of deep representations. We apologize for the clarity issues in question. To fully address your concerns, we respond to the questions below and summarize them into a minor revision. We will appreciate further advice and discussions based on the revision.\\n\\n1.\\tOn the results of Section 3. \\n\\nThe \\u201csimilarity\\u201d here is not the \\u201cvisual similarity\\u201d between data. As we have pointed out in Section 5, what really matters is the similarity between the nature of tasks, i.e. both images and labels matter.\\n\\nNote that in the sense of granularity, ImageNet, Webcam and Caltech are general classification tasks for mainly coarse-grained objects, while CUB, Stanford Cars are fine-grained classification tasks. In this sense, Webcam and Caltech are more similar to ImageNet, while CUB and Stanford Cars are dissimilar.\\n\\nWe provide the results of training from scratch below, which are also added to Section B.3 (appendix) in the revision. The decreased percentage is calculated by dividing the error reduced in fine-tuning with the error of training from scratch. \\n_____________________________\\nDataset test error decreased percentage 1/sqrt(n) \\\\|W-W(0)\\\\|_F\\nWebcam 26.39\\t\\t98.29% 3.54\\nStanford Cars 46.25\\t\\t48.75% 5.95\\nCaltech-101 22.14 79.36% 2.34\\nCUB-200 52.70 60.15% 3.90\\n\\nComparing the results of training from scratch, the improvement of generalization error is consistent with our observation.\\n\\n2.\\tQuantifying how the knowledge is preserved during fine-tuning with Frobenius norm.\\n\\n- Correctness is guaranteed. A line of works have shown that the Frobenius norm of the deviation is directly related to the generalization error [1][2][3]. As proved in our theoretical analysis, the Frobenius norm is further upper bounded by a term quantifying the relationship between the pretraining and target tasks (See Theorem 2 in Section 6.3). Thus, the Frobenius norm of deviation is a crucial part in transfer learning, in that it behaves as the link between generalization error and pretrained-target relationship. \\n\\n- We agree that how well the transferred network performs at old tasks from the original dataset is also a reasonable measure of knowledge preservation, but it is a more suitable measure for \\u201ccatastrophic forgetting\\u201d that was extensively studied in those problems [4][5]. Our measure is more suitable for transfer learning, with attention focused on how much useful knowledge can be preserved to boost the target tasks.\\n\\n- Also note that the norm is an upper bound of the generalization error, which means small norm of deviation guarantees good transfer performance with high probability, but parameters with a fixed distance does not necessarily perform similarly. For example, if fine-tuning to different target datasets results in models with similar norm of deviation, these models will perform differently, but the generalization error bound will be similar. These results correspond well to our theoretical analysis (Section 6.3), in that the final accuracy will be influenced both by the norm of deviation and the intrinsic property of the target tasks.\\n\\n3.\\tAbout experiments in Figures 5 and 6.\\n\\nExperiments in Figure 5 and 6 are carried out on multiple datasets with multiple trials. Figure 5 (b) (c) is carried out on Stanford Cars. Figure 6 is carried out on CUB. Due to the space limitation, we only report results on one dataset in the main paper, and defer results on other datasets to the Section B.1 and B.2 (appendix) in the revision.\\n\\nExperiments in Figure 6 are inspired by [6], a well-known technique for analyzing Batch Normalization, and we follow its protocols. We add detailed implementation of this experiments to the Section A (appendix) in the revision.\\n\\nFigure 5a was taken from (He et al. 2018) with proper citation. To remove any possible concern, it is now replaced with another figure to show that transferring from a proper dataset indeed accelerates convergence.\"}", "{\"title\": \"Response to AnonReviewer #4 (Part 1)\", \"comment\": \"Many thanks for your helpful and insightful comments! All reviewers unanimously believe that this work is dedicated to an interesting and important open problem: understanding the transferability of deep representations. Please find our responses to your concerns below. We also attach a minor revision to reflect these responses, and provide the codes to guarantee reproducibility. We will be interested in any further feedback and discussions.\\n\\n1a) Elaboration on how this work \\u201cmay stimulate further algorithmic advances\\u201d.\\n\\nThis paper tries to understand the transferability of deep representations from both empirical and theoretical perspectives, and sheds insights on the feasibility of transfer. Basically, the benefit of pretrained representations is two-fold: better generalization and faster optimization, both of which indicate the directions of future fine-tuning algorithms. \\n\\n- Generalization. Our empirical and theoretical results on generalization error (Sections 3 & 6.3) indicate that the knowledge preserved, as quantified by the norm of the deviation, is determined by the intrinsic similarity between pretraining and target tasks. Note that, the norm of deviation is the consequence, and the similarity between tasks is the cause. Hence, a direct control of this norm term by regularization [1] is not necessarily helpful for drawing the tasks close, especially for dissimilar target tasks where negative transfer could happen. In this sense, we should rethink previous methods that control the norm term explicitly. Further algorithms for fine-tuning should consider implicit regularization to restrict the networks from escaping the flat region of pretrained landscape. And more importantly, seeking pretrained models with larger flat regions to enable safe transfer.\\n\\n- Optimization. Our empirical and theoretical results on optimization (Sections 4 & 6.2) uncover the major cause of accelerated convergence \\u2013 improved loss landscapes. However, current optimization methods (e.g. SGD with momentum) for fine-tuning is simply the same as those of training from scratch. Future optimization methods for fine-tuning should harness the landscape information with specific statistics reflecting the similarity between pretraining and target tasks.\\n\\n- Feasibility. Our results in Section 5.2 reveal an interesting issue of current pretrained models: while pretraining, instead of better performance on the pretraining dataset, a more realistic goal is the transferability of the pretrained models (e.g. less over-fitting, larger flat region). Algorithms for pretraining should take this point into consideration.\\n\\n1b) On the experimental methodology and details.\\n\\nWe add all necessary details of implementation in Section A (appendix) in the revision. In particular, the code is provided to reproduce all results.\\n\\nFor fine-tuning, we train ResNet-50 on each dataset for 200 epochs. The visualization method is the same as filter normalization [2]. The resolution of each visualization image is 200 * 200, and the step size is fixed for each figure. Other implementation details can also be found in Section A (appendix).\\n\\n1d) and 1e)\\n\\nWe rescale Figure 8 in the revision, so the results are now easy to read. Also, we have fixed several minor issues you raise in the revision.\\n\\n2a) About the results of training from scratch in Table 1 and Figure 2.\\n\\nWe definitely agree that the generalization error is also related to the intrinsic property of the dataset itself. Since the focus of this work is transfer learning (fine-tuning), we mainly provide results related to the similarity of tasks.\\n\\nIn fact, the conclusion corresponds well with our theoretical analysis. In Section 6.3, we have already pointed out that for some fine-grained dataset, fine-tuning from another dataset may be less effective, since the target data is \\u201csingular\\u201d, i.e. the target task itself is very difficult.\\n\\nTo this end, we further provide the results of training from scratch below and also in Table 2 in Section B.3 (appendix) in the revision. The decreased percentage is calculated by dividing the error reduced in fine-tuning with the error of training from scratch.\\n_____________________________\\nDataset test error decreased percentage 1/sqrt(n) \\\\|W-W(0)\\\\|_F\\nWebcam 26.39\\t\\t 98.29% 3.54\\nStanford Cars 46.25\\t\\t 48.75% 5.95\\nCaltech-101 22.14 79.36% 2.34\\nCUB-200 52.70 60.15% 3.90\\n\\nComparing the results of training from scratch, the improvement of generalization error is still consistent with our observation.\"}", "{\"title\": \"Response to AnonReviewer #1\", \"comment\": \"We really appreciate your insightful comments! All reviewers unanimously believe that this work is dedicated to an interesting and important open problem: understanding the transferability of deep representations.\\n\\nRegarding the theoretical results, we recognize there is still gap between our analysis and realistic networks.\\n\\n- Analysis of transfer learning relies heavily on the analysis of standard supervised learning. Exact analysis for realistic architectures remains an open problem for standard deep neural networks, let alone the scenario of transfer learning. \\n\\n- Current over-parametrization based techniques are very popular and implicative of future development on multilayer and other settings, so we analyze the property of transfer learning based on over-parametrization. To our best knowledge, this work provides the first theoretical analysis for transfer learning (fine-tuning) in deep neural networks.\\n\\nWe will delve into the theoretical property of realistic deep networks in our future work.\"}", "{\"title\": \"Response to AnonReviewer #2 (Part 2)\", \"comment\": \"4.\\tAbout the figures in Section 5.1\\n\\nWe apologize for the presentation issues in this section, and have fixed these problems in the revision.\\n\\n- The networks for the digit experiment are pretrained on various versions of MNIST and transferred to SVHN. The scaling of the generalization error is corrected in the revision and is now easy to read. \\\\psi refers to the convolutional kernel in LeNet. \\n\\n- The \\u201cVarying Labels with fixed input\\u201d experiments are carried out on transferring from Caltech-101 to Webcam. This figure is now presented in the revision. \\n\\n- The description of transferring from ImageNet and Places to MIT indoors and CUB200 are presented in Figure 8(c)-(d).\\n\\nThanks again for the constructive reviews, which were sincerely appreciated and carefully addressed in the revision. We hope these one-to-one responses are the right answers to your questions.\\n\\n--\\n[1] Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. NIPS, 2017.\\n[2] Sanjeev Arora, Rong Ge, Behnam Neyshabur, and Yi Zhang. Stronger generalization bounds for deep nets via a compression approach. arXiv preprint arXiv:1802.05296, 2018.\\n[3] Colin Wei, and Tengyu Ma, Data-dependent Sample Complexity of Deep Neural Networks via Lipschitz Augmentation, arXiv preprint arXiv:1905.03684, 2019.\\n[4] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell, Overcoming catastrophic forgetting in neural networks, PNAS 2017 114 (13) 3521-3526, 2017\\n[5] Michalis K. Titsias, Jonathan Schwarz, Alexander G. de G. Matthews, Razvan Pascanu, Yee Whye The, Functional Regularisation for Continual Learning, arXiv preprint arXiv:1901.11356, 2019\\n[6] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? NeurIPS, 2018.\"}", "{\"title\": \"Response to AnonReviewer #4 (Part 2)\", \"comment\": \"2b) Why is the Frobenius norm normalized by the square root of the number of target samples?\\n\\nThis normalization is in line with the generalization error in statistical learning theory [3]. The generalization error always contains the inverse of the square root of the sample number. In practice, even for the same dataset, different number of training samples will lead to different deviation from initialization. That is reasonable since the model has to deviate more to fit more samples. We confirm the validity of this theoretical bound by providing additional results of fine-tuning on CUB-200 in the table below.\\n _____________________________\\n Sample number 500 1000 2000 4000 5994 (full)\\n\\t\\\\|W-W(0)\\\\|_F 55.1 76.5 112.4 160.7 184.4\\n\\nThe concern on duplication of examples seems a misunderstanding. First, in most generalization theories, the data should not be degenerate, i.e. there are no two identical samples [4]. Second, the training samples in real-world applications are sampled i.i.d. from a continuous distribution. And the probability of sampling one data point twice is 0. \\n\\n2c) Regarding the permutation symmetry of the weights in Figure 4.\\n\\nWe totally agree with the insightful points raised by the reviewers. In fact, our results in Figure 4 correspond well to these points. More specifically:\\n\\n- The permutation symmetry of the weights. We show that, even fine-tuning for the same target dataset, based on models pretrained from different random initializations will converge to very different local minima.\\n\\n- Permutation symmetry broken with pretraining. We show that, the fine-tuned weights stay relatively close to the pretrained model even when fine-tuned to different datasets.\\n\\nHowever, the key points we want to uncover here are none of the above. What we focus on is that, given the above facts, how to understand the transferability of deep representations. Our insights are:\\n\\n- Due to the permutation symmetry, it is of little sense to quantify the preserved knowledge by the norm of deviation, if not conditioned on a particular pretrained model. Hence both our empirical and theoretical analyses are conditioned on a given pretrained model, such that the norm of deviation is informative to the generalization error bound.\\n\\n- We show that pretrained representations can implicitly restrict the fine-tuned weight matrices to stay near the pretrained weights. Since the pretrained dataset is sufficiently large and of high-quality, transferring their representations will lead to flatter minima located in large flat basins. This will finally help the generalization performance.\\n\\n2d) The experiment in Figure 6.\\n\\nThis experiment follows the same protocol as [5], a well-known technique for analyzing Batch Normalization. It is carried out on CUB-200, a classification task.\\n\\n- The magnitude of gradient is considered in both [5] and our paper since the smoothness of the loss landscape is a local property. Gradient is the proper scale for analyzing the smoothness. \\n\\n- We believe fixing step size along the gradient is also a reasonable scenario. To reflect on your advice, we further add results on Stanford-cars with fixed step size along the gradient in Section B.2 (appendix) in the revision. Results are similar for both scenarios. \\n\\n- The overall loss magnitude has little to do with the loss variation, since initially the loss is the same (log(#class) for cross entropy), and both converge to less than 1e-3.\\n\\n2e) About the figures in Section 5.1.\\n\\nWe apologize for the improper scaling and arrangement of those figures. The problems are now fixed in the revision. \\\\psi refers to the convolutional kernel in LeNet. W denotes the weight of the fully-connected layer. The results of transferring from Caltech-101 to Webcam is now provided in Figure 8.2 in the revision. \\n\\nThank you again for the constructive reviews, which were very helpful for the revision. And the time and detail put into these reviews were sincerely appreciated.\\n\\n--\\n[1] Xuhong Li, Yves Grandvalet, and Franck Davoine. Explicit inductive bias for transfer learning with convolutional networks. ICML, 2018\\n[2] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. NeurIPS, 2018\\n[3] Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of machine learning. 2012.\\n[4] Simon S. Du, Xiyu Zhai, Barnabas Poczos, and Aarti Singh. Gradient descent provably optimizes over-parameterized neural networks. ICLR, 2019.\\n[5] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? NeurIPS, 2018.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"I thank the authors for the clarifications and the modifications to the paper. However, I still lean towards rejection. While the authors provided detailed explanations on some of my points here, most of these are still not in the paper, so the reader would still probably be confused. There are new figures in the appendix, but they are not referred to from the text. The paper feels like a hastily compiled collection of weakly related, somewhat anecdotal and not necessarily clearly explained and motivated experiments. While I believe there is interesting content, with the current presentation style it is really difficult to digest. As an advice, I think it may be better to not squeeze in all the results into one paper, but rather focus on some aspects and analyze them really well, with clear explanation, motivation, and preferably with demonstrating practical gains that result from the analysis - not just hypothetical, but supported by experiments.\\n\\n---\\n\\nThe paper studies transfer of representations learned by deep networks across datasets and tasks. Namely, the paper analzes the standard setup with pre-training on one dataset and further fine-tuning on a different dataset. Results include both experiments and theory. On the experimental side, the focus is mainly on TODO. On the theory side, an analysis of transfer in two-layer fully connected networks, based on the ICML-2019 work of Arora et al., is proposed. \\n\\nI lean towards rejecting the paper, since the presentation and the technical quality are somewhat substandard. This is mainly based on evaluation of the experimental results, since I am not an expert in this subfield of theory and was therefore not able to check the statements and the proofs thoroughly. On the presentation side, many details are missing, which often makes understanding difficult and will likely lead to the results not being reproducible. On the experiments side, the issue is that they are quite anecdotal and have multiple confusing or questionable details, as specified below.\", \"pros\": \"1) An interesting topic of trying to understand generalization and transfer in deep learning\\n2) Multiple types of experiments, including visualizations of loss landscapes at convregence and at initialization, plots of Hessian eigenvalues, measuring the deviation of the weights from their initial values, measuring the variance of the gradients of the weights, measuring the transfer between different datasets, measuring the transfer performance depending on the durarion of pre-training.\\n3) Theoretical analysis. As mentioned, I cannot really judge about the quality of the results.\", \"cons\": \"1) The presentation is quite confusing. \\n1a) The paper includes many different experiments as well as theory, and it is not very clear how these all come together and what message do they give to the reader. The paper states at one point that it \\\"may stimulate further algorithmic advances\\\", and it would be great if there was a bit more elaboration on this. \\n1b) Experimental methodology is not presented in the main paper and not referred to. Some of it is described in the appendix, but also not too detailed, for instance the duration of ResNet training is not clear, the details of loss landscape visualization are confusing (for instance, the phrase \\\"... i.e. 0.1\\u00d7 gradient and a total of 200 \\u00d7 200 grids\\\"), \\n1c) The paper is over 9 pages, which is more than the recommended length of 8 pages.\\n1d) Scaling in Figure 8(b) is quite suboptimal, it is impossible to read the test accuracy results.\\n1e) Minor issues:\\n - No color-code in Figure 3 (unclear what values do the colors correspond to) and it does not seem to be referred to in the text. \\n - Page 5: \\\"predictive\\\" -> \\\"predictable\\\"?\\n - Page 6 \\\"While pretraining on sufficiently large datasets...\\\" - I do not think the experiments tell anything about the dependence of the effect on the size of the dataset, so this phrasing is not justified\\n\\n2) I found many of the experiments confusing or unconvincing. This is partially affected by the aforementioned issues with presentation.\\n2a) In Table 1 and Figure 2, it is unclear if difference in generalization between the dataesets is due to similarity to ImageNet (by the way, ImageNet is only mentioned in the caption of Table 1, but not in the text) or due to the inherent properties of the datasets (perhaps some are naturally more difficult or prone to overfitting). I think numbers for training from scratch would be helpful for disambiguating the two cases.\\n2b) It is unclear why is the norm difference normalized by the square root of the number of target examples. This choice is not justified and it seems it can very noticeably affect the results in counter-intuitive way. For instance, if I understand right, if one duplicates each example in the training set, these values will change, which seems somewhat counter-intuitive. Would it make more sense to normalize by the initial norm of the weights?\\n2c) In Figure 4, it is unclear if the permutations of the weights are accounted for. The output of a deep network is invariant under a wide variety of weight permutations, and it is natural that networks trained from different random initializations may converge to weights permuted in different ways. In order to meaningfully compare the weight vectors of these networks, one would have to first \\\"align\\\" them, which is a non-trivial task. Another issue I have with the results on Figure 4 is that I don't find them all that counter-intuitive: it seems natural that the weights stay relatively close to the pre-trained network when fine-tuned for that (paritally because the aforementioned permutation symmetry is already largely broken during pre-trianing).\\n2d) It is unclear which downstream task is analyzed in Figure 6. Moreover, the plot seems to mix two factors: magnitue of the gradient and the smoothness of the loss landscape. Would it not make more sense to make fixed-size steps in the direction of the gradient? Moreover, I am wondering if the difference between the loss values is simply because the overall loss magnitude is larger for the randomly initialized network?\\n2e) A few parts of subsection 5.1 are confusing. It is not clear where does the interpretation of Figure 8 follow from (perhaps partially because figure 8(b) is partially difficult to read). \\\\psi and W are not defined. What does it move that the weight matrices \\\"show no improvement\\\"? Where are the results on Caltech-101 and Webcam? Why is Food-101 mentioned in the text, but CUB-200 shown in the plots?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": [\"The paper gives an extensive empirical and somewhat limited theoretical analysis for the transferability in DNNs. It is shown that transferred models tend to have flatter minima with improved Lipschitzness in the loss function when good choices of pretraining are made.\", \"The paper is well written and well-organized. Notations and claims are clear.\", \"This paper presents an interesting line of research, that in my opinion, would be interesting to many researchers in the field, and could attract many follow up works.\", \"Empirical analysis in sections 3 and 4 are interesting, and give good sense of generalization and optimization landscape. Analyzing the Frobenius norm of the deviation between fine-tuned network and fixed network is reasonable.\", \"The theoretical analysis seems like a good start, but it is not sufficient in general. There seems to be gap between the network architectures used in empirical evaluations and the theoretical results. However, analyzing transferability is an important topic that needs to be evaluated more. This paper presents interesting new steps towards that goal. That being said, I would be interested to see theoretical results for more general cases alongside experiments on different types of applications.\", \"Overall, the paper presents a novel approach for evaluating transferability, that I think would be interesting to many researchers in this field. The theoretical results are still limited, and should be investigated more.\"]}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"N.B Since this paper exceeds 8 pages, it was held to a higher standard as per the instructions.\\n\\nThe main goal of this paper is to concretely demonstrate some key properties of transfer in Deep Neural Networks through experiment and theory. By transfer, here they specifically mean the ability for pre-trained networks to be 'updated' to new, similar datasets either completely (all parameters are updated while being initialised by the pre-trained network parameters) or partially (all but the last few layers are kept constant at pre-trained parameter values). There are of course other ways of carrying out transfer learning, but this paper focusses on these methods.They attempt to assess the viability of such a process in its improvements to generalisation and improvements to the loss landscape. In addition, the authors attempt to assess when this type of transfer is viable. The majority of the paper focus on experimental results, while the final 2 pages present some theoretical work that explains those results. \\n\\nI believe that this work is well motivated. As the paper suggests, there have been several advances in the use of transfer learning that showcase its benefits. That being said, there is a lack of work that systematically tries to explain why these benefits are seen, and how we can better make use of them. This paper tries to fill that gap. \\n\\nHowever, while it goes a way in trying to do so, I am not convinced that this paper sufficiently addresses what it sets out to do. This is why I recommend a weak reject, and a summary of my reasons for this are as follows:\\n\\n1) Section 3: In this section, the authors try to show that transferred networks tend to have better generalisation when the dataset being transferred to is similar to that the network was pre-trained on (ImageNet dataset). The results are shown in Table 1, and they assess this by showing that for new datasets that are more visually similar to the ImageNet dataset, the generalisation error is lower (e.g the Webcam dataset shows lower gen. error than the Stanford Car dataset). I believe that the authors are saying that they are 'visually' similar. However, the ImageNet dataset has subsets that are similar to the Webcam dataset (geological formation), and the Cars dataset (wheeled vehicle). As such, while the goal of the experiment is interesting, it is not clear how interpretable the results are, nor the validity of the conclusions raised. \\n\\nIn addition to this, I would have liked to see the generalisation error of a randomly initialised network for each of the datasets. This would have been an interesting control to see whether the pre-training does indeed improve generalisation. \\n\\nFurther, the authors use the Frobenius norm between the original pre-trained parameters and the final parameters as a measure of how much knowledge is preserved. I am not convinced that this is a sufficiently representative measure of this. I think the extent to which knowledge is preserved is indicated by how well the new, transferred network performs at old tasks from the original dataset. Simply measuring a distance between the parameters doesn't show this. Also consider the fact that it isn't true that networks with parameters a fixed distance away from the original parameters will have similar behaviour. \\n\\nFigure 3 isn't mentioned anywhere in the text!\\n\\n2) Section 4: Here, the authors show that pretrained networks lead to flatter, smoother loss landscapes when compared to randomly initialising. This is shown in Figure 5 and Figure 6 mainly. Figure 5 depicts the loss landscapes, and directly shows what this section is claiming. Figure 6 further solidifies this claim by showing that the change in loss at each gradient update is smaller when compared to a randomly initialised network. That being said, the experimental details of this setup is quite sketchy; what dataset is Figure 5 and 6 transferring to, having been pre-trained with ImageNet? Has this been tested between multiple different datasets, multiple times, to show that the conclusions are consistent? Further, it looks like Figure 5a was taken from another paper, why?\\n\\n3) Section 5: This section tries to answer the question of when transfer learning (as defined by this paper) is viable. Section 5.1 was quite difficult for me to read because I could not understand how the experimental setup described in the text. For example, Section 5.1 says that the network was pretrained on the MNIST dataset and transferred to the SVHN dataset, whereas Figure 8b states the complete opposite. If I assume that the text is correct, the generalisation error in Figure 8b is very difficult to read. In addition to this, I am not sure what the norm of phi-phi(0) is. I also still have my reservations to the use of the norm between parameters, as mentioned above. \\n\\nIn the the section 'Varying Labels with fixed input', the authors mention the use of Caltech-101 and Webcam data, but this isn't mentioned in Figure 8, instead, it mentioned CUB-200, which isn't mentioned in the text. They also mention conclusions from experiments using the Food-101 and Places datasets, but don't show these results anywhere. \\n\\nSection 5.1 asks important questions, but the authors haven't shown results that can properly answer them. \\n\\nThat being said, Section 5.2 shows the very interesting result that pre-training after a certain number of epochs starts showing diminishing returns in terms of performance of the transferred network.\"}" ] }

[ "Hae Beom Lee", "Taewook Nam", "Eunho Yang", "Sung Ju Hwang" ]

[ "latent features", "training examples", "learning", "generalization", "low errors", "meta dropout", "generalization meta dropout", "machine", "model", "unseen test examples" ]

{ "note_id": [ "N6YZNAvJX", "HkgSRBhniH", "BJgDya-vsB", "HJe3wYbPoS", "ByemJtbwsB", "rJgYVv-PoB", "ryg4oUWPsH", "S1e1sc_GqB", "rJxKsYxycS", "H1lRMzRpYr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731139, 1573860812720, 1573489887128, 1573488996181, 1573488859171, 1573488433295, 1573488283967, 1572141718603, 1571912097516, 1571836438249 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1735/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1735/Authors" ], [ "ICLR.cc/2020/Conference/Paper1735/Authors" ], [ "ICLR.cc/2020/Conference/Paper1735/Authors" ], [ "ICLR.cc/2020/Conference/Paper1735/Authors" ], [ "ICLR.cc/2020/Conference/Paper1735/Authors" ], [ "ICLR.cc/2020/Conference/Paper1735/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1735/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1735/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper proposes a type of adaptive dropout to regularize gradient based meta-learning models. The reviewers found the idea interesting and it is supported by improvements on standard benchmarks. The authors addressed several concerns of the reviewers during the rebutal phase. In particular, revisions added results against other regularization mthods. We recommend that further attention is given to ablations, in particular the baseline proposed by Reviewer 1.\", \"title\": \"Paper Decision\"}", "{\"title\": \"still not convinced\", \"comment\": \"I would like to thank the authors for the detailed response.\\n\\nIt\\u2019s indeed true that the formulas in Section 3.2 are correct, I apologize for the confusion. But I agree with other reviewers that the connection between the proposed method and variational inference is not significant. There is no variational approximation in the proposed method, only the Jensen lower bound. Equation 7 could be safely skipped. \\n\\nI spent significant time deliberating whether the baseline results provided in the paper and in the rebuttal provide enough evidence to say that meta-learning input-dependent noisy regularization is useful. Let me share my thoughts.\", \"the_proposed_method_can_be_seen_as_introducing_3_changes_to_the_model_and_the_learning_algorithm\": \"\", \"the_first_change_to_the_model\": \"element-wise multipliers z^(l), as well as the parameters \\\\phi that are used to compute z^(l)\", \"the_second_change_to_the_model\": \"adding a form of multiplicative noise\", \"a_change_to_the_algorithm\": \"the parameters \\\\phi are now treated differently and receive a different kind of derivative\", \"meaningful_subsets_of_the_three_proposed_changes_include\": \"(1), whereby \\\\phi and \\\\theta are treated identically and no noise is used. I am not sure if this is what the authors tried in their rebuttal, they did not say if they kept noise on. This is an important baseline because it isolates the change to the model. \\n(1, 2), which is to my best understanding what the authors did in the rebuttal when they tried \\u201c\\\\phi treated as \\\\theta\\u201d\\n(2): \\u201cFixed Gaussian\\u201d in Table 3\\n(1, 3): \\u201cDeterministic Meta-Dropout\\u201d in Table 3\\n(2, 3): \\u201cIndependent Gaussian\\u201d in Table 3\\n\\nTo justify that all these 3 changes are needed, one needs to show in several experiments how the combination of all 3 performs better than subsets. The paper performs 4 experiments: 1-shot Omniglot, 5-shot Omniglot, 1-shot MiniImageNet, 5-shot MiniImageNet.\\n\\nHere\\u2019s my summary of the results:\", \"1_shot_omniglot\": \"the proposed method works significantly better than baselines\", \"5_shot_omniglot\": \"using just (2) performs as well as the proposed method\", \"1_shot_imagenet\": \"the proposed method works significantly better than baselines, but the 42.08 accuracy that is reported in the rebuttal for (1, 2) baseline looks not very trust-worthy. Why would there be such a deterioration in this case, when in OmniGlot 1-shot case this baseline almost approached Meta-Dropout?\", \"5_shot_imagenet\": \"the 95% confidence intervals overlap for (1,3), (2,3) and the proposed method\\nAdditionally, to my best understanding, the baseline (1) was not tried. That said, I still can\\u2019t be sure that the main source of improvement is not the change in the model. \\n\\nGiven this summary of the results, I am not convinced that the paper clearly shows the benefit of meta-learning input-dependent noisy regularization. \\n\\nLastly, I disagree with the authors that the paper is clear w.r.t. what exactly \\\\phi is and how exactly is used. I did not find this information in Figure 2 and \\u201cForm of the noise\\u201d paragraph.\"}", "{\"title\": \"Response to Reviewer #3 (1/2)\", \"comment\": \"We appreciate your constructive comments. We respond to each comment as follows.\\n\\n1. The paper is somewhat incremental considering that Li et al, (2017) and Balaji et al., (2018) have already proposed meta-learning parameter-wise learning rates and parameter-wise regularization coefficient respectively.\\n\\n- This is a critical misunderstanding. The two papers you mentioned are not even superficially similar to our model. First Meta-SGD (Li et al., 2017) aims to meta-learn the element-wise learning rate, which is completely orthogonal to our model that meta-learns input-dependent perturbation of latent features. It also largely underperforms our method (See Table 1). \\n\\nMeta-Reg (Balaji et al., 2018) meta-learns the hyperparameter for a global regularizer while our Meta-Dropout directly learns to perturb each feature in an input-dependent manner. and aims to tackle a completely different problem of domain generalization. Nonetheless, to the best of our efforts, we tried to import the MetaReg idea to the conventional few-shot classification setting, such that for each task, its inner-optimization objective includes L1 regularization whose coefficients are element-wisely meta-learned. The results are as follows.\\n \\nModels\\t\\t \\tOmni-1shot\\tOmni-5shot\\tmimg-1shot\\tmimg-5shot\\nMAML\\t\\t \\t95.23+-0.17\\t98.38+-0.07\\t49.58+-0.65\\t64.55+-0.52\\nMetaReg (Ours)\\t\\t95.28+-0.15\\t98.85+-0.06\\t49.76+-0.67\\t65.42+-0.53\\nMeta-dropout\\t\\t96.63+-0.13\\t98.73+-0.06\\t51.93+-0.67\\t67.42+-0.52\\n\\nAs shown, Meta-dropout largely outperforms MetaReg, demonstrating the effectiveness of our framework, which meta-learns the input-dependent perturbation function.\\n\\n\\n2. It seems like the choice of the particular method for adding the noise was performed using the test set.\\n\\n- This is a complete misunderstanding. We never used a meta-test set for training or hyperparameter tuning. Test in the paper refers to the instances *simulating* test instances within the meta-training dataset, which is a common terminology in meta-learning.\\n\\n\\n3. Table 2 contains some results named \\u201cAdd.\\u201d, which I guess stands for additive noise. I did not find an explanation of what is the specific method for adding noise used in this case. \\n\\n- We apologize for the confusion. This is indeed an additive noise version of our meta-dropout. We updated the revision with the full description of the method in Appendix B. Since we compare against this additive version in the main table (Table 2), which is already an ablation study of the two, it is excluded from the Ablation study in Table 3. \\n\\n\\n4. Overall, it seems that paper falls short of clearly proving that back-propagating through MAML to the noise parameters is helpful. \\n\\n- The slightly good performance of \\u201cDeterministic Meta-dropout\\u201d and \\u201cFixed Gaussian\\u201d does not hurt our claim on the effectiveness of multiplicative noise, because multiplicative noise, by definition, consists of two parts: deterministic multiplication and pure random noise. For example, Bernoulli dropout consists of Bernoulli retain probability $0 \\\\leq p \\\\leq 1$ (deterministic multiplication) and actual random sampling (pure random noise). In this vein, we can say that \\u201cDeterministic Meta-dropout \\u201cdemonstrates the effectiveness of meta-learning the probability p, and \\u201cFixed Gaussian\\u201d shows the effectiveness of injecting (multiplying) pure random noise N(0,I) on each feature location.\\n\\nOverall, our meta-dropout combines the two components, which are complementary each other, into a novel input-dependent form of perturbation function. In this regard, we have clearly demonstrated that meta-learning input-dependent multiplicative noise is beneficial for improving generalization, jointly as well as component-wisely.\"}", "{\"title\": \"Summary of updates in the revision\", \"comment\": \"We thank all reviewers for their constructive comments. Here we briefly mention what have been updated in the revision. For more detailed explanations, please refer to the response to each reviewer.\\n\\n1. New experimental results on adversarial robustness: we replaced the FGSM attack with PGD attacks with $L_1$, $L_2$, $L_\\\\infty$ norm, and included more baselines (e.g. Mixup, VIB, and Information Dropout). The results show that our meta-dropout yields deep neural networks that are significantly more robust to adversarial attacks than baselines, regardless of the types of attacks used ($L_1$, $L_2$, $L_\\\\infty$) on few-shot classification tasks. We believe that these results will significantly strengthen our work, as we have shown that our meta-dropout not only achieves models that generalize better, but are more robust. While these findings are from few-shot classification tasks, we believe that they are meaningful, since existing models for adversarial learning achieved robustness at the expense of generalization accuracy and most of them do not generalize across different types of attacks. Please see the paragraph \\u201cAdversarial robustness\\u201d and Figure 5 in Page 7, as well as our response to the R1\\u2019s comment below for more detailed explanations.\\n\\n2. In Section 3.2, based on R1\\u2019s comment, we toned down on our claims. We changed \\u201cLearning to regularize variational inference\\u201c into \\u201cConnection to variational inference\\u201d and corrected corresponding sentences that look overstated. \\n\\n3. In Table 3, for each baseline, we added in the indication on whether it has the following properties: Random sampling, Learned multiplication, and Input-dependency. This allows us to better see the effect of each component to the generalization performance. We also denote that the input-dependent meta-learning of multiplicative noise with stochasticity as the core components of our meta-dropout. We provide more in-depth analysis on \\u201cAblation study\\u201d paragraph in Page 8.\"}", "{\"title\": \"Response to Reviewer #3 (2/2)\", \"comment\": \"5. It seems like Equation 7 is wrong because y_i is missing from the second argument of the KL divergence term. The transition to Equation 8 is therefore also wrong, and as far as I can understand, the whole argument breaks down.\\n\\n- This is a critical misunderstanding. Equation 7 is correct. In Equation 7, the second argument p(z|x) is called a \\u201cconditional prior\\u201d, which does not observe the label y since p(z|x) is part of the generative process. See the Equation 4 of Sohn et al. [1] which is exactly the same as the Equation 7 of our paper.\", \"it_seems_that_you_are_confused_with_the_following_decomposition_of_log_evidence\": \"Log evidence = ELBO + $KL[q(z|x,y)\\\\|p(z|x,y)]$,\\t\\t (a)\\nELBO = E[Log-likelihood] - $KL[q(z|x,y)\\\\|p(z|x)]$\\t\\t(b)\\n \\nThe log evidence term is decomposed of ELBO and the KL term involving $p(z|x,y)$ in Eq. (a). However, what Eq.7 in the main paper describes is the actual ELBO expression in Eq.(b), which contains the KL divergence between the approximate posterior and the prior $p(z|x)$.\", \"reference\": \"[1] Sohn et al., Learning Structured Output Representation using Deep Conditional Generative Models, In NIPS, 2015.\\n \\n\\n6. Line 7 in Algorithm 1 in Appendix A (which by the way should really be in the main text) does not make sense. The second sentence of the abstract is not implied by the first, the usage of \\u201cthus\\u201d does not seem appropriate. The intro should probably mention L1 and L2 regularization as well. \\u201cmeta-droput\\u201d, \\u201crobustenss\\u201d: typos in many places.\\n\\n- Thank you for pointing them out. We updated the revision based on your suggestions. \\n\\n\\n7. Figure 4 visualization is not clear.\\n\\n- Figure 4 visualizes the perturbations generated by our Meta-Dropout and the decision boundary for classification. Could you provide us more specific comments on which part of Figure 4 is not clear? \\n\\n\\n8. The architectural change required to add noise is not explained in the paper (i.e. what is \\\\phi and how it\\u2019s used).\\n\\n- The architectural change and \\\\phi are clearly explained in Figure 2 and the paragraph \\u201cform of the noise\\u201d at the end of page 4.\\n\\n\\n9. Additional baselines\\n\\n- We conduct experiments on additional baselines as requested:\\n(1) MetaReg: MAML + L1 regularization\\n(2) MAML (R3): \\\\phi treated as \\\\theta, as R3 requested\\n\\nModels\\t\\t\\t\\tOmni-1shot\\tOmni-5shot\\tmimg-1shot\\tmimg-5shot\\nMAML\\t\\t\\t\\t95.23+-0.17\\t98.38+-0.07\\t49.58+-0.65\\t64.55+-0.52\\nMetaReg (Ours)\\t\\t95.28+-0.15\\t98.85+-0.06\\t49.76+-0.67\\t65.42+-0.53\\nMAML (R3)\\t\\t\\t96.15+-0.15\\t98.69+-0.07\\t42.08+-0.63\\t62.82+-0.53\\nMeta-dropout\\t\\t96.63+-0.13\\t98.73+-0.06\\t51.93+-0.67\\t67.42+-0.52\\n \\nAs shown, all the suggested baseline models largely underperform Meta-Dropout.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We sincerely appreciate your constructive comments. We respond to your main concerns below:\\n\\n1. The Deterministic Meta-Dropout performs better than baseline MAML - maybe it's an effect of a larger number of parameters in the model?\\n\\n- To demonstrate that strong generalization performance of Meta-Dropout is not the effect of using larger number of model parameters, we doubled the number of channels for the base model and report its performances (MAML(x2)).\\n\\nModels\\t\\t #param.\\tOmni-1shot\\tOmni-5shot\\tmimg-1shot\\tmimg-5shot\\nMAML\\t\\t x1\\t \\t95.23+-0.17\\t98.38+-0.07\\t49.58+-0.65\\t64.55+-0.52\\nMAML(x2)\\t x4\\t \\t94.96+-0.16\\t98.36+-0.08\\t48.19+-0.64\\t65.84+-0.52\\nMeta-SGD x2\\t \\t96.16+-0.14\\t98.54+-0.07\\t48.30+-0.64\\t65.55+-0.56\\nMeta-dropout x2\\t \\t96.63+-0.13\\t98.73+-0.06\\t51.93+-0.67\\t67.42+-0.52\\n\\nThe number of parameters of MAML(chx2) is four times of that of MAML, while Meta-dropout is only doubled. Nonetheless, MAML(chx2) does not improve on MAML, demonstrating that the effectiveness of meta-dropout does not simply come from using larger number of parameters. Meta-SGD also doubles the number of parameters in the base MAML model, but is significantly outperformed by Meta-dropout. \\n\\nWe want to emphasize that Deterministic meta-dropout is also one of our models, and that its good performance does not hurt our claim on the effectiveness of the multiplicative noise. This is because meta-dropout consists of two parts: meta-learned deterministic multiplicative perturbation and random noise. Thus the deterministic meta-dropout still \\u201clearns to perturb\\u201d, although not random, and is actually a core component of meta-dropout (See Table 3 in the revision). Please also see our response to the Reviewer #3, comment #4.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We really appreciate your constructive comments. We respond to each comment as follows.\\n\\n1. Meta dropout does not regularize the variational framework because there is no variational inference framework.\\n\\n- Thank you for your comment. We agree with you that the current lower bound is not a variational form due to the assumption of q=p. In Section 3.2, we toned down the original expression \\u201cLearning to regularize variational inference\\u201c into \\u201cConnection to variational inference\\u201d, and corrected the corresponding sentences. Still, there exists a clear connection between standard variational inference and our learning framework. Thus we believe that discussion in Section 3.2 will be helpful to readers who want to understand the meaning of learning objective Eq.(2) in depth.\\n\\n2. Improving adversarial robustness experiment.\\n\\n- Thank you for the helpful suggestion. During the rebuttal period, we conducted additional experiments on adversarial robustness as you suggested:\\n\\na) We replaced the previous FGSM attack with stronger PGD attack (200 iter.), with $L_1$, $L_2$, and $L_\\\\infty$ norm constraints. \\n\\nb) We included more baselines (e.g. Mixup, VIB, and information dropout), and show that our meta-dropout largely and consistently outperforms all of them.\\n\\nc) We added more detailed descriptions of the adversarial meta-learning baseline and in-depth analysis on the results.\\n\\nd) We further show that the learned perturbation from our Meta-dropout also generalize across different types of adversarial attacks with $L_1$, $L_2$, and $L_\\\\infty$ attacks. The generalization to different types of attacks is an important problem in adversarial learning, and most existing models fail to achieve this goal. \\n\\nPlease see the corresponding section in the revision. We believe that the adversarial robustness part of our paper has become much stronger than before, thanks to your suggestion.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposes meta dropout, which leverages adaptive dropout training for regularizing gradient based meta learning models, e.g., MAML and MetaSGD. Experiments on few shot learning show that meta dropout achieves better performance.\\n\\nOverally, I think this paper is well motivated and experiments on few shot learning are impressive. I have only two major concerns.\\n\\n1. Sec 3.2. According to my understanding, Meta dropout introduces a learnable prior for latent $z$, but the training objective does not require posterior inference and thus no variational inference is needed. I think it is ok to say that meta dropout tries to optimize a lower bound of log p(Y|X;\\\\theta,\\\\phi^*), but meta dropout does not regularize the variational framework because there is no variational inference framework.\\n\\n2. Experiments on adversarial robustness can be further improved. (1) the settings and the analysis of adversarial robustness experiment can be discussed in details. For example, how to build ''adversarial learning baseline'' in meta learning settings and why the result implies the perturbation directions for generalization and robustness relates to each other; (2) how other regularization methods (e.g., Mixup, VIB and Information dropout) perform on adversarial robustness? Does Meta dropout performs better than them? (3) FGSM is a quite weak adversarial attack method, which makes evaluating adversarial robustness on FGSM may be misleading. I suggest trying some other STOA attack methods (e.g., iterative methods).\", \"some_typos\": \"Page 3, Regularization methods, 3rd line, ````wwwdiscuss\\nPage 7, 2nd line from the bottom, FSGM->FGSM\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors propose to meta-learn, using MAML, the mean of an elementwise, input-dependent, multiplicative noise to improve generalization in few-shot learning.\\nThe motivation is that meta-learning the noise allows to learn how to best perturb examples in order to improve generlization. This claim is supported by ample experimental evidence and comparisons against many baselines, as well as additional ablation studies w.r.t design choices of the algorithm itself. The paper is well written and easy to read. Consequently, I think this is a nice paper and should be accepted. \\n\\nEdit (leaving everything else unchanged for now): After reading R3's assessment, I agree with them that it's worrying that the Deterministic Meta-Dropout performs better than baseline MAML - maybe it's an effect of a larger number of parameters in the model?\", \"edit\": \"Thank you for your response.\\n\\nI will leave my score as is.\\n I would strongly encourage the authors to incorporate the baseline \\\"(1)\\\" as proposed by R3 in a future version of the paper as I agree with them that this is a relevant baseline.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper proposes learning to add input-dependent noise to improve the generalization of MAML-style meta-learning algorithm. The proposed method is evaluated on OmniGlot and miniImageNet. The paper reports improvements upon MAML, MAML with meta-learned parameter-wise learning rates, as well as a few regularization methods that are based on input/hidden state perturbations (Mixup, Variational Information Bottleneck). An ablation study also compares the proposed meta-dropout algorithm with a number of modifications, such as a fixed noise, input-independent noise, etc. It is furthermore shown that meta-dropout somewhat improves the model\\u2019s robustness against an adversarial attack.\\n\\nThe paper is somewhat incremental considering that Li et al, (2017) and Balaji et al, (2018) have already proposed meta-learning parameter-wise learning rates and parameter-wise regularization coefficient respectively. One difference from the methods above is that in the proposed method noise is controlled by the input. The ablation however shows that in 5-shot classification case simply adding non-trainable noise works quite well. \\n\\nIt seems like the choice of the particular method for adding the noise was performed using the test set. If it\\u2019s true, this is methodologically wrong: model selection should be performed on a development set (or meta-development) set. Futhermore, Table 2 contains some results named \\u201cAdd.\\u201d, which I guess stands for additive noise. I did not find an explanation of what is the specific method for adding noise used in this case. Such additive noise is also missing from ablation experiments. \\n\\nOverall, it seems that paper falls short of clearly proving that back-propagating through MAML to the noise parameters is helpful. The \\u201cDeterministic Meta-Dropout\\u201d performs better than baseline MAML, and arguably, meaning that some part of the improvement upon MAML can be due to the architectural differences and not due to noise. \\u201cIndependent Gaussian\\u201d and \\u201cWeight Gaussian\\u201d baselines perform worse than non-trainable noise (\\u201cFixed Gaussian\\u201d). Learning the variance for the noise is shown to be detrimental. There is just too much confusion in the results, the improvements are not very robust. \\n\\nThe paper writing is okay, but there are serious issues. I am not sure I understand the argument in Section 3.2 that meta-dropout performs variational inference. It seems like Equation 7 is wrong because y_i is missing from the second argument of the KL divergence term. The transition to Equation 8 is therefore also wrong, and as far as I can understand, the whole argument breaks down. Line 7 in Algorithm 1 in Appendix A (which by the way should really be in the main text) does not make sense.\", \"other_issues\": [\"the second sentence of the abstract is not implied by the first, the usage of \\u201cthus\\u201d does not seem appropriate\", \"the intro should probably mention L1 and L2 regularization as well\", \"in Section 3.1 there is a forward reference to Equation 5, makes understanding the text quite hard\", \"\\u201cmeta-droput\\u201d, \\u201crobustenss\\u201d: typos in many places\", \"Figure 4 visualization is not clear.\", \"the architectural change required to add noise is not explained in the paper (i.e. what is \\\\phi and how it\\u2019s used)\", \"no comparison to meta-learned L1 regularization\", \"a baseline is missing in which \\\\phi is treated as a part of \\\\theta and trained with vanilla MAML\"]}" ] }

[ "Xinyu Zhang", "Qiang Wang", "Jian Zhang", "Zhao Zhong" ]

[ "Automatic Data Augmentation", "Adversarial Learning", "Reinforcement Learning" ]

{ "note_id": [ "A0pUgJ715c", "HJx91T8hjr", "SJgRylr2sr", "HygeSPE3jr", "r1el_7gojr", "SJepTa1jsS", "S1lbi609iS", "ryeaPUKKsH", "B1l0K2oLjS", "rJlknssIjH", "BkxK-ssIjr", "SJxZB33S5S", "SkeDSmkQ5r", "rkg8m0BKtr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731110, 1573838050120, 1573830630140, 1573828407877, 1573745512514, 1573744068710, 1573739929053, 1573652068724, 1573465222069, 1573464999080, 1573464833272, 1572355129200, 1572168510685, 1571540509879 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1734/Authors" ], [ "ICLR.cc/2020/Conference/Paper1734/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1734/Authors" ], [ "ICLR.cc/2020/Conference/Paper1734/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1734/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1734/Authors" ], [ "ICLR.cc/2020/Conference/Paper1734/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1734/Authors" ], [ "ICLR.cc/2020/Conference/Paper1734/Authors" ], [ "ICLR.cc/2020/Conference/Paper1734/Authors" ], [ "ICLR.cc/2020/Conference/Paper1734/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1734/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1734/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper proposes a method to learn data augmentation policies using an adversarial loss. In contrast to AutoAugment where an augmentation policy generator is trained by RL (computationally expensive), the authors propose to train a policy generator and the target classifier simultaneously. This is done in an adversarial fashion by computing augmentation policies which increase the loss of the classifier. The authors show that this approach leads to roughly an order of magnitude improvement in computational cost over AutoAugment, while improving the test performance.\\nThe reviewers agree that the presentation is clear and that the proposed method is sound, and that there is a significant practical benefit of using such a technique. As most of the concerns were addressed in the discussion phase, I will recommend acceptance of this paper. We ask the authors to update the manuscript to address the remaining (minor) concerns.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"Thank you very much. The helpful discussion improves the paper substantially. We will consider your suggestion.\"}", "{\"title\": \"Response to Authors\", \"comment\": \"Thank you for your rebuttal and insightful comments. I think my main concerns have been addressed, and as such I will adjust my score accordingly.\\n\\nI'm not sure if there is time left to respond in the rebuttal period, but I was wondering if the learned augmentation policies could be distilled into a simple script such as done for AutoAugment in [1]? Since training the policy network is still quite computationally expensive and likely out of reach for the majority of practitioners, a data augmentation script with pretrained policies would likely be very well received.\\n\\n[1] https://github.com/DeepVoltaire/AutoAugment\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"Thanks for your positive comments on our work.\\n\\n>>> Response to Q1:\\nFrom our point of view, extreme learning rate or entropy penalty values may cause rapid convergence, which is unexpected. However, we do not think that equals to the collapsing results you worried about. We didn\\u2019t make specific experiments on those extreme hyperparameters, but we would like to make a deduction based on our understanding of the work: \\nFirst, all magnitudes are set in a moderate range as mentioned before, which protect the network from collapsing during training even if they are picked at boundary.\\nThen, with the training processes, loss of target network decreases gradually, which weaken the reward signal while converging policy network. Hence, extreme values are avoided during training.\\n\\nIn Figure 6, the probability distribution of the parameters is visualized over 600 epochs (Appendix A1). We think this long-epoch training has made the policy network steadily converged, which can be observed from Figure 6(b). This indicates that the policy network may not collapse to extreme values given more epochs. \\n\\n>>> Response to Q2: \\nWe think the sensitivity analysis of the hyperparameters is benefit to consolidate our work, which would be considered in our future work due to the space and time limits.\\nThe setting of the hyperparameters is almost the same as AutoAugment. The extensive results have shown that it also works well on different datasets in our method, including CIFAR-10/100 and ImageNet. Hence, we have reason to believe that it probably could be used for other similar datasets. However, without exhaustive experiments, it is very difficult for us to determine whether it is suitable for any new dataset. If it is unfortunate that the hyperparameters are needed to be re-tuned for new datasets, we think AutoAugment perhaps will face the same problem during the policy search as well. According to the efficiency comparison between AutoAugment and our method in Table 5, we believe that the tuning of the hyperparameters for new datasets would not negate the gain in efficiency over AutoAugment. \\n\\nHope our response answers your questions.\"}", "{\"title\": \"Response to Authors\", \"comment\": \"Thank you for adding the additional plots, they are exactly what I was looking for.\\n\\nI am surprised that this method works so well without the policy network collapsing to just predicting extreme values. Although it does seem to be slowly converging towards using more extreme augmentations as training progresses. This leads me to a few more questions:\\n\\n1. Would the policy network collapse to extreme values if the learning rate is too large/entropy penalty value too small, or if the number of epochs is increased?\\n\\n2. How sensitive is the approach to the values of the controller's learning rate and the entropy penalty value? Can the same values used in the paper be used for any new dataset, or would they need to be re-tuned? Needing to tune these hyperparameters for each new dataset would be quite computationally expensive, and might negate any gains in efficiency over AutoAugment.\"}", "{\"title\": \"Response to authors\", \"comment\": \"Thanks for making these changes, they address my comments.\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"Thank you for your valuable comments and suggestions. Hope your concerns can be well addressed in the new revision.\\n\\n>>> Response to \\u201cWhether the policy network is actually learning useful policies?\\u201d:\\nWe have visualized the probability distribution of the parameters in the learned augmentation policies on CIFAR-10 over training epochs in the new revision, and also added the analysis as follows:\\n\\n\\u201cAs shown in Figure 6(a) and 6(b), we further visualize the probability distribution of the parameters of the augmentation policies learned with PyramidNet+ShakeDrop on CIFAR-10 over time. \\u2026. This indicates that our method does not simply learn the transformations with the extremes of the allowed magnitudes to spoil the target network.\\u201d \\n\\nThe revision above indicates that the policy network does learn useful policies along with the training process rather than simply learn the policies with the extremes of the allowed magnitudes to spoil the target network.\"}", "{\"title\": \"Response to Authors\", \"comment\": \"Thank you for taking the time to address my concerns. Please see some comments below.\\n\\n>>> Response to \\u201cWhether the policy network is actually learning useful policies?\\u201d:\\nThe main reason I am skeptical about this goes back to the original AutoAugment [1] paper. In that paper they constrain the range of magnitudes for some transformation such that they do not produce images that might be out of the distribution of normal training images. This is a reasonable precaution to take. However, it turns out that those ranges were very important, since it was later shown that one did not even need to learn useful policies; randomly selecting some transformations from within the restricted range of magnitudes was good enough [2]. In essence, the very expensive policy search computation can be skipped if the pool of available transforms is restricted to a useful range, in which case we can revert to simple random sampling of augmentations.\\n\\nMy main concern then is that the policy network in Adversarial AutoAugment is simply learning to apply transformations that are towards the extremes of the allowed ranges, since extreme transforms would be the most likely to fool the classifier. If we can achieve similar performance gains to Adversarial AutoAugment by simply restricting the augmentation pool to more extreme transformations, then we can again skip the computationally expensive search.\\n\\nIf you can show that Adversarial AutoAugment is not simply learning to apply transformations that are on the extreme ranges of the allowed magnitudes, I would feel better about accepting this paper.\\n\\n[1] Cubuk, Ekin D., Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. \\\"Autoaugment: Learning augmentation policies from data.\\\" CVPR (2019).\\n\\n[2] Cubuk, Ekin D., Barret Zoph, Jonathon Shlens, and Quoc V. Le. \\\"RandAugment: Practical data augmentation with no separate search.\\\" arXiv preprint arXiv:1909.13719 (2019).\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"We want to express our deep gratitude to for the constructive suggestions and positive comments on the novelty, motivation, and performance. We will explain your concerns point by point and hope you find our new revision satisfactory.\\n\\n>>> Response to \\u201cWhether the policy network is actually learning useful policies?\\u201d:\\n1)\\tIn this paper, the generalization of the target network is improved by finding the best data augmentation policy to perform label-preserved image transformation. Although adversarial augmentation policies are generated by the policy network, the extreme and unmeaning image transformation that will almost destroy all the image information, such as making the entire image black or white, should not be included in the search space. This can be regarded as the guideline to determine the acceptable range of the magnitude of all the operations. How to handle extreme image transformations will be considered in our future works.\\n2)\\tWe think that the ablation study has shown the effectiveness of adversarial augmentation policies. Through training the target network with adversarial augmentation policies, our method achieves the best test error, which indicates that the policy network can generate adversarial policies which are more adaptive to the training process, rather than apply extreme transformations which always fool the target network.\\n\\n>>> Response to \\u201cgrammatical errors and typos\\u201d:\\nSincerely apologize for these mistakes. We have corrected these grammatical errors and typos, and revised the paper accordingly. Following your kind advice, we have also invited some native speakers to proofread the paper carefully. Wish you find our new revision satisfactory. Thanks again for pointing out our mistakes.\\n\\n>>> Response to \\u201cclaim about weak transferability of human-designed augmentation policies\\u201d: \\nWe are very sorry for the misunderstanding. We just want to claim that human-designed augmentation policies sometimes show the weak transferability across different datasets, which is also presented in [1]. We have modified it in the new version.\\n\\n>>> Response to \\u201cNo citation for the original GAN paper\\u201d: \\nThank you for the reference, which has been included in the new version.\\n\\n>>> Response to \\u201cNo computation time comparison with PBA\\u201d:\\nBecause the experiment analysis on ImageNet dataset is not provided in PBA, we don\\u2019t compare computation time with it in Table 5. Although PBA is 1000x faster than AutoAugment in term of the searching cost, the cost of target network training is still non-trivial. Besides, we think that the comparison of computation time on large-scale tasks is more meaningful. Hence, we compare the total computing cost and time overhead in training ResNet-50 on ImageNet dataset between AutoAugment and our method.\\n\\n>>> Response to \\u201cCIFAR-10 results are not necessarily state-of-the-art\\u201d: \\nWe are very sorry for the omission. Due to the trivial difference, we think that the Top1 error of 1.36% on CIFAR-10 is comparable to the SOTA results in [2]. Meanwhile, we demonstrate that our overall performance is largely robust (insensitive) to the random factor due to the small variances on five runs in the revised version.\\n\\n>>> Response to \\u201ccitation error\\u201d:\\nApologize for the typo. We have fixed it in the revised version. \\n \\n[1] Cubuk, Ekin D., Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. \\\"Autoaugment: Learning augmentation policies from data.\\\" CVPR (2019).\\n\\n[2] Wistuba, Martin, Ambrish Rawat, and Tejaswini Pedapati. \\\"A Survey on Neural Architecture Search.\\\" arXiv preprint arXiv:1905.01392 (2019).\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"We want to express our deep gratitude to for the constructive suggestions and positive comments on the novelty, motivation, and performance.\\n\\n>>> Response to \\u201cincremental contributions\\u201d: \\nWe are very sorry for the misunderstanding. There are several reasons that strongly support for our substantial contributions.\\na)\\tTraditional GANs are used to enlarge datasets through directly synthesizing new images. Although our method also plays in a GANs setting to jointly optimize target network and policy search, our goal is to find the best augmentation policy to perform label-preserved image transformations, rather than synthesize new images.\\nb)\\tThe exciting performance achieved by AutoAugment shows the potential benefits of automated data augmentation for training DNNs. However, the process of augmentation policy search in AutoAugment is very computationally expensive, and these augmentation policies learned on proxy tasks are not guaranteed to be optimal on target tasks. In this paper, we propose an adversarial framework to address these drawbacks. Through reusing the computing resource of target network training, the computing cost of policy search can be extremely degraded, as shown in Table 5. Directly learning policies on target tasks avoids the performance degradation caused by the transfer from proxy tasks to target tasks.\\n\\n>>> Response to \\u201cBenefits of the GAN approach against pre-training policies\\u201d: \\nWe are very sorry for confusing you. We think that the reported results have shown the benefits of the GAN approach against pre-training policies. Taking ResNet-50 trained on ImageNet dataset as an example, AutoAugment can achieve a drop of 1.32% in Top-1 error compared to baseline with the 15000 GPU*hours of consumption. However, our method achieves a drop of 3.09% in Top-1 error compared to baseline only with the (1280-160=1120) GPU*hours of additional consumption, which is even 1.77% better than AutoAugment. The transferability further shows that our method is more computing-efficient.\\n\\n>>> Response to \\u201cMore complete results\\u201d: \\na)\\tThe confidence interval with all the reported values and the Top-5 results on ImageNet have been provided in the new revision. We also list the table below:\", \"table_3\": \"Top-1 / Top-5 test error(%) on ImageNet\\n+-------------+--------------+--------------+-----+--------------------------------+\\n| Model | Baseline | AutoAugment | PBA | Our Method |\\n+-------------+--------------+--------------+-----+--------------------------------+\\n| ResNet-50 | 23.69 / 6.92 | 22.37 / 6.18 | - | 20.60$\\\\pm$0.15 / 5.53$\\\\pm$0.05 |\\n+-------------+--------------+--------------+-----+--------------------------------+\\n| ResNet-50-D | 22.84 / 6.48 | - | - | 20.00$\\\\pm$0.12 / 5.25$\\\\pm$0.03 |\\n+-------------+--------------+--------------+-----+--------------------------------+\\n| ResNet-200 | 21.52 / 5.85 | 20.00 / 4.90 | - | 18.68$\\\\pm$0.18 / 4.70$\\\\pm$0.05 |\\n+-------------+--------------+--------------+-----+--------------------------------+\\n\\nb)\\tWe have tried our best to evaluate our method on the same networks as Cubuk19 except AmoebaNet.\\nThis is because our baseline training of AmoebaNet cannot achieve the reported accuracy in the original paper. \\nThe problem is also reported in [1].\\n\\n>>> Response to \\u201ccouple the learned policies\\u201d: \\nGenerally, we don\\u2019t couple the learned policies again after target network training. This is because our method can directly perform the augmentation policy search along with target network training with a relatively low computing cost.However, to make our method more computing-efficient, we have shown the transferability of the learned policies between datasets and architectures in the new version.\\n\\n[1] Chen Lin, Minghao Guo, Chuming Li, Wei Wu, Dahua Lin, Wanli Ouyang, and Junjie Yan. Online\\nhyper-parameter learning for auto-augmentation strategy. CoRR, abs/1905.07373, 2019.\"}", "{\"title\": \"Response to Review #4\", \"comment\": \"We want to express our deep gratitude to for the constructive suggestions and positive comments on the novelty, motivation, and performance.\\n\\n>>> Response to \\u201cevaluating transferability\\u201d: \\nThanks for your constructive advice. We have added the transferability analysis of the learned policies in the new version. Although the augmentation policies are dynamically changed in an adversarial manner, the learned policies can still transfer well to different datasets and architectures. Through directly applying the learned dynamic policies to various networks, the competitive performance is also achieved.\\n\\n>>> Response to \\u201ca measure of uncertainty\\u201d: \\nFollowing your kind suggestion, we have stated the confidence interval with all the reported values in the new version. We also list the table below:\", \"table_1\": \"Top-1 test error(%) on CIFAR-10\\n+-------------------------+----------+--------+-------------+------+---------------+\\n| Model | Baseline | Cutout | AutoAugment | PBA | Our Method |\\n+-------------------------+----------+--------+-------------+------+---------------+\\n| Wide-ResNet-28-10 | 3.87 | 3.08 | 2.68 | 2.58 | 1.90$\\\\pm$0.15 |\\n+-------------------------+----------+--------+-------------+------+---------------+\\n| Shake-Shake (26 2x32d) | 3.55 | 3.02 | 2.47 | 2.54 | 2.36$\\\\pm$0.10 |\\n+-------------------------+----------+--------+-------------+------+---------------+\\n| Shake-Shake (26 2x96d) | 2.86 | 2.56 | 1.99 | 2.03 | 1.85$\\\\pm$0.12 |\\n+-------------------------+----------+--------+-------------+------+---------------+\\n| Shake-Shake (26 2x112d) | 2.82 | 2.57 | 1.89 | 2.03 | 1.78$\\\\pm$0.05 |\\n+-------------------------+----------+--------+-------------+------+---------------+\\n| PyramidNet+ShakeDrop | 2.67 | 2.31 | 1.48 | 1.46 | 1.36$\\\\pm$0.06 |\\n+-------------------------+----------+--------+-------------+------+---------------+\", \"table_2\": \"Top-1 test error(%) on CIFAR-100\\n+------------------------+----------+--------+-------------+-------+----------------+\\n| Model | Baseline | Cutout | AutoAugment | PBA | Our Method |\\n+------------------------+----------+--------+-------------+-------+----------------+\\n| Wide-ResNet-28-10 | 18.80 | 18.41 | 17.09 | 16.73 | 15.49$\\\\pm$0.18 |\\n+------------------------+----------+--------+-------------+-------+----------------+\\n| Shake-Shake (26 2x96d) | 17.05 | 16.00 | 14.28 | 15.31 | 14.10$\\\\pm$0.15 |\\n+------------------------+----------+--------+-------------+-------+----------------+\\n| PyramidNet+ShakeDrop | 13.99 | 12.19 | 10.67 | 10.94 | 10.42$\\\\pm$0.20 |\\n+------------------------+----------+--------+-------------+-------+----------------+\", \"table_3\": \"Top-1 / Top-5 test error(%) on ImageNet\\n+-------------+--------------+--------------+-----+--------------------------------+\\n| Model | Baseline | AutoAugment | PBA | Our Method |\\n+-------------+--------------+--------------+-----+--------------------------------+\\n| ResNet-50 | 23.69 / 6.92 | 22.37 / 6.18 | - | 20.60$\\\\pm$0.15 / 5.53$\\\\pm$0.05 |\\n+-------------+--------------+--------------+-----+--------------------------------+\\n| ResNet-50-D | 22.84 / 6.48 | - | - | 20.00$\\\\pm$0.12 / 5.25$\\\\pm$0.03 |\\n+-------------+--------------+--------------+-----+--------------------------------+\\n| ResNet-200 | 21.52 / 5.85 | 20.00 / 4.90 | - | 18.68$\\\\pm$0.18 / 4.70$\\\\pm$0.05 |\\n+-------------+--------------+--------------+-----+--------------------------------+\\nThe small variances on 5 models demonstrate that our overall performance is largely robust (insensitive) to the random factor.\\n\\n>>> Response to \\u201cgrammatical errors and typos\\u201d: \\nApologize for these errors and typos. We have corrected these grammatical errors and typos, and revised the paper accordingly.\\n\\n>>> Response to \\u201cthe breadth of the claim\\u201d: \\nThanks for pointing out it. We have modified the claim in the new version.\\n\\n>>> Response to \\u201cFigure 4\\u201d: \\nWe have added the axis units on the figure.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper describes a method to learn data augmentation policies using an adversarial loss. It builds on the AutoAugment method. In AutoAugment, an augmentation policy generator is trained by reinforcement learning. At each iteration, a classifier network is trained from scratch based on the current augmentation policy, and its validation accuracy is used as the reward signal. This is extremely costly because it requires to train a complete network for every training step of the policy generator. Instead, the current paper proposes to train the policy generator and the classifier simultaneously. The policy generator is trained adversarially to find augmentation policies that increase the loss of the classifier. This leads to a significant speedup compared to classic AutoAugment.\\n\\nThe presentation of the algorithm and the results is very clear. The proposed method yields improved performance compared to AutoAugment at ~1/10 of the computational cost, which is impressive. Although the paper only evaluates on two datasets (CIFAR and ImageNet), the idea is likely applicable very generally. I recommend this paper for publication, but have some comments that should be addressed:\", \"major_comments\": [\"It would be good to evaluate how well the learned policies transfer between datasets and architectures. Adversarial AutoAugment still comes with a significant computational cost compared to hand-crafted augmentation, so transfer of policies would be useful. AutoAugment is transferable by design, so any competing algorithm should evaluate transferability.\", \"The authors state that all results are mean of 5 initializations, which is great. Please use these replicates to compute a measure of uncertainty (SEM or confidence interval) and state this with all values in the tables.\"], \"minor_comments\": [\"Overall, there are many grammatical errors and typos that sometimes require interpretation and reduce clarity. Please proof-read carefully.\", \"Abstract sentence \\u201c... can simultaneously optimizes\\u2026\\u201d has grammatical issues.\", \"Second sentence of introduction has grammatical issues.\", \"Third sentence of intro: \\u201cet al.\\u201d is used for persons, use \\u201cetc.\\u201d for things.\", \"Contribution section: \\u201c...our proposed method outperforms all previous augmentation method.\\u201d Please be careful with the breadth of your claims. You do not compare against *all* previous augmentation methods.\", \"Figure 4: Please add units and/or refer to Table 5 in the legend.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": [\"The authors propose a method for adversarial data augmentation which jointly trains the target network and the augmentation policy network. The authors claim that such learning set-up prevents overfitting and reduces computational cost with respect to competitors. Finally they provide extensive results and show that outperform the state-of-the-art in multiple datasets.\", \"The rationale behind the idea is properly introduced and justified\", \"The formulation is clear and sound\", \"The results show improvements over competitors in terms of accuracy and computational cost\", \"The contributions seem very incremental. Applying GANs for data augmentation is not new and regarding the policy search the authors strongly base their approach in Cubuk19. The main differentiator wrt the work of Cubuk is to jointly train both target and policy learner in a GAN setting, rather than learning the policies a priory with a fixed target network.\", \"I miss further discussion regarding the benefits of the GAN approach against pre-training policies. The overfitting argument seems pretty weak given the relatively marginal gains of accuracy.\", \"It would make the experiments more complete if the Top-1 and Top-5 results were provided, as well as running experiments on the same networks as the Cubuk19 paper.\", \"How coupled are the policies learned to the specific dataset after training?. The work by Cubuk19 shows transferability properties that could somehow diminish the gain achieved by this work regarding computational cost.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes a technique called Adversarial AutoAugment which dynamically learns good data augmentation policies during training. An adversarial approach is used: a target network tries to achieve good classification performance on a training set, while a policy network attempts to foil the target network by developing data augmentation policies that will produce images that are difficult to classify. To train the policy network, each mini-batch is augmented multiple times with different policies sampled from the policy network. Each augmented mini-batch is passed through the target network to produce a corresponding training loss, which is used as the training signal for the policy network. Experimental results are shown for CIFAR-10, CIFAR-100, and ImageNet datasets, where Adversarial AutoAugment outperforms competing methods (AutoAugment and Population Based Augmentation) on a variety of model architectures.\\n\\nIn its current state, I would tend towards rejecting this paper. The overall structure, the figures, and the experimental results are very nice, but there are two major issues that are holding it back. First, I am skeptical that the policy network is actually learning useful policies. Secondly, there are many grammatical errors in the paper which hamper readability. Upon reading the first paragraph of the paper, my initial impression was already quite negative, simply due to the number of grammatical errors. Fixing these would strengthen the paper considerably, and I would increase my score accordingly if properly addressed.\", \"primary_concerns\": \"1) One of my main concerns with this paper is this line here: \\\"To guarantee the convergence during adversarial learning, the magnitude of all the operations are set in a moderate range\\\". Since the policy network has no incentive to select transformations that the target network can still learn from, I assume that the ranges are required so that the policy network cannot choose to apply extreme transformations which always fool the target network, such as setting brightness to 0 to make the entire image black. How are acceptable ranges determined? If cross validation is required, then this becomes very much like the original hand-tuning of data augmentations that we wanted to avoid in the first place. \\n\\nAdditionally, I think it would be useful to to see a plot of the magnitude of each transformation versus training epoch, similar to Figure 4a in [1]. If the policy network simply learns to use the most extreme augmentations available in order to fool the target network, then this may indicate that the gain in performance is from tuning the magnitude ranges, and not from the policy network selecting good policies.\\n\\n2) The paper could benefit greatly from some revision of the grammar. I would recommend either having a friend or colleague read it over, or even using an automated grammar checking program, such as Grammarly. For example, in the first paragraph alone there are several sentences that could be improved:\\n\\\"Massive amount of data promotes the great success\\\" -> \\\"Massive amounts of data have promoted the great success\\\"\\n\\\"when more supervised data available\\\" -> \\\"when more supervised data is available\\\"\\n\\\"or better data augmentation method adapted\\\" -> \\\"or a better data augmentation method is adopted\\\"\\n\\\"which can automated learn\\\" -> \\\"which can automatically learn\\\"\\n\\\"there still requires tens of thousands of GPU-hours consumption.\\\" -> \\\"tens of thousands of GPU-hours of computation are still required\\\"\", \"things_to_improve_the_paper_that_did_not_impact_the_score\": \"3) In the first paragraph it is claimed that data augmentation policies have weak transferability across different tasks and datasets. I do not fully agree with this claim, since papers such as AutoAugment [2] have shown that learned policies are highly transferable to new datasets, and augmentation strategies such as CutMix have been shown to be highly effective for a variety of tasks. \\n\\n4) No citation for the original GAN paper, despite multiple mentions of adversarial learning, and GANs themselves.\\n\\n5) There is no computation time comparison with PBA, which is about 1000x faster than Autoaugment, and therefore roughly 100x faster than Adversarial AutoAugment.\\n\\n6) CIFAR-10 results are not necessarily state-of-the-art. The 1.36% error rate claimed by the paper is surpassed by work in [3], which achieved 1.33% using a combination of AutoAugment and mixup. \\n\\n7) Citation for the CIFAR-10 dataset incorrectly refers to the Adam optimizer paper [4].\", \"references\": \"[1] Daniel Ho, Eric Liang, Ion Stoica, Pieter Abbeel, and Xi Chen. Population based augmentation: Efficient learning of augmentation policy schedules. ICML, 2019.\\n\\n[2] Cubuk, Ekin D., Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. \\\"Autoaugment: Learning augmentation policies from data.\\\" CVPR (2019).\\n\\n[3] Wistuba, Martin, Ambrish Rawat, and Tejaswini Pedapati. \\\"A Survey on Neural Architecture Search.\\\" arXiv preprint arXiv:1905.01392 (2019).\\n\\n[4] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015\", \"edit\": \"The authors have addressed the majority of my concerns, and as such I have increased my score from a 3 to a 6.\"}" ] }

[ "Rohan Taori", "Achal Dave", "Vaishaal Shankar", "Nicholas Carlini", "Benjamin Recht", "Ludwig Schmidt" ]

[ "robustness", "distribution shift", "image corruptions", "adversarial robustness", "reliable machine learning" ]

{ "note_id": [ "4WZsUiW01", "BkeiMMzBsB", "SkgTxwWSoS", "HyepRLZrsH", "rJgUJQbSoS", "SkxmTz-rir", "Bkl69MZHoB", "BkgJfUqRYH", "Hke8aQMpKS", "HkxRhsD5YS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731081, 1573360147177, 1573357300783, 1573357268519, 1573356254111, 1573356219258, 1573356180781, 1571886598708, 1571787710483, 1571613621961 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1733/Authors" ], [ "ICLR.cc/2020/Conference/Paper1733/Authors" ], [ "ICLR.cc/2020/Conference/Paper1733/Authors" ], [ "ICLR.cc/2020/Conference/Paper1733/Authors" ], [ "ICLR.cc/2020/Conference/Paper1733/Authors" ], [ "ICLR.cc/2020/Conference/Paper1733/Authors" ], [ "ICLR.cc/2020/Conference/Paper1733/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1733/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1733/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors show that models trained to satisfy adversarial robustness properties do not possess robustness to naturally occuring distribution shifts. The majority of the reviewers agree that this is not a surprising result especially for the choice of natural distribution shifts chosen by the authors (for instance it would be better if the authors compare to natural distribution shifts that look similar to the adversarial corruptions). Moreover, this is a survey study and no novel algorithms are presented, so the paper cannot be accepted on that merit either.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response\", \"comment\": \"We thank the reviewer for their comments and respond below to specific concerns:\\n\\n1) We disagree with the reviewer\\u2019s statement that ImageNetV2 is not distribution shift. The argument made by the ImageNetV2 paper is that despite the authors\\u2019 best efforts, ImageNetV2 is a slightly different distribution than the original ImageNet, and this suffices to cause the accuracy drop. Specifically, the authors of the ImageNetV2 paper conclude that\\n\\n\\u201cin spite of our efforts to match the original dataset creation process, the distribution gap is still our leading hypothesis for the accuracy drops. This demonstrates that it is surprisingly hard to accurately replicate the distribution of current image classification datasets.\\u201d [1]\", \"hence_imagenetv2_is_an_interesting_and_challenging_distribution_shift_for_two_reasons\": \"1) the shift is not easily formally characterized, in contrast to adversarial examples and common corruptions; and 2) this shift is a result of a real-world process, rather than a synthetic modification.\\n\\n\\n2) We agree with the reviewer that it is not a priori clear that adversarial robustness will be helpful outside a fully adversarial setting. Nevertheless, adversarial robustness has been connected to broader claims about robustness in various papers in the literature. Please refer to the list compiled at the end of our response to reviewer 1 for multiple papers making such claims.\\n\\nMoreover, we also find the question\\n\\n\\u201cDoes robustness on synthetic datasets and synthetic generations of such perturbations transfer to perturbations (of the same kind) but occurring in natural world\\u201d\\n\\nvery interesting. Including ImageNet-Vid-Robust in our testbed serves precisely this purpose, as the distribution shift is inspired by adversarial examples: the neighborhood of k frames contains images that are visually highly similar, and the evaluation metric is designed as the worst case performance in this neighborhood. This allows us to study comparable distribution shifts from both the synthetic and the natural world perspective. Unfortunately, we find that the synthetic robustness intervention does not meaningfully improve the robustness on the natural distribution shift. We believe that this lack of transfer from synthetic to natural distribution shift demonstrates an important shortcoming of current robustness interventions.\\n\\nMoreover, we would like to note that distribution shifts occurring in the real world do not always come with a precise characterization, as for example in ImageNetV2. In the absence of a precisely defined distribution shift, it becomes hard or impossible to design corresponding synthetic measures accurately. Thus, a natural question is whether existing robustness interventions at least partially transfer to unforeseen but non-adversarial changes in the data distribution. This is why evaluating current robustness measure on ImageNetV2 is interesting and important.\\n\\n\\n3) Finally, we would also like to address the reviewer\\u2019s comment\\n\\n\\u201cFinally, it seems like the novelty/contribution of this work is very small. The paper uses existing datasets and existing robustness measures.\\u201c\\n\\nWe disagree with this characterization of our work. Meta studies such as ours are a backbone of empirical research. The fact that our paper does not introduce new datasets or methods is deliberate and shared by almost all meta studies in the empirical sciences. Meta studies are intended to lead to reliable empirical information, as is especially needed in machine learning today. Rejecting a meta study on grounds that it lacks novel methodology or datasets goes against standard scientific practices.\\n\\n\\n\\n[1] Do ImageNet Classifiers Generalize to ImageNet? \\nBenjamin Recht, Rebecca Roelofs, Ludwig Schmidt, Vaishaal Shankar\", \"https\": \"//arxiv.org/abs/1902.10811, ICML 2019\"}", "{\"title\": \"Response to points 3 and 4 raised by the reviewer\", \"comment\": \"3) On video robustness vs JPEG robustness: There is indeed a correlation between the two robustness measures. However, this point is more nuanced since much of the correlation is driven by the lp-adversarially robust models. As pointed out in Section 5 of our paper, the models operate in a low accuracy regime where they are dominated by interpolating between a random classifier and a high accuracy model. So these models do not offer effective robustness in an interesting regime, and hence the correlation between video robustness and JPEG robustness is also substantially diminished.\\n\\nMore broadly, we disagree with the reviewer\\u2019s conjecture that \\u201cthe main conclusion of this paper might just be driven by the averaging of a large number of non-predictive perturbations and a smaller number of more predictive perturbations\\u201d. As can be seen from Figure 1, no robustness intervention currently offers substantial effective robustness on the natural distribution shifts we studied. Hence there can also be no highly predictive perturbations since there is simply no robustness to predict.\\n\\nNevertheless, we agree with the reviewer that these points and other scatter plots in the appendix of our paper are worth discussing in more detail and will do so in an updated version of the paper.\\n\\n\\n4) On ImageNet-P: We agree that ImageNet-P would make our testbed more complete, and we will work to incorporate the dataset. But as pointed out in our response to point 1 of the reviewer (see above), it is worth noting that our testbed is already substantially larger than any previous robustness evaluation. We refer the reader to point 1 above for a longer discussion of this issue.\"}", "{\"title\": \"Response to the first two points raised by the reviewer\", \"comment\": \"We would like to thank Reviewer 3 for their detailed comments. Below we respond to the specific concerns:\\n\\n\\n1) Regarding ImageNet-A: First, we would like to emphasize that our testbed is already substantially larger than any previous work on robustness evaluations. Due to size and quick growth of the robustness literature, it is nearly impossible to include all proposed evaluation settings. Hence we had to prioritize some datasets and models over others, and unfortunately some prior work could not be included in our submission. Nevertheless, we believe that our testbed already allows us to draw interesting conclusions about the state of robustness research in machine learning.\\n\\nWe look forward to continuing to maintain and grow our testbed for the community. In particular, we will work on incorporating ImageNet-A, ImageNet-P, hue, and translation perturbations as suggested by the reviewers. Moreover, we welcome suggestions for additional datasets or models for our testbed. However, since these evaluations take many days to run and involve collaborators in industry for some of the models, the results may not be complete in time for the rebuttal period. \\n\\nAdditionally, evaluating on ImageNet-A will not obviate the need for our measure of effective robustness. Though models currently have low accuracy on ImageNet-A, models with higher original accuracy will naturally have higher accuracy on ImageNet-A as well (as can be seen in [1]). This would represent a confounding factor, as it would now be unclear whether the gains in accuracy come from the model being more accurate or whether the gains come from extra robustness properties the model may exhibit. Thus, a metric that corrects for gains coming from simply having a higher original accuracy is preferable. Since empirically we find that original accuracy is an almost perfect predictor of natural distribution shift accuracy for many models (Figure 1), we can measure effective robustness as deviation away from this fit.\\n\\n\\n2) On synthetic perturbations from [2]: Of the synthetic perturbations mentioned by the reviewer, we currently do have \\u201csaturate\\u201d in our testbed (please see the plots on page 21, column 2). In our experiments, we find that saturate has no correlation with effective robustness on ImageNet-Vid-Robust. We believe that the discrepancy here is due to different measures of robustness.\\n\\nIn particular, the correlation plots in Figure 4 of [2] do not account for standard accuracy as a confounder. The authors consider video robustness as accuracy within k frames of an anchor frame given that the anchor frame was correctly classified. While this definition does somewhat account for models with higher standard accuracies, it is natural to expect that models with higher standard accuracy are still more likely to predict the neighboring frames correctly given that the anchor frame was predicted correctly. Thus, standard accuracy will be correlated with video robustness. Moreover, standard accuracy is also correlated with corruption accuracy, and hence corruption accuracy will be correlated with video robustness as well.\\n\\nHowever, this correlation does not mean that robustness to saturation *causes* robustness on videos. For instance, our testbed contains a model trained on saturation as data augmentation. While this model is highly robust to saturation (exhibiting only a 1% drop from standard accuracy to accuracy under saturations, compared to a baseline model exhibiting a 12% drop), it is no more video robust than a baseline without the saturation training (the saturation-trained model still experiences an 18% video robustness drop, compared to a baseline model exhibiting a 19% drop). This example further shows the need for our measure of effective robustness as it explicitly corrects for the confounding effect of standard accuracy.\\n\\n\\n[1] Hendrycks et al. Natural Adversarial Examples. https://arxiv.org/abs/1907.07174\\n[2] Gu et al. Using Videos to Evaluate Image Model Robustness. https://arxiv.org/abs/1904.10076\"}", "{\"title\": \"Papers connecting corruption robustness to broader robustness issues\", \"comment\": \"Benchmarking Neural Network Robustness to Common Corruptions and Perturbations\\nDan Hendrycks, Thomas Dietterich\", \"https\": \"//arxiv.org/abs/1808.08750, NeurIPS 2018\\n\\n\\u201cOne of the most remarkable properties of the human visual system is its ability to generalise\\nrobustly... Even if one never had a shower of confetti before, one is still able to effortlessly recognise objects at a carnival parade. Naturally, such generic, robust mechanisms are not only desirable for animate visual systems but also for solving virtually any visual task that goes beyond a well-confined setting where one knows the exact test distribution already at training time. Deep learning for autonomous driving may be one prominent example: one would like to achieve robust classification performance in the presence of confetti, despite not having had any confetti exposure during training time. Thus, from a machine learning perspective, general noise robustness can be used as a highly relevant example of lifelong machine learning requiring generalisation that does not rely on the standard assumption of independent, identically distributed (i.i.d.) samples at test time.\\u201d\"}", "{\"title\": \"Papers connecting adversarial robustness to broader robustness issues\", \"comment\": \"On Evaluating Adversarial Robustness\\nNicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, Aleksander Madry, Alexey Kurakin\", \"https\": \"//arxiv.org/abs/1706.06083, ICLR 2018\\n\\n\\u201cComputer vision presents a particularly striking challenge: very small changes to the input image can fool state-of-the-art neural networks with high confidence [28, 21]. This holds even when the benign example was classified correctly, and the change is imperceptible to a human. Apart from the security implications, this phenomenon also demonstrates that our current models are not learning the underlying concepts in a robust manner.\\u201d\"}", "{\"title\": \"The literature has made several connections between adversarial robustness and broader robustness claims.\", \"comment\": \"We thank the reviewer for their comments and address each point in turn:\\n\\n1) We agree with the reviewer that it is not a priori clear that adversarial robustness will be helpful outside a fully adversarial setting. Nevertheless, adversarial robustness has been connected to broader claims about robustness in various papers in the literature. Below, we compile a list of several making such claims and also cite papers connecting non-adversarial corruption robustness to robustness more broadly.\\n\\n\\n2) Regarding the importance of the effective robustness measure: Figures 1, 3, and 4 do rely on the effective robustness metric. In particular, the scatter plots show the effective robustness as the deviation of various models away from the linear fit indicated by the red line. The figures are interesting precisely because they clearly visualize the effective robustness. An important phenomenon in these three figures is that (almost) all models lie close to the linear fit. In particular, this means that the models have effective robustness close to zero. \\n\\nFigures 1, 3, and 4 can show the effective robustness via a scatter plot by using both axes as accuracies on two different test sets. While this is useful for visualizing effective robustness as deviation away from the linear fit, it also makes it impossible to plot effective robustness as a function of other quantities (there is no axis left to visualize other quantities). Hence we formally defined the effective robustness measure so we can visualize how effective robustness relates to other quantities such as robustness to various perturbations or to adversarial attacks.\\n\\nOur response to point 2 of Reviewer 3 offers another example of how effective robustness clarifies whether certain robustness interventions are helpful for making models truly more robust.\\n\\n\\n3) We will fix the caption of Figure 1.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper studied an interesting question, whether the gain of robustness from synthetic distribution shifts can be transferred/generalized to the robustness under natural distribution shifts. It was shown that in the context of natural distribution shifts, no current robustness intervention can really outperform standard models without a robustness intervention. The main strength of this paper is its extensive experimental study. However, I still have concerns on this work.\\n\\n1) The authors mentioned \\\"an implicit assumption underlying this research direction is that robustness to such synthetic distribution shifts will lead to models that also perform more reliably on natural distribution shifts.\\\" However, I am uncertain about this point. Let us take adversarial robustness as an example, I am NOT surprising that the robustness on crafted adversarial examples cannot be generalized to the robustness over natural distribution shifts. Thus, I do not think that adversarial robustness obeys the 'implicit assumption'. In other words, the studied problem should be better connected to adversarial robustness.\\n\\n2) The significance of 'effective robustness' is not clear. It seems that most of insights were learnt from Figure 1, 3 and 4. Do they rely on the metric 'effective robustness'? \\n\\n3) In Figure 1, 'Trained with more' -> 'Trained with more data'\\n\\n############# Post-feedback ################\\nThanks for the response. I am still not fully convinced by the significance of the findings in this work.\\nIn the paper, the authors highlighted that \\\"our results show that current robustness gains on synthetic distribution shifts do not transfer to improved robustness on the natural distribution shifts presently available as test sets.\\\" I am not surprising at this point, since the synthetic shift, e.g., introduced from adversarial examples, may only characterize the short-cut shift for misclassification. Thus, robustness learnt from this synthetic distribution shift might not transfer to the natural distribution shift. \\n\\nI decide to keep my original score.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"============================================ Update after rebuttal =========================================\\n\\nI thank the authors for their detailed rebuttal. I'm in general satisfied with the authors' response to my concerns, so as promised, I'm happy to increase my score.\\n\\n========================================================================================================\\n\\nThis paper considers the relationship between various measures of synthetic robustness and two distinct measures of natural robustness in large scale image classification models. The authors argue that the synthetic robustness measures considered in this paper are not predictive of natural robustness when the effect of baseline accuracy is subtracted. I think, if true, this is an important message that researchers in this area need to be aware of. However, I have a number of questions and concerns about the results. I would be happy to increase my score if the authors could address some of these issues:\\n\\n1) Another important recent benchmark not mentioned in the paper is ImageNet-A (https://github.com/hendrycks/natural-adv-examples). I would encourage the authors to include this benchmark among their natural robustness measures (in addition to ImageNetV2 and ImageNetVidRobust). The advantage of this dataset is that because it samples from the error distribution of a high-performing ImageNet model, to a large extent, it already comes with the baseline subtracted, so it essentially obviates the need for the indirect effective robustness measure introduced here. The raw accuracies on ImageNet-A would be directly interpretable and they would also answer the question \\u201cwhy should we care?\\u201d in a more visceral way, because even the Instagram trained state-of-the-art ImageNet models seem to achieve a mere 17% accuracy on this benchmark: https://arxiv.org/abs/1907.07640\\n\\n2) There seems to be a direct conflict between the main conclusion of this paper (that synthetic robustness measures do not predict natural robustness) and an opposite conclusion reached in an earlier paper (https://arxiv.org/abs/1904.10076) where the authors claim that robustness against synthetic perturbations like translation, hue, and saturation are actually highly predictive of video robustness (not sure if this would generalize to ImageNetV2). As far as I can see, these particular perturbation types are not included among the synthetic perturbations considered in this paper. Can you please clarify this discrepancy? \\n\\n3) Relatedly, looking at the scatter plots of effective robustness vs. robustness against individual perturbations in the appendix, especially for the video robustness measure, some of the correlations seem to be pretty significant (for example, video robustness vs. jpeg compression robustness, p. 19). So, I am wondering to what extent the main conclusion of this paper might just be driven by the averaging of a large number of non-predictive perturbations and a smaller number of more predictive perturbations. \\n\\n4) Also, ImageNet-P perturbations are not included in the paper. If the authors want to make their claims more reliable, I would encourage them to consider these among their synthetic perturbations as well. The translation perturbation in ImageNet-P, in particular, would be particularly important to consider for video robustness given the results from the arxiv pre-print mentioned in 2) above.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary: This paper tries to evaluate whether robustness \\u2018interventions\\u2019 such as robustness to adversarial examples and other synthetic distribution shifts on natural distribution shifts.\\nOverall this paper is generally well written and focuses on an important direction of understanding when the toy/synthetic settings used for convenience actually transfer to the real world. However, I have some concerns with the current version of this work.\", \"decision\": \"I vote for rejecting this paper. My main concerns are regarding both the motivation of this work and the quality and soundness of the claims.\\n\\n\\u2014 Soundness: My major concern is with the datasets that are used for \\u201cnatural\\u201d distribution shift. ImageNetV2 is essentially from the same distribution as the original Imagenet. It\\u2019s unclear what the distribution shift is and whether there is a distribution shift. In fact, the paper that produces Imagenet V2 tries very hard to minimize the distribution shift. Hence it seems unreasonable to use this dataset for a shift. Similarly, the other dataset used is from Imagenet Videos. Here again, it\\u2019s unclear what the shift actually is, and hence using these datasets to measure performance to natural distribution shifts seems questionable. \\n\\n\\u2014 Motivation: I am a little confused about the motivation of this work as presented. It does not seem unreasonable that robustness to one kind of shift doesn\\u2019t transfer to another. It seems like the right questions to ask would be the following: Do the synthetic robustness goals seem like they would occur in natural settings? Does robustness on synthetic datasets and synthetic generations of such perturbations transfer to perturbations (of the same kind) but occurring in natural world. \\n\\nFinally, it seems like the novelty/contribution of this work is very small. The paper uses existing datasets and existing robustness measures.\"}" ] }

[ "Jinlong Liu", "Yunzhi Bai", "Guoqing Jiang", "Ting Chen", "Huayan Wang" ]

[ "DNN", "generalization", "GSNR", "gradient descent" ]

{ "note_id": [ "DfU1kBE-ej", "rye6IOi42B", "HJgI26Q3iH", "S1xhXUI9jB", "SJxVhP5FoS", "B1ehzS5For", "B1lyj_jm5B", "HkxvaNRhFB" ], "note_type": [ "decision", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review" ], "note_created": [ 1576798731052, 1574381653041, 1573825966156, 1573705252200, 1573656491634, 1573655827967, 1572219030731, 1571771582980 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1732/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1732/Authors" ], [ "ICLR.cc/2020/Conference/Paper1732/Authors" ], [ "ICLR.cc/2020/Conference/Paper1732/Authors" ], [ "ICLR.cc/2020/Conference/Paper1732/Authors" ], [ "ICLR.cc/2020/Conference/Paper1732/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1732/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Spotlight)\", \"comment\": \"Quoting a reviewer for a very nice summary:\\n\\n\\\"In this work, the authors suggest a new point of view on generalization through the lens of the distribution of the per-sample gradients. The authors consider the variance and mean of the per-sample gradients for each parameter of the model and define for each parameter the Gradient Signal to Noise ratio (GSNR). The GSNR of a parameter is the ratio between the mean squared of the gradient per parameter per sample (computed over the samples) and the variance of the gradient per parameter per sample (also computed over the samples). The GSNR is promising as a measure of generalization and the authors provide a nice leading order derivation of the GSNR as a proxy for the measure of the generalization gap in the model.\\\"\\n\\nThe majority of the reviewers vote to accept this paper. We can view the 3 as a weak signal as that reviewer stated in his review that he struggled to rate the paper because it contained a lot of math.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"Paper Summary\\n\\nThis paper introduces a quantity termed the \\\"one-step generalization ratio\\\" . They derive approximate relations between OSGR and GSNR then show experimental results demonstrating the validity of these approximations, thus linking GSNR and a quantity related to generalization. They investigate the empirical value of GSNR during training of a neural network on Cifar10 with real labels vs. random labels. The final section derives a relation which attempts to explain the correlation between the size of the expected gradient and the learning of features.\\n\\nReview\\n\\nFirst, I should note that I am not well-versed in this specific line of work (GSNR ,as other papers were cited with this concept), and thus I cannot readily assess whether the contributions are \\\"well-placed in the literature\\\".\\n\\nFor the derivations and theory, I checked as much of the derivations as possible, although I'm not familiar with all the assumptions used throughout, but they seem relatively reasonable and supported by the experiments. I.e. 2.3.1 seems reasonable at the start of training. But I'm not sure how to evaluate some of the limiting arguments, like taking the step size to zero. \\n\\nThe paper addresses an important topic and is interesting. The arguments seem well reasoned and logical. The experiments addressed the veracity of of the approximate relations derived in the first section as well as some interesting trends related to the GSNR quantity during training of some simple models. \\n\\nSmall Concerns \\n\\nOn line 188 it is mentioned that GSNR is the \\\"reason\\\" for a particular phenomenon or explains a phenomenon. It certainly seems reasonable that it is correlated or an indicator of the underlying phenomenon, but it doesn't seem (at least intuitively) that GSNR is the cause or reason for the generalization concepts being examined.\"}", "{\"title\": \"We have submitted a final revision of the Discussion Stage 2.\", \"comment\": \"Hi, we have submitted the final revision. Changes in this revision are mainly about spelling errors.\\nThanks.\"}", "{\"title\": \"We have submitted a second revision.\", \"comment\": \"Hi, we have submitted a second revision where\\n(1) we changed \\\"eq. 17\\\" to \\\"eq. 19\\\" in the plot labels of Figure 1, Figure 5 and Figure 6;\\n(2) we removed the definitions of expectation of g_D and g_D' from the beginning of section 2.3, to improve the readability.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"Thank you for your responsible review and comments.\\n\\nWe have resubmitted a revision and keep improving the overall presentation, where\\n(1) the paper has been largely revamped to improve its quality of presentation; \\n(2) notations are changed to improve readability and consistency, and a notation table is added in Appendix C; \\n(3) all figures are re-generated to make them more readable. \\n(4) additional experiments on CIFAR10 and toy dataset are included.\\n\\nWe conducted the experiments on CIFAR10 and toy dataset, in Appendix A.2 and A.3 respectively. Briefly, for CIFAR10 we used a deeper CNN with batch normalization layer, and for the toy dataset which is constructed by a very simple way, we used a 2-layers MLP. The experiment settings on both dataset are similar to that on MNIST, and we got similar observations.\\n\\nAssumption 2.3.1 usually holds true in the early training stage. Multi-epoch training undoubtedly separates the training and testing per-sample gradient distributions significantly, which will break the stringent Assumption 2.3.1 in the late training stage. However, the experiments on Toy-DNN, MNIST and CIFAR10 together indicate that, for common network structures (MLP or CNN), with the commonly used step size of 0.001, the strongly positive correlation of LHS and RHS of equation (19)(in the revision) remains until training converges even when the Assumption 2.3.1 no longer holds. Therefore, the relation that \\u201cthe larger GSNR during training leads to better generalization ability\\u201d is valid during the whole training process with or without Assumption 2.3.1. We have also improved presentation of Assumption 2.3.1 to make it more intuitive.\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"Thank you for your responsible review and comments.\\n\\nWe have resubmitted a revision and keep improving the overall presentation, where\\n(1) the paper has been largely revamped to improve its quality of presentation; \\n(2) notations are changed to improve readability and consistency and a notation table is added in Appendix C;\\n(3) all figures are re-generated to make them more readable.\\n(4) additional experiments on CIFAR10 and toy dataset are included.\\n\\nYour main concerns are addressed below.\\n\\nThe definition of SNR on Wikipedia is the ratio between squared mean and variance, which ensures it is positive.\\n\\nAssumption 2.3.1 is needed for the derivation. It usually holds true in the early training stage. Multi-epoch training undoubtedly separates the training and testing per-sample gradient distributions significantly, which will break the stringent Assumption 2.3.1 in the late training stage. However, the experiments on Toy-DNN, Mnist and CIFAR10 together indicate that with the commonly used step size of 0.001, the strongly positive correlation of LHS and RHS of equation(19) (in the revision) remains until training converges even when the Assumption 2.3.1 no longer holds. Therefore, the relation that \\u201cthe larger GSNR during training leads to better generalization ability\\u201d is valid during the whole training process with or without Assumption 2.3.1. We have also improved presentation of Assumption 2.3.1 to make it more intuitive.\\n\\nWe have \\\\sigma represents \\u201cthe standard deviation of gradient mean over a training dataset of size n\\u201d whereas \\\\rho represents \\u201cthe standard deviation of parameters\\u2019 gradient of a single sample\\u201d. Under Assumption 2.3.1, it can be proved that they differ by a factor of n, which is the dataset size. \\\\mu is the same, please refer to equation 7 in the first version or equation 6 in the revision. We give a proof here.\\n$$E_{D\\\\sim \\\\mathcal Z^n}[\\\\mathbf g_D(\\\\mathbf\\\\theta)]=E_{(x_0,y_0)\\\\sim \\\\mathcal Z,(x_1,y_1)\\\\sim \\\\mathcal Z,...,(x_n,y_n)\\\\sim \\\\mathcal Z}(\\\\frac1n\\\\sum_{i=1}^n\\\\mathbf g_i(\\\\mathbf \\\\theta))=\\\\frac1n\\\\sum_{i=1}^nE_{(x_i,y_i)\\\\sim \\\\mathcal Z}(\\\\mathbf g_i(\\\\mathbf \\\\theta))=E_{(x,y)\\\\sim \\\\mathcal Z}(\\\\mathbf g(x,y,\\\\mathbf \\\\theta))\\\\equiv\\\\tilde{\\\\mathbf g}(\\\\mathbf\\\\theta)$$\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper defines the quantity of \\\"gradient SNR\\\" (GSNR), shows that larger GSNR leads to better generalization, and shows that SGD training of deep networks has large GSNR. It tells a great story on why SGD-trained DNNs have good generalization.\\n\\nThis topic is highly relevant to this conference.\\n\\nHowever, I struggle to rate this paper, since I feel swamped with math. It is hard work to read this paper, and I can honestly say that I could semi-confidently follow until about Eq. (8). To even get there, I had to scroll back and forth to remember the definitions of the various symbols. The math may be very well correct, but it is infeasible to verify (or follow) it fully. It does not make it easier that one cannot really search a PDF for greek symbols with indices etc. Someone who reads theoretical papers all day long might do better here.\\n\\nThis is the reason I rate the paper Weak Reject.\", \"some_feedback_points\": \"Section 2.1:\\n\\nEq. (1): It seems the common definition of SNR is the ratio of mean standard deviation. Your SNR is its square. This should be explained.\\n\\nI think it would help the reader a lot to give some intuitive meaning to the GSNR value. Can you, in Section 2.1, explain with examples what typical (or extreme) values would be?\\n\\nAssumption 2.3.1:\\n\\nThis is dropped on the reader without any motivation. It is also confusing: \\\"we will make our derivation under the non-overfitting limit approximation\\\" conflicts with \\\"In the early training stage,...\\\" So is this whole derivation only true in the early stages?\\n\\nAssumption 2.3.1 seems to address a thought I had when reading this: At the end of the training, I would expect mu_q(theta) to be zero (the definition of convergence). At the start, it is arbitrary as it entirely depends on the initial values. So this paper must look at some part between the two extremes to make sense. Is it? Is this assumption related?\\n\\nWhat is the difference between \\\\sigma and \\\\rho? Seems one is on the data distribution and one on a sampled set. But then why is \\\\mu the same in both cases (Eq. (1) vs. Eq. (5))?\", \"all_plots\": \"The plot labels are far too small to be readable.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"In this work, the authors suggest a new point of view on generalization through the lens of the distribution of the per-sample gradients. The authors consider the variance and mean of the per-sample gradients for each parameter of the model and define for each parameter the Gradient Signal to Noise ratio (GSNR). The GSNR of a parameter is the ratio between the mean squared of the gradient per parameter per sample (computed over the samples) and the variance of the gradient per parameter per sample (also computed over the samples). The GSNR is promising as a measure of generalization and the authors provide a nice leading order derivation of the GSNR as a proxy for the measure of the generalization gap in the model. After the derivation, experimental results on MNIST are presented and suggest that empirically there is a relation between the generalization gap of neural network trained by gradient descent and the GSNR quantity given in the paper. Next the author analyze the GNSR of DNNs as opposed with shallow models or other learning techniques and observe that the GSNR differs when using random labels (lower GSNR) as compared with true labels and exhibits different behavior along training for DNNs and gradient descent.\\n\\n Assumption 2.3.1 is not necessarily the most realistic in the current training paradigm since multi-epoch training indeed separates the training and testing per-sample gradient distributions significantly. \\n\\nOverall the paper gives a fresh (as far as I know) and nice idea on generalization of neural networks. The derivation requires quite a few stringent assumption (the leading order analysis, on the step size, assumption 2.3.1 on the test and train gradient distributions) the experiments do suggest that the theory is valid to an extent, especially during the early parts of training. In contrast, the presentation of the paper distracts from the work and needs additional cleaning up. Also, the experimental section 2.4 would benefit additional empirical analysis on other datasets and additional experiments, as well as more thorough explanation of the experiments. The writing of the paper needs additional proofreading as currently it is easy to spot typos and grammar errors along the paper. I currently vote weak reject, for solid content and not-quite-ready presentation.\\n\\nAdditional experiments and careful proofreading should definitely enhance the paper and get it to the level of publications, so I am willing to change my decision if the authors improve the writing and the overall presentation. I like the idea presented in the paper and encourage the authors to resubmit a more tidy draft.\\n\\nsmaller presentation issues and typos/grammar issues:\", \"figure_1\": \"name X and Y labels with more meaningful names instead of RHS,LHS\\nFigure 3(c): add legend\", \"line_64\": \"consists >>> which consists\", \"line_73\": \"we refer R(z) >>> we refer to R(z)\\nline 79 futher >>> further overall sentence needs more work\\nline 92 distribution becomes >>> distributions become\", \"line_100\": \"using common notation summing over n samples, the sum should possibly start from i=1 to n instead of i = 0 to n\\nline 132 include >>> includes (also possibly rephrase sentence)\\nline 136 vefity >>> verify\", \"line_137\": \"M number >>> M\\nline 157 the data points closely distributed >>> the data points are closely distributed\\nline 162 thorough analytically >>> thorough analytical \\n\\n\\n\\n\\n-------- Update After Rebuttal --------\\n\\nI have read the comments the authors made and reviewed the paper's revision. \\nThe presentation has improved a lot (Even though there is still room to go). \\nNonetheless at this time I choose to update my score to a weak accept, as I think the authors bring a fresh (as far as I know) and interesting idea with regards to empirical generalization.\"}" ] }

[ "Tanmay Gangwani", "Jian Peng" ]

[ "Imitation learning", "Reinforcement Learning", "Inverse Reinforcement Learning" ]

{ "note_id": [ "Qb4Lm7XiRW", "BJebQ1o2iS", "ryeRChQ2iB", "HkxzZ1dioB", "BygOeTvssB", "BkgygtDjoH", "r1gZxPviiS", "S1lu8BPijr", "SJg24Pj6tS", "B1lITN63FS", "SJeATxLnYr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798731023, 1573855001428, 1573825750054, 1573777146072, 1573776624066, 1573775590986, 1573775081112, 1573774671717, 1571825459562, 1571767486261, 1571737798479 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1731/Authors" ], [ "ICLR.cc/2020/Conference/Paper1731/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1731/Authors" ], [ "ICLR.cc/2020/Conference/Paper1731/Authors" ], [ "ICLR.cc/2020/Conference/Paper1731/Authors" ], [ "ICLR.cc/2020/Conference/Paper1731/Authors" ], [ "ICLR.cc/2020/Conference/Paper1731/Authors" ], [ "ICLR.cc/2020/Conference/Paper1731/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1731/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1731/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper addresses the setting of imitation learning from state observations only, where the system dynamics under which the demonstrations are performed differs from the target environment. The paper proposes to circumvent this dynamics shift with an algorithm whereby the target policy is trained to imitate its own past trajectories, re-ranked based on the similarity in state occupancies as judged by a WGAN critic.\\n\\nThe reviewers found the paper to be clearly written and enjoyable. The paper improved considerably through reviewers feedback. Notably, a behavior cloning from observations (BCO) baseline was added, which was stronger than the authors expected but still helped highlight the strength of the proposed method by comparison. R1 had a particularly productive multiple round exchange, clarifying the description of previous work, clarifying the details of the proposed procedure and strengthening the presentation of empirical evidence.\\n\\nThis work compellingly addresses an important problem, and in its final form is a polished piece of work. I recommend acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for further feedback.\", \"comment\": \"1- \\\"Depiction of MaxEnt-IRL\\\"\\n\\nOur corrections to the background on MaxEnt-IRL gave the wrong impression that MaxEnt makes design choice assumptions on the reward function. Our description also somewhat blurred the lines between the original MaxEnt-IRL formulation and recent works using maximum likelihood with energy-based policies. We agree with the reviewer that it\\u2019s very important to be crisp about established principles like MaxEnt-IRL, and clearly outline the transition to algorithms that are *inspired* by MaxEnt-IRL, such as AIRL. \\n\\nWe have strived to do this by modifying Section 2.1. Please see the changed text in blue. We describe the MaxEnt-IRL formulation in terms of the feature matching constraint and mention the solution to the optimization problem as per [1]. We then allude to recent methods [2-5] which are motivated by maximum entropy policies and employ an energy-based distribution, with energy parameterized by a neural network.\\n\\nWe are very grateful to the reviewer for their help on this subtle issue.\\n\\nWe would also like to present some clarification on reference to the \\u201csecond ambiguity\\u201d. We did not mean to say that the \\u201csecond ambiguity\\u201d is specific to MaxEnt-IRL. The IRL framework, in general, has the following 2 ambiguities, also brought up in recent works such as [4,5]:\\n\\n a. Many expert policies could explain a set of provided demonstrations\\n b. Many reward functions lead to the same optimal policy [Reward shaping ambiguity] \\n\\n[1] Brian D Ziebart. Modeling purposeful adaptive behavior with the principle of maximum causal entropy\\n[2] Finn et al. A connection between generative adversarial networks, inverse reinforcement learning, and energy-based models\\n[3] Haarnoja et al. Reinforcement learning with deep energy-based policies \\n[4] Fu et al. Learning robust rewards with adversarial inverse reinforcement learning\\n[5] Yu et al. Multi-Agent Adversarial Inverse Reinforcement Learning\\n\\n\\n2- \\u201cI still can not see the parametrization of the value-function inside the AIRL-discriminator\\u201d\\n\\nWe do not have separate reward and value (V) networks inside the AIRL discriminator. A single network--with 3 layers of 64 hidden units and tanh nonlinearity--is used to parameterize the advantage (or shaped rewards). If the reviewer is alluding to the original AIRL paper with separate reward and value networks to learn *disentangled rewards*, then we would like to note that the separate parameterization there serves a very specific purpose -- which is to learn rewards that are not shaped by environment dynamics, and can therefore be used for other related tasks/environments. For our paper, the goal is to learn a performant policy in the imitator MDP using demonstrations from a different expert MDP. Our rewards for the imitator could be shaped by the imitator MDP; we are not concerned about that aspect. Of course, it is straightforward to include a separate reward and value parameterization in our method as well, but we found the single network to work sufficiently well.\\n\\n\\n3- AIRL and SA-GAIL plots\\nThese are now provided in Appendix 7.8. We would combine all plots in the final revision.\"}", "{\"title\": \"Thank you for the revision\", \"comment\": \"Thank you for the revision. I think some parts have been clearly improved.\\nIt is now clear how exactly the buffer is updated, and the empirical results showing the improvement of the lower of the lower bound / Wasserstein-distance are assuring.\\n\\nHowever, some of my comments / question are not yet adequately addressed:\\n\\n1) Regarding the depiction of MaxEnt-IRL:\\nPlease note that MaxEnt-IRL is a specific algorithm proposed my Ziebart et al. (2008) with the following properties:\\n- Based on the principle of maximum entropy it aims to find the MaxEnt distribution that matches given, observed features in expectation.\\n- It learns a reward function that is linear in these features. This form of reward function is _not_ a design choice! The weights correspond to the Lagrangian multiplier w.r.t. the feature matching constraint. MaxEnt-IRL is _not_ ill-posed, the Lagrangian dual is convex in the weights. \\n\\nYes, there a several methods that, often inspired by the MaxEnt-IRL dual, aim to maximize the likelihood of a Boltzmann-policy. These methods may not be well-posed and may further be hard to justify by the principle of MaxEnt. I understand that you use MaxEnt-IRL as umbrella term for these method. However, you need to understand that using the name and reference of MaxEnt-IRL and claiming that it parametrizes the advantage function and that it is ill-posed by having some \\\"second ambiguity\\\" produces a completely wrong depiction of that algorithm.\\n\\n2) I did not ask why AIRL and SA-GAIL are shown. I asked why _only the mean_ is shown.\\n\\n3) I still can not see the parametrization of the value-function inside the AIRL-discriminator.\"}", "{\"title\": \"[Part 2 of 2] Response to AnonReviewer1. Thank you for your comments!\", \"comment\": \"continued...\\n\\n6- Regarding \\u201cDescription of MaxEnt-IRL using advantage does not seem correct\\u201d\\n\\nWe acknowledge that this might cause some confusion in the reader\\u2019s mind, and therefore, we have added a paragraph in Section 2.1 to explain our parameterization (Please see the text in blue in Section 2.1). Our approach for directly parameterizing advantage is motivated by the fact that we use AIRL as our MaxEnt-IRL method; AIRL recovers the advantage function of the expert. But we explain that other parameterizations are possible. Further justification for using the advantage is also provided.\\n\\nConcerning the normalizer when using parametric advantage, it would be 1 only when $a_w$ \\u201cconverges\\u201d to the actual advantage, which is the optimal point of the AIRL min-max training. For a general parameterization (without external constraints), we need to have the normalizer, which is a function of $w$.\\n \\n\\n7- Regarding \\u201cPresentation of experiments\\u201d \\n\\nLearning curves for GAIfO and unnecessary tables - We had added the GAIfO learning curves in Appendix 7.7. We appreciate this suggestion on improving the presentation of our results. We are going to remove the unnecessary tables from the final manuscript and replace them with plots from Appendix 7.7\\n\\nWhy are AIRL and SA-GAIL shown - They were added to show that while these methods work well when the dynamics match, the performance degrades under dynamics mismatch. A less important point was to convince the reader that we have implemented the baseline methods to the best of our ability, as shown by the good performance in the same dynamics scenario. We can move these to the Appendix.\\n\\nShow performance of experts in plots - This is now included in the plots in Appendix 7.7. We use PPO with the real environmental rewards for each of our MDPs, and plot the final performance of these \\u201coracle\\u201d experts. We stress that we have state-only demonstrations *only* from the expert in the unmodified MuJoCo MDP. The oracle experts for modified environments (changed gravity, density, friction) are only to get the final performance for the purpose of plotting.\\n\\n\\u201cNumber of evaluations\\u201d - This was mentioned in the second paragraph of Section 5. All our experiments average 8 independent runs with random seeds.\\n\\n\\u201cMeaning of the error region\\u201d - Thanks for pointing this out. This has been rectified. We plot the mean and standard deviation.\\n\\n\\u201cAIRL reward/value-function network architecture\\u201d - These were already mentioned in Appendix 7.2. We would be happy to provide more details if required.\"}", "{\"title\": \"[Part 1 of 2] Response to AnonReviewer1. Thank you for your comments!\", \"comment\": \"1- Regarding \\u201cLack of clarity on the update mechanism of the buffer\\u201d\\n\\nThank you for expressing this concern. We acknowledge that we may have condensed a bit too much when describing the buffer update in Section 3. We have now added Appendix 7.3 with further details on the update mechanism of the buffer. It mathematically defines the buffer distribution in terms of delta measures, and from there motivates using the trajectory \\u2018score\\u2019 (as defined in our main section) as a way to reduce the Wasserstein distance between expert-states and buffer-states. For completeness, we also outline the min-heap based priority queue algorithm. We would be happy to include any further details the reviewer may suggest.\\n\\n\\n2- Regarding \\u201cI2L is derived based on a lower bound but it seems that it does not even ensure improvement on that lower bound\\u201d\\n \\nWe have added Appendix 7.4 for this. The figures therein plot that gap between the original objective and lower bound, for all the environmental settings considered in the paper, and show that the gap generally reduces as the buffer distribution is updated over the iterations of the I2L algorithm. Please see our description and plots in Appendix 7.4, and also our response to AnonReviewer3.\\n\\n\\n3- Regarding \\u201cIt does not seem to ensure that the Wasserstein distance decreases compared to the last iteration\\u201d\\n\\nWe have added Appendix 7.5 with plots which show the estimated Wasserstein distance between expert and buffer (state) distributions, over the course of training. We observe that this value generally decreases over time, for all the environmental settings considered in the paper. Please see our description and plots in Appendix 7.5. \\n\\n\\n4- Regarding \\u201cPlease precisely state when the trajectories are removed from the trajectory buffer and the heap of the priority queue. May the same trajectories be used during several iterations?\\u201d\\n\\nWe hope that the details in Appendix 7.3 provide a clearer picture of when and how the buffer trajectories are updated. In summary, in each iteration of the I2L algorithm (Algorithm 1 in the paper), line 8 is where the update to the buffer happens. The exact rules for the update are in Appendix 7.3 (the algorithm box therein). We also motivate the update rule there.\\n\\n\\n5- Regarding \\\"Please elaborate: under which assumptions does the update of the trajectory buffer ensure that the W1-distance of the individual trajectories decreases with respect to the expert's trajectories?\\\" \\n\\nThe update mechanism for the buffer is designed such that empirical 1-Wasserstein distance between expert and buffer (state) distributions reduces. Specifically, using a priority-queue buffer, and adding trajectories using \\u2018score\\u2019 as the priority ensures the reduction. This is mathematically shown in Appendix 7.3 and empirically in Appendix 7.5\\n\\n\\ncontinued below...\"}", "{\"title\": \"Response to AnonReviewer3. Thank you for your comments!\", \"comment\": \"1- Regarding question on the error of the lower bound.\\n\\nWe thank the reviewer for this question, for it motivated us to do some empirical analysis on the convergence of the lower bound to the original likelihood objective (Proposition in the paper). We have added Appendix 7.4 in the revision with our observations. Therein, we detail how we estimate the values of the original objective and the lower bound for different buffer distributions. The figures in Appendix 7.4 plot that gap between the original objective and lower bound, for all the environmental settings considered in the paper, and show that the gap generally reduces as the buffer distribution is updated over the iterations of the I2L algorithm. A better lower bound in turn leads to improved gradients for updating the AIRL discriminator, ultimately resulting in more effective policy gradients for the imitator. We also believe that the lower bound we obtained in this proposition, although quite simple to prove, is non-trivial. For example, if the function approximation ($a_w$) is linear, the lower bound becomes tight and cannot be further improved.\"}", "{\"title\": \"Response to AnonReviewer2. Thank you for your comments!\", \"comment\": \"1- Regarding \\u201cBCO baseline\\u201d\\n\\nWe have now added Appendix 7.6 comparing I2L to BCO. Since the BCO code is not publicly available, we implemented the algorithm ourselves to the best of our ability using the description provided in the paper. We implement the BCO(alpha) version from the paper since it is shown to be better than vanilla BCO. Appendix 7.6 provides background details on BCO and the results. We observe the barring two situations (Ant with no dynamics mismatch, and Ant with 2x density), BCO is unsuccessful in learning high-return policies. This is potentially due to the difficulties in learning a robust inverse dynamics model, and the compounding errors problem inherent to BC. Similar performance for BCO is also reported by [1], in their Figure 3.\\n\\nThe fact that BCO works even on Ant is surprising enough. We thank the reviewer for pointing this out!\\n\\n[1] Generative Adversarial Imitation from Observation, Torabi et. al.\\n\\n\\n2- Regarding \\u201climitations of I2L\\u201d\\n\\nEnsuring sufficient diversity in the trajectories generated by the imitator agent (Line 5, Algorithm 1) might be important in certain situations. This is because these trajectories form the candidate set from which entries are added to the priority-queue buffer, based on the \\\"score\\\" as defined in Section 3. Consider a degenerate case where the agent gets stuck in a local optimum, thereby producing similar trajectories from a small part of the state-action space. These trajectories might not have an adequate score (priority) to enter the buffer, and the algorithm may get stuck. The issue can be potentially alleviated by incorporating techniques for efficient exploration for RL agents, such as those used in single-agent (e.g. intrinsic motivation) and population-based exploration methods. For the MuJoCo environments evaluated in our paper, we find that the exploration induced by a standard Gaussian parameterization for the stochastic policy suffices. But we would add this line of thought to our revised paper, discussing it as possible future work. \\n\\nFor training the Wasserstein critic, we do not expect I2L to add any new challenges, beyond those already encountered by prior approaches which use adversarial training.\"}", "{\"title\": \"General response to the reviewers\", \"comment\": \"We would like to thank the anonymous reviewers for their comments and constructive feedback. We address each reviewer's comments individually and summarize the major additions to the revision here:\\n\\n1. Added Appendix 7.3 with further details on the buffer B, and its update mechanism\\n2. Added Appendix 7.4 with some empirical evaluation of the lower bound used in the paper\\n3. Added Appendix 7.5 with plots to show a reduction in Wasserstein distance between expert-states and buffer-states\\n4. Added Appendix 7.6 on comparison to BCO baseline\\n5. Added Appendix 7.7 with missing learning curves\\n\\nAll changes are in blue text in the revision.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper proposes an imitation method, I2L, that learns from state-only demonstrations generated in an expert MDP that may have different transition dynamics than the agent MDP. I2L modifies the existing adversarial inverse RL algorithm: instead of training the disciminator to distinguish demonstrations vs. samples, I2L trains the discriminator to distinguish samples that are close (in terms of the Wasserstein metric) to the demonstrations vs. other samples. This approach maximizes a lower bound on the likelihood of the demonstrations. Experiments comparing I2L to a state-only GAIL baseline show that I2L performs significantly better under dynamics mismatch in several low-dimensional, continuous MuJoCo tasks.\\n\\nOverall, I enjoyed reading this paper. A few comments:\\n\\n1. It would be nice to include a behavioral cloning (e.g., BCO) baseline in the experiments. Your point in Section 4 that BC can suffer from compounding errors is well taken, but in my experience, BC can perform surprisingly well on some of the MuJoCo benchmark tasks, even from a single demonstration trajectory. Prior work showing relatively poor results for BC on MuJoCo tasks usually sub-samples demonstrations to intentionally exacerbate the state distribution shift encountered by BC.\\n\\n2. It would be nice to discuss potential failure cases for I2L. For example, how dependent is the method on the diversity of trajectories \\\\tau generated by the agent in line 5 of Algorithm 1? Are there conditions in which training the critic network to approximate the Wasserstein metric is harder than prior methods like Stadie et al. (2017) and Liu et al. (2018)?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The submission considers the imitation learning with environment change. In the new environment (where we we aim to conduct imitation learning), the expert demonstrations are unavailable, making the objective function (5) unavailable. To deal with this, the authors derive a lower bound to replace (5).\\n\\nIn the experiment section, the performance of the proposed method is clearly demonstrated. Environment change is a challenging case for existing imitation learning methods, while the proposed one works. \\n\\nWhile empirically, the performance of the proposed method is justified, I am curious how tight the bound used to replace (or approximate) (5) is. The submission offers few discussions on the error may induced by replacing (5) with the lower bound. Will such an error decrease or converge to 0, as we iterate according to the algorithm?\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary:\\nThe manuscript considers the problem of imitation learning when the system dynamics of the agent are different from the dynamics of the expert. The paper proposes Indirect Imitation Learning (I2L), which aims to perform imitation learning with respect to a trajectory buffer that contains some of the previous trajectories of the agent. The trajectory buffer has limited capacity and adds trajectories based on a priority-queue that prefers trajectories that have a similar state distribution to the expert. Similarity is hereby measured by the score of a WGAN-critic trained to approximate the W1-Wasserstein distance between the previous buffer and the expert distribution. By performing imitation learning with respect to a trajectory buffer, state-action trajectories of the agent's MDP are available, which enables I2L to apply AIRL (Fu et al. 2017). By using those trajectories for the transition buffer that have state-marginals close to the expert's trajectory, I2L produces similar behavior compared to the expert. I2L is compared to state-only GAIL, state-action-GAIL and AIRL on four MuJoCo tasks with modified dynamics compared to the expert policy. The experiments show that I2L may learn significantly better policies if the dynamics of agent and the expert do not match.\", \"decision\": \"I think that the submission is below borderline in its current state. The main reason for rejecting would be the insufficient justification of some algorithmic choices, improper presentation of the experiments and the description of MaxEnt-IRL, which seems quite wrong in my opinion. However, I think that the work is interesting and sufficiently novel and could be accepted if the mentioned issues were adequately addressed.\", \"supporting_arguments\": \"- Novelty / Significance: Imitation learning from state-only observations under dynamic mismatch is an important problem and a promising approach for training robots by non-experts. The idea of performing imitation learning with respect to carefully updated trajectory buffer seems simple and effective. Although self-imitation has been applied in reinforcement learning (references in submission), I am not aware of a similar approach in imitation learning.\\n\\n- Soundness / Correctness:\\n1. The description of MaxEnt-IRL seems quite wrong. The paper claims in Section 2.1. that MaxEnt-IRL maximizes the likelihood of the policy by optimizing the advantage function and thus only learns shaped rewards. However, the referenced paper (Ziebart et al. 2008) optimizes the weights of a linear reward function which in general does not correspond to the advantage of the learned policy. Also, the policy is not (only) proportional to the exponentiated advantage but equal to it and the normalizer of the trajectory distribution (Eq. 1) would be therefore 1.\\n\\n2. I think that the selection of trajectories for the buffer is not sufficiently well motivated. I2L is derived based on a lower bound but it seems that it does not even ensure improvement on that lower bound. It is not fully clear to me how the trajectory buffer is updated (see \\\"Clarity\\\") but it does not seem to ensure that the Wasserstein distance decreases compared to the last iteration. The trajectories are chosen greedily, i.e., without considering the remaining trajectories in the buffer which can be especially problematic for multimodal demonstrations. \\n\\n- Clarity / Presentation:\\nThe paper is well-written in general and only has few typos. It is not clear to me how exactly the trajectory buffer is updated. The submission merely states that the buffer is a priority-queue structure of fixed lengths, where the trajectory priorities are given by the average state score based on the Wasserstein critic. This description leaves several open questions (see \\\"Questions\\\") and further does not seem well motivated.\\n\\n\\n- Evaluation:\\nThe presentation of the experimental results seems odd. For the imitation learning experiments (no dynamic mismatch) the (apparently) same runs of I2L are compared to GAIL-S (Table 1) and GAIfO (Table 2) in separate tables. For the experiments under dynamic mismatch, the comparison with GAIL-S are shown with learning curves (Figure 3-5) and shaded error bars, whereas the results of GAIfO are only shown in in terms of final mean performance in Table 2. This presentation may make the impression or hiding learning curves or confidence intervals. Both tables could be removed by adding a single curve to each of the plots in Figure 3-5 and 7. It is also not clear to me, why only the mean of the final performance for AIRL and SA-GAIL is shown in Figure 3-5. Instead, the performance of the expert--which is currently not shown--would be better suited as a baseline level. The paper also does not seem to mention the number of evaluations and the meaning of the error region in the figures. Furthermore, some hyper-parameters, e.g., the architectures for the AIRL reward/value-function networks are not presented.\", \"questions\": [\"Please precisely state when the trajectories are removed from the trajectory buffer and the heap of the priority queue. May the same trajectories be used during several iterations?\", \"Please elaborate: under which assumptions does the update of the trajectory buffer ensure that the W1-distance of the individual trajectories decreases with respect to the expert's trajectories?\"], \"minor_comments\": \"\", \"typo\": \"\\\"upto a constant\\\"\", \"post_rebuttal_update\": \"I increased my rating to weak accept because the authors addressed my main concerns by a) giving additional information on the update of the buffer, b) improving the presentation of the experiments, c) fixing the description of MaxEnt-IRL and d) by showing empirically that both the lower bound increases / Wasserstein-distance decreases.\"}" ] }

[ "Yifan Hou", "Jian Zhang", "James Cheng", "Kaili Ma", "Richard T. B. Ma", "Hongzhi Chen", "Ming-Chang Yang" ]

[ "use", "graph information", "graph data", "graph neural networks", "measuring", "neural networks", "gnns", "representation", "limited", "much performance gnns" ]

{ "note_id": [ "FFSNqtyup1", "B1eu28SVsB", "B1lwIIHVjr", "r1x6NrBEsH", "BJxsAEBNor", "ryl-NEH4jS", "H1ep9tkO9H", "ryxJ4-ExqH", "SJlJQTncFH" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730992, 1573308080078, 1573307983264, 1573307701428, 1573307603436, 1573307433283, 1572497812992, 1571991846870, 1571634454545 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1730/Authors" ], [ "ICLR.cc/2020/Conference/Paper1730/Authors" ], [ "ICLR.cc/2020/Conference/Paper1730/Authors" ], [ "ICLR.cc/2020/Conference/Paper1730/Authors" ], [ "ICLR.cc/2020/Conference/Paper1730/Authors" ], [ "ICLR.cc/2020/Conference/Paper1730/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1730/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1730/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"Two reviewers are positive about this paper while the other reviewer is negative. The low-scoring reviewer did not respond to discussions. I also read the paper and found it interesting. Thus an accept is recommended.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Reporting Lambda f and Lambda l values in the main paper\", \"comment\": \"Thank you for your positive feedbacks and the suggestion. We have now included a separate table (Table 2) in the main paper to report and discuss Lambda f and Lambda l values of the datasets. Indeed, this should make the analysis in Section 4.2 easier to follow as the two smoothness metrics are one of our key contributions to measure and understand the use of graph information in GNNs.\"}", "{\"title\": \"[Part 3] Responses to the reviewer's three suggestions\", \"comment\": \"Response to Suggestion 1: Thank you for your suggestion. We believe one of the main problems is that we did not clearly highlight our two main contributions, and thus the first contribution, i.e., the smoothness metrics are proposed primarily to understand how the use of graph information may (or may not) benefit GNNs (as detailed in our response to the first concern above), was not clear to readers. To address your concern, we have revised our paper to clearly state the purposes of proposing the smoothness metrics, which should have strengthened the motivation for the smoothness metrics.\\n\\nIn addition, we have also added another comparison result, CS-GNN (no topology), so that the difference between CS-GNN (no topology) and GAT is mainly in the use of the smoothness metrics in CS-GNN\\u2019s attention mechanism. The results reported in Table 3 show that CS-GNN (no topology) still achieves significant improvements over GAT on the PubMed and BGP graphs. The improvements are mainly because the use of smoothness enables CS-GNN to increase the gain of positive information and reduce negative noisy information from these two graphs.\", \"response_to_suggestion_2\": \"Thank you for your suggestion. We did not include CS-GNN in Table 1 because Section 2.1 only introduces the background and we do not find it easy to include CS-GNN there without introducing some details of our method first, and thus this may make readers even more confused. Instead, we now list the key differences between CS-GNN and GAT clearly in Section 3.3 of the revised paper. First, the attention mechanism of CS-GNN is similar to the multiplicative attention mechanism, while GAT is similar to the additive attention mechanism. Compared with the traditional multiplicative attention mechanism, we introduced two weight matrices instead of one to compute the attention coefficients. Second, we consider the smoothness metrics in our attention mechanism, which is unique in CS-GNN and leads to significant performance improvements over existing GNNs (e.g., GAT) for processing graphs such as PubMed and BGP. Third, we also include the use of nodes\\u2019 local topology features in computing the attention coefficients to further improve the performance. (Remark: if the reviewer does not think the last point is significant enough, we may consider to remove it. Please also see our response to Suggestion 3 below.)\", \"response_to_suggestion_3\": \"We have conducted experiments on CS-GNN without using local topology features (LTF), denoted as CS-GNN (w/o LTF). The results are presented in Table 3 in the revised paper, which shows that using LTF does improve the performance of CS-GNN, but not significantly. However, the results do reveal the effectiveness of the smoothness metrics in CS-GNN, because the difference between CS-GNN (w/o LTF) and GAT is mainly in the use of the smoothness metrics in CS-GNN\\u2019s attention mechanism. As shown in Table 3, CS-GNN (w/o LTF) still achieves significant improvements over GAT on the PubMed and BGP graphs.\\n\\nWe also found that the use of LTF in other GNNs also only leads to marginal performance improvement, e.g., using LTF in GCN even has worse performance in a few cases. If the reviewer does not feel the improvements obtained using LTF very significant, we would consider not to put the use of side information as a subsection in Section 3.2, but only briefly discuss it at the end of Section 3.1. Thank you for suggesting the ablation experiments that helped us see the effects of LTF!\"}", "{\"title\": \"[Part 2] Response to the reviewer's concern on the marginal performance improvements\", \"comment\": \"Next, we would like to address your concern that our experimental results show marginal improvement over existing GNN methods. Actually we believe the improvements made by CS-GNN is quite significant, but maybe the numbers are buried in the table as we squeezed many things in Table 2, which has now been divided into 3 tables. CS-GNN outperforms GAT by 9.65% and 17.14% on the PubMed and BGP graphs. CS-GNN outperforms GraphSAGE by 8.98%, 9.15% and 6.32% on the Citeseer, Cora and BGP graphs. CS-GNN outperforms GCN by 6.23%, 12.78%, 11.45% and 26.26% on the Citeseer, Cora, PubMed and BGP graphs. Although in some other cases the performance may seem to be marginal, there is still improvement in every case and we would like to remark that such marginal improvements are actually considered \\u201cacceptable\\u201d in related work. For example, Velickovic et al. (2018) reported that GAT improves over GCN only for 2.26%, 1.97% and 0.00% on Citeseer, Cora and PubMed. But this is understandable as it is very challenging for a method to outperform other methods by a large margin in every case. Thus, while some of our improvements may be marginal, we would like to highlight that there are actually quite a number of other cases that CS-GNN does obtain significant improvements over existing methods. We understand these numbers could have been buried in the big table in the previous version, and we hope that by breaking the table into three smaller tables, the results are clearer now.\"}", "{\"title\": \"[Part 1] Response to the reviewer's concern on the similarity with GAT\", \"comment\": \"Thank you for your comments and the detailed suggestions. We try to address your main concerns below and hopefully this will make the contributions of our paper clearer.\\n\\nFirst, we would like to discuss your concern that our method is close to GAT. While our GNN model, i.e., CS-GNN, is also an attention-based method like GAT, we remark that our contributions are not just a new GNN model itself. In addition to CS-GNN, another main contribution of our work is the use of the two smoothness metrics to help researchers understand what and how much graph information can benefit an existing GNN model (not just our model). For example, we found that existing GNNs, e.g., GAT, achieve good performance for graphs with large feature smoothness and small label smoothness. However, they do not perform well for graphs with small feature smoothness or large label smoothness because they fail to obtain sufficient useful information or obtain too much negative information from neighboring nodes. \\n\\nSuch understanding is very important as until now there is limited knowledge about why and in what situations GNNs can outperform other methods. For example, we show in Table 3 that GAT outperforms MLP in some cases, but GAT has significantly worse results than MLP for processing graphs with small feature smoothness or large label smoothness. Also, label propagation can have similar or even better performance than GNN methods such as GCN and GraphSAGE for processing some graphs. \\n\\nIn our paper, we used the two smoothness measures to explain the performance of different methods on different graph datasets. And we believe the insights obtained in our study are valuable to researchers, as they can also use the smoothness measures to better understand the performance of different GNNs on different graphs (e.g., designing better models, building benchmarks). As currently there is a lack of metrics to measure the quantity and quality of information a GNN obtains from graph data, our smoothness metrics may also inspire other researchers to develop better measures to gain more comprehensive understanding. \\n\\nOur second contribution is the CS-GNN model, which improves the use of graph information using the smoothness values. Although CS-GNN is also an attention-based method, there are some key differences between CS-GNN and GAT as we will explain in our response to Suggestion 2 below. In addition, CS-GNN actually achieves quite significant performance improvements over GAT as we will explain in our next response.\\n\\nWe admit that we did not make our contributions clear in the paper, and this has obscured the primary purpose of the smoothness metrics. We have now stated in the 2nd paragraph of Section 1 the two main contributions of the paper. We hope the reviewer would find this clearer now. Thank you for raising this concern.\"}", "{\"title\": \"Revisions according to the reviewer's suggestions\", \"comment\": \"Thank you for your positive feedbacks and all your suggestions. We have made the following changes to our paper according to your suggestions:\\n\\n1. Rephrased all long sentences in the paper.\\n\\n2. Re-organized some subsections and tables. We separated Section 2.2 into two subsections and changed the title of Section 2.2. But we didn\\u2019t move paragraph 3 of Section 2.1 to Section 3.3 because paragraph 3 only discusses the background of existing GNNs listed in Table 1. Instead, we rewrote Section 3.3 to mainly discuss the key differences between existing GNNs and CS-GNN. \\n\\n3. Added a new section in Appendix G to discuss the differences between the KLD and mutual information (also see our answer to your question below).\\n\\n\\nAnswers to other questions are given as follows.\", \"q1\": \"Is there a particular reason for using the KLD instead of mutual information?\", \"a1\": \"We use the KLD instead of MI because of the following reason. In an information diagram, mutual information I(X; Y) can be seen as the overlap of two correlated variables X and Y, which is a symmetric measure. In contrast, KLD(X||Y) can be seen as the extra part brought by X to Y, which is a measure of the non-symmetric difference between two probability distributions. Considering the node classification task, the information contributed by neighbors and the information contributed to neighbors are different (i.e., non-symmetric). Thus, we use the KLD instead of MI.\\n\\nWe remark that, although some existing GNN works [1, 2] use MI in their models, their purposes are different from our work. MI can be written as I(X, Y) = D_{KL}(P(X,Y)||P(X) \\\\times P(Y)), where P(X,Y) is the joint distribution of X and Y, and P(X), P(Y) are marginal distributions of X and Y. From this perspective, we can explain the mutual information of X and Y as the information loss when the joint distribution is used to approximate the marginal distributions. However, this is not our purpose in node classification.\\n\\nWe included the above discussion in Appendix G of the revised paper as some readers may be interested to know.\", \"q2\": \"In 2nd sentence of section 2.2, what does \\u201cnode\\u2019s own information\\u201d mean? If it is the individual node\\u2019s features then why it naturally the representation vector h_v (which is aggregated with neighboring nodes?)\", \"a2\": \"The \\u201cnode\\u2019s own information\\u201d here refers to the node\\u2019s information before the aggregation with neighboring nodes takes place. That is, here h_v = h_v^{(0)} = x_v, i.e., the initial value of h_v or the feature vector x_v of v. To avoid confusion, we have modified the sentence as: We consider the context c_v of a node v as the node\\u2019s own information, which is initialized as the feature vector x_v of v.\\n\\nThank you again for your suggestions, which have helped make the paper flow clearer and highlight our contributions.\\n\\n[1] Petar Velickovic, William Fedus, William L. Hamilton, Pietro Li{\\\\`{o}}, Yoshua Bengio and R. Devon Hjelm. Deep Graph Infomax. In ICLR 2019.\\n\\n[2] Anonymous Authors. Utilizing Edge Features in Graph Neural Networks via Variational Information Maximization. Submitted to ICLR 2020.\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary\\n\\nThe paper proposes two graph smoothness metrics for measuring the usefulness of graph information. The feature smoothness indicates how much information can be gained by aggregating neighboring nodes while the label smoothness assesses the quality of this information. The authors show that Graph Neural Networks (GNNs) work best for tasks with high features smoothness and low label smoothness by utilizing information from surrounding nodes which also tends to have the same label. Based on these two metrics, the authors introduce a framework, called Context-Surrounding Graph Neural Network (CS-GNN), that utilizes important information from neighboring nodes of the same label while reduce the disturbance from neighboring nodes from different classes. The results demonstrate considerable improvement across 5 different tasks. \\n\\nStrength\\n\\nThe authors advocate for better understanding of the use of graph information in learning, which is both an important and interesting problem. Two graph smoothness metrics appears to be intuitive and reflect common situations in graph-based data (1) features from neighboring nodes contribute differently to target node representation (2) neighboring nodes information sometimes causing disturbance if node with different labels tend to be connected. The paper provides some theoretical analysis that supports this claim and thorough experiments that show the correlation between the two proposed metrics with the performance of GNNs. \\n\\nWeakness\\n\\nWhile the paper is reasonably readable, there is certainly room for improvements in the clarity of the paper. First, I would suggest the authors to avoid too much word repetition as well as long, obscure sentences. For example, the first sentence of section 2.2 can be rewritten as \\u201cGNNs usually contains an aggregation step to collect neighboring information and a combination step that merges this information with node features.\\u201d The flow of the paper is also hard to follow and need some rearrangement. For instance, paragraph 3 of section 2.1 can be pushed until section 3.3. Another suggestion about the flow is to separate section 2.2 into two subsections for features smoothness and label smoothness (also the title of this section need to be refined). Finally, the results would be more clear if separated into different tables or subsections/paragraphs.\\n\\nQuestions\\n\\n* Is there a particular reason for using the KLD instead of mutual information?\\n* In 2nd sentence of section 2.2, what does \\u201cnode\\u2019s own information\\u201d mean? If it is the individual node\\u2019s features then why it naturally the representation vector h_v (which is aggregated with neighboring nodes?)\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposes two smoothness metrics to measure the quantity and quality of the graph information that GNNs employ in the model. With the smoothness measures, this paper proposes a new GNN model to improve the use of graph information. Overall, the paper is well-organized and clearly written. The main concern is the novelty, since the proposed method is pretty close to the graph attention network (GAT), except using the two smoothness metrics and a slightly different way to compute the attention coefficients. Experimental results show marginal improvement over existing GNN methods on the node classification task. Given these aspects, it is not that convincing that the introduced smoothness metrics are necessary. I would like to recommend a weak reject for this paper.\", \"suggestions_to_improve_the_paper\": \"1) It would be better to provide more convincing evidence and motivation for the smoothness metrics, either in theory or empirical analysis.\\n\\n2) When describing the proposed method, organize the key differences in a more clear way. For example, add the proposed method directly in Table 1 and summarize the key differences in bullet points. Currently, this important part is deferred to Section 3.3, which may cause confusion for the readers to understand the paper.\\n\\n3) In Section 3.2, this paper claims that the proposed method can easily include side information on graphs to improve performance, by using the topology feature in the last fully connected layer for class prediction. However, this technique can also be used in existing GNN models, and it is not clear why this is described as something unique for the proposed method. Also, there is no corresponding ablation experiments to compare the performance of the proposed method with and without using local topology features.\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"The authors study how neighbor information on graphs can be used in Graph Neural Networks. It proposes measures on whether the data in neighboring nodes are useful in terms of labels or features. It also provides a new Graph Neural Network algorithm that is a modification of attention-based models incorporating the derived label and feature smoothness measures. The paper demonstrates the usefulness of these measures and algorithms with several different baselines from different families. The writing is mostly smooth, and the authors seem to provide enough detail of the experiments performed.\\n\\nThe proposed measures look simple but effective (as demonstrated in Table 2). The paper compares different techniques and shows that when neighboring labels are not smooth, techniques such as label propagation does not help. I do recommend Table 4 Lambda f and Lambda l values to be included in the main paper (though text mentions). When incorporated into the attention-based GNNs, Figure 1 also shows that smoothness parameters of the techniques also improve.\"}" ] }

[ "Yu-Xiong Wang", "Yuki Uchiyama", "Martial Hebert", "Karteek Alahari" ]

[ "few-shot learning", "meta-learning" ]

{ "note_id": [ "hq08Bo79qc", "BylL9a5hsS", "Hklm7p93jS", "rJggDiqhoB", "HkeTeIBiiB", "ByxBNMRKjr", "SJlzWIKX5r", "Hklpg8yk9r", "rJl3L37nYB", "HkeCv6fj_B" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "comment", "official_review", "official_review" ], "note_created": [ 1576798730956, 1573854605845, 1573854491225, 1573854039847, 1573766644961, 1573671468529, 1572210170368, 1571907061279, 1571728468427, 1570610533931 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1729/Authors" ], [ "ICLR.cc/2020/Conference/Paper1729/Authors" ], [ "ICLR.cc/2020/Conference/Paper1729/Authors" ], [ "ICLR.cc/2020/Conference/Paper1729/Authors" ], [ "ICLR.cc/2020/Conference/Paper1729/Authors" ], [ "ICLR.cc/2020/Conference/Paper1729/AnonReviewer1" ], [ "~Temporary_Temp1" ], [ "ICLR.cc/2020/Conference/Paper1729/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1729/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper describes a new approach to meta-learning with generating new useful examples.\\n\\nThe reviewers liked the paper but overall felt that the paper is not ready for publication as it stands.\\n\\nRejection is recommended.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thank you for the comments and interest in our approach.\", \"comment\": \"1. \\u2026 imposing that Performance(Query | Augmented data) = Performance(Query | Support):\\n\\nWe are not imposing that Performance(Query | Augmented data) = Performance(Query | Support). Instead, we target Performance(Query | Augmented data) = Performance(Query | Larger set of real data). This is explained in the paragraph \\u201cSoft precision-inducing hallucinator\\u201d in Section 4, page 5. Specifically, given an initial large training set S^*_{train}, which contains n^\\u2217 examples for each of the m classes, we randomly sample n (n << n^\\u2217 ) examples per class, and obtain a subset S_{train}. From S_{train}, the hallucinator G generates n^\\u2217 examples per class as S^G_{train}. This produces two training sets: S^*_{train} with real examples and S^G_{train} with hallucinated examples, where both contain the same number of examples. Importantly, note that S^G_{train} is hallucinated from the subset S_{train} instead of the initial large set S^*_{train}, and because n << n^\\u2217 , we rule out the trivial hallucinator.\\n\\n2. Would you have some visualizations of the hallucinated images:\\n\\nOur hallucination is performed in the pre-trained feature space, and visualizing them directly is not intuitive. Instead, we provided a t-SNE plot of hallucinated samples in Figure 3. In the appendix A.8, we have included an additional visualization of hallucinated examples in the pixel space, using the nearest neighbor real image in the feature space, corresponding to each hallucinated feature sample.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"We thank the reviewer for the comments. The comments focus mostly on the generality of our approach and additional comparison. We address all these points as follows.\\n\\n1. \\u2026 applying their proposed technique to meta-learning based techniques that do not involve learning a metric-embedding space and instead learn the learning procedure via nested optimization, e.g. MAML and its variant:\\n\\nIn principle, our meta-learning with hallucination framework requires that the learner (i.e., the classification model h) is differentiable with respect to the examples in the augmented training set, thus allowing us to back-propagate the final loss and update not just the parameters of the classification model, but also the parameters of the hallucinator. This is true for many meta-learning algorithms, including the optimization-based meta-learning approaches such as MAML.\\n\\nTo show the generality of our framework, we apply it to MAML. Implementation details of this integration are provided below. We have included the result of MAML, \\u2018MAML w/ G' which corresponds to hallucination-based MAML (i.e., hallucination framework following Wang et al., 2018), and \\u2018MAML w/ G + PECAN\\u2019 (i.e., our proposed hallucination framework) in Table 4, and updated the submission accordingly. As shown in the summary below, our approach is generic and applies to MAML as well.\\n\\nMethod n=1 n=5\\nMAML (ResNet-10) (Finn et al., 2017) 54.69\\u00b10.89 66.62\\u00b10.83\\nMAML w/ G (Wang et al., 2018) 56.37\\u00b10.63 68.91\\u00b10.57\\nMAML w/ G + PECAN (Ours) 58.39\\u00b10.37 71.36\\u00b10.44\", \"implementation_details\": \"Following the procedure described in our submission, we use a ResNet-10 architecture as the feature extractor. We perform meta-learning over these pre-trained features. The embedding architecture of the MAML classification network h is composed of two-layer MLPs. The architecture of the generator G remains the same as in our submission.\\n\\nIn each meta-training episode, we sample a batch of few-shot classification tasks. For each of the tasks, following the process in Figure 2, we sample training sets S_train^\\u2217 and S_train, sample a test test S_test, hallucinate S_train^G, and obtain an augmented training set S_train^aug. In the MAML inner loop, for each task, conditioning on its S_train^aug (S_train^G, or S_train^*), we adapt the parameters of h^cls (h^G, or h^real) using few gradient updates. For each task, the adapted h^cls (h^G, or h^real) is evaluated on the corresponding S_test. In the MAML outer loop, we average the classification objective on S_test across the batch of tasks. In a similar way, we compute the collaborative objective in the pre-trained feature space. The final loss is used to update the initial MAML model and the hallucinator.\\n\\n2. Table 4 \\u2026 the results of PMN w/G:\\n\\nWe have included the result of PMN w/ G (Wang et al, 2018) in Table 4, and updated the submission accordingly. Our approach significantly outperforms Wang et al., 2018. Here, we summarize the comparison:\\n\\nMethod n=1 n=5\\nPMN w/ G (Wang et al, 2018) 62.28\\u00b10.53 78.28\\u00b10.62\\nPMN w/ G + PECAN (Ours) 63.93\\u00b10.40 80.58\\u00b10.29\\n\\n3. \\u2026 hallucinating examples for regression tasks:\\n\\nIn our paper, we mainly focus on few-shot classification tasks. However, our approach is general and applies to few-shot regression tasks as well. When extending our approach to address regression tasks, we need to slightly modify the design of the hallucinator G. Specifically, in classification tasks, the hallucinator G takes as input a seed example (x, y) and outputs a hallucinated example (x\\u2019, y\\u2019)=(G (x, z), y), where the class label y\\u2019=y, indicating that the hallucinated example belongs to the same class as the seed example. In contrast, in regression tasks, y becomes a continuous quantity, thus we cannot enforce the constraint y\\u2019=y. To address this issue, we modify the hallucinator G to directly generate the tuple (x\\u2019, y\\u2019) = G (x, y, z). This is achieved by the concatenation of x and y as input to G.\\n\\nWe incorporated our hallucinator with MAML and evaluate on the sinusoidal regression task proposed in (Finn et al., 2017). Our \\u2018MAML w/ G + PECAN\\u2019 outperforms MAML and \\u2018MAML w/ G\\u2019 for the regression task as well. Here we briefly summarize the mean squared error comparison:\\n\\nMethod n=5\\nMAML (Finn et al., 2017) 0.84\\nMAML w/ G (Wang et al, 2018) 0.67\\nMAML w/ G + PECAN (Ours) 0.52\\n\\nWe have revised the submission accordingly to emphasize classification tasks, especially in the summary of contributions in Section 1, as well as showing our generality to regression tasks in the appendix A.7.\"}", "{\"title\": \"General Response\", \"comment\": \"We thank the reviewers for their positive comments --- our \\u201cnovel and interesting idea\\u201d (R3) to few-shot learning achieves \\u201cconsistent improvement in performance versus various state-of-the-art meta-learning algorithms\\u2019\\u2019 (R1) with \\u201cmany experiments\\u201d (R1) and \\u201cquite nice and thorough ablation studies\\u201d (R1). We address all the comments below and have also updated the submission accordingly.\\n\\nIn the revised submission, we: (1) emphasized few-shot classification and included few-shot regression results in the appendix A.7, (2) included additional comparisons suggested by the reviewers, and (3) included visualization of the hallucinated examples suggested by the public comment in the appendix A.8.\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"We thank the reviewer for the comments. The comments focus mostly on the clarity of some details and additional comparison. We address all these points as follows.\\n\\n(1) In Figure 2 \\u2026 why not generate the S_{train} directly and then measure your precision-inducing loss over the real set S_{train} and S^G_{train}:\\n\\nThis is because S^G_{train} is hallucinated from S_{train}. If we measure the precision-inducing loss over S^G_{train} and S_{train}, we will then end up with a trivial hallucinator which only needs to memorize S_{train} and is not useful for improving classification performance. In contrast, in our approach, we measure the precision-inducing loss over S^G_{train} and a much larger real set S^*_{train}. By doing so, the hallucinator is forced to generate additional useful examples from the small training set S_{train}, so that the classifier trained on them matches the performance of a classifier trained on a larger set of real examples S^*_{train}. Please refer to Section 1, paragraph 4, page 2, and Section 4, paragraph \\u201cSoft precision-inducing hallucinator\\u201d, page 5, for more details.\\n\\nIn fact, introducing a large set of real examples S^*_{train} as learning guidance is one of our key contributions: we not only extract shared knowledge across a collection of few-shot learning tasks, but also leverage additional knowledge in large-sample models to guide hallucination and few-shot learning.\\n\\n(2) For Function 2 \\u2026 could we compute the distance on the probability vectors that contains the logit for ground-truth label:\\n\\nOur final classification objective consists of the soft precision-inducing loss (i.e., Function 2), which is measured in the absence of ground-truth labels, and the hard precision loss, which is the standard cross-entropy classification loss measured with respect to ground-truth labels only (the last two sentences below Function 2, in the paragraph \\u201cSoft precision-inducing hallucinator\\u201d, Section 4, page 5).\\n\\nBy doing so, we: (i) separate the impacts of the predictions associated with ground-truth labels and other entries in prediction probability vectors, and (ii) ensure that the soft precision-inducing loss is not dominated by the predictions associated with ground-truth labels, given that they have already contributed to the hard precision loss.\\n\\nEmpirically, in the ablation study of \\u201cChoice of similarity measure in soft precision-inducing loss\\u201d and Table 3 (page 8), we showed that our loss function without ground-truth labels consistently and significantly outperformed the loss with ground-truth labels, for the cosine as well as other measures.\\n\\n(3) Some latest work \\u2026 \\u201cFew-shot Learning via Saliency-guided Hallucination of Samples\\u201d (Ref1) and \\u201cEdge-Labeling Graph Neural Network for Few-shot Learning\\u201d (Ref2) \\u2026 compare with these two methods:\\n\\nWe have included the comparison with these two methods on miniImageNet in Table 4, and updated the submission accordingly. Our approach significantly outperforms these methods. Here we summarize the results:\\n\\nMethod n=1 n=5\\nSalNet (Ref1) 57.45\\u00b10.88 72.01\\u00b10.67\\nEGNN+Transduction (Ref2) NA 76.37\\nOurs 63.93\\u00b10.40 80.58\\u00b10.29\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"We thank the reviewer for the comments. The comments focus mostly on the novelty and the clarity of some details. We address all these points as follows.\\n\\n1. The novelty ... hallucinating in Wang et al., 2018 ... knowledge distillation (Hinton et al., 2015):\\n\\nWhile we agree with the reviewer that our approach falls into the category of meta-learning with data hallucination (Wang et al., 2018), it is not a trivial extension to any of the existing work. Different from Wang et al., 2018, which learns the data hallucinator solely through end-to-end training with the classifier and without any additional constraints, we investigated critical and unexplored properties (i.e., precision and collaboration) that the data hallucinator should satisfy. We instantiated these properties as novel loss functions to improve the precision of the hallucinated examples, and enhance the collaboration between the hallucinator and the classifier. In addition, from a broader perspective, our approach not only extracts shared knowledge across a collection of few-shot learning tasks as most of the existing meta-learning approaches normally do, but also leverages additional knowledge in large-sample models to guide hallucination and few-shot learning. As we have shown in the experiments, our approach consistently and significantly outperforms Wang et al., 2018 for different meta-learning approaches and across different datasets.\\n\\nWhile our soft precision-inducing loss is related to knowledge distillation (Hinton et al., 2015), as mentioned in Section 4, paragraph \\u201cSoft precision-inducing hallucinator\\u201d, page 5, it is different in three important ways. First, knowledge distillation typically focuses on model compression, which trains a shallow \\u201cstudent\\u201d neural network by mimicking the output of a deep \\u201cteacher\\u201d model. These two models are trained on the same data but are of different capacities. In contrast, our soft precision-inducing loss compares two models of the same capacity but trained on different types of data: a classifier trained on real data, and another trained on hallucinated examples. Second, our formulation measures the similarity of two probabilities in the absence of ground-truth labels, which is different from the formulation in Hinton et al., 2015. Third, in the ablation study of \\u201cChoice of similarity measure in soft precision-inducing loss\\u201d and Table 3 (page 8), we empirically showed that our loss function significantly outperformed the loss used in Hinton et al., 2015, as well as other similarity measures.\\n\\n2. how to use the collaborative objective on the hallucinator \\u2026 the soft precision-inducing loss (l_hal^pre) for hallucinator:\\n\\nThe soft precision-inducing losses are computed in a similar way for l_hal^pre and l_learner^pre. The difference lies in that l_learner^pre is a loss in the classifier embedding space, while l_hal^pre is in the pre-trained feature space. This is also noted in Review #1: the collaborative objective applies \\u201cdirect early supervision in the feature space in which the hallucination is conducted in addition to in the classifier embedding space.\\u201d\\n\\nMore precisely, we first pre-train a deep convolutional network using a standard cross-entropy loss, and use it to extract image features. Meta-learning is then performed over these pre-trained features (the last paragraph in Section 3, page 4). Without loss of generality, we take prototypical networks (PN) as an example (the last paragraph in Section 4, page 6). PN learns a new embedding space \\u03a6 on top of the pre-trained feature space X, and uses a non-parametric nearest centroid classifier to assign class probabilities for a test example based on its distances from class means in \\u03a6. The embedding architecture of PN is composed of two-layer MLPs (implementation details in Section A.1).\\n\\nIn each meta-training episode, after sampling S_train^\\u2217, S_train, and S_test and hallucinating S_train^G in the pre-trained feature space X, we perform nearest centroid classification, and produce the collaborative objective L_hal on S_test, including the soft precision-inducing loss l_hal^pre. We then feed the examples to the PN learner, obtain their embedded features in \\u03a6, perform nearest centroid classification, and produce the classification objective L_learner on S_test, including the soft precision-inducing loss l_learner^pre (the last paragraph in Section 4, page 6).\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper describes a method that builds upon the work of Wang et al. It meta-learns to hallucinate additional samples for few-shot learning for classification tasks. Their two main insights of this paper are to propose a soft-precision term which compares the classifiers' predictions for all classes other than the ground truth class for both a few-shot training set and the hallucinated set and b) to introduce the idea of applying direct early supervision in the feature space in which the hallucination is conducted in addition to in the classifier embedding space. This allows for stronger supervision and prevents the hallucinated samples from not being representative of the classes. The authors show small, but consistent improvement in performance on two benchmarks: ImageNet and miniImageNet with two different network architectures versus various state-of-the-art meta-learning algorithms with and without hallucination. The authors have adequately cited and reviewed the existing literature. They have also conducted many experiments (both in the main paper and in the supplementary material) to show the superior performance of their approach versus the existing ones. Furthermore their ablation studies both for the type of soft precision loss and for their various individual losses are quite nice and thorough.\\n\\nOverall the contribution of this paper is incremental over Wang et al and is mainly in the introduction of their new loss terms to regularize the hallucination process. This is clearly evident from Table 2 (comparing rows 1 and 3), where much of the performance gain is attained by including the l_learner^cls term versus the collaborative loss term (comparing row 1 and row 4).\\n\\nFurthermore, I would like the author to answer the following two questions:\\n\\n1. The authors claim that their method is general applicable to all meta-learning methods and can be combined with them. Yet, the meta learning methods that they apply it to: prototypical networks, prototypical matching networks and cosine classifiers are all metric-learning-based meta-learning techniques. I would like the authors to outline the procedure (and preferably also show experiments) for applying their proposed technique to meta-learning based techniques that do not involve learning a metric-embedding space and instead learn the learning procedure via nested optimization, e.g. MAML and its variants.\\n\\n2. In Table 4, I would like to see the results of PNM w/G or in other words the results of (Wang et al, 2018)'s method in comparison to the authors' proposed method.\\n\\n3. The authors make no attempt to solve the problem of hallucinating examples for regression tasks. This is fine as it is perhaps outside he scope of their current work. However, I would like the authors to fully qualify their claims everywhere in the paper and restrict the contribution of their work to classification tasks only.\"}", "{\"title\": \"Clarification on constraint\", \"comment\": \"Thanks for the work, and congratulations on achieving such good results!\\n\\nI have concerns regarding the objective function. \\n\\nAs far as I understand, you are imposing that Performance(Query | Augmented data) = Performance(Query | Support). If this is the case, then a trivial solution would be when the augmented data = support data (i.e. the hallucinator is essentially an autoencoder). \\n\\nAm I missing something? Would you have some visualisations of the hallucinated images?\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a general meta-learning with hallucination framework called PECAN. It is model-agnostic and can be combined with any meta-learning models to consistent boost their few-shot learning performance.\\n\\nThere are two key points for the proposed model. On the one hand, the authors introduce a novel precision-inducing loss which encourages the hallucinator to generate examples so that a classifier trained on them makes predictions similar to the one trained on a large amount of real examples. On the other hand, the authors introduce a collaborative objective for the hallucinator as early supervision, which directly facilitates the generation process and improves the cooperation between the hallucinator and the learner.\\n\\nOn the whole, the paper is well-written, and the proposed idea is novel and interesting.\", \"i_have_some_following_major_concerns_about_the_paper\": \"(1) In Figure 2, the authors first sample the training set S^*_{train}, which contains n^* examples for each of the m classes, and then they randomly sample n examples per class, and obtain a subset S_{train}. Why not generate the S_{train} directly and then measure your precision-inducing loss over the real set S_{train} and S^G_{train}? I hope the authors explain it in their paper.\\n(2) For Function 2 in the paper, why compute the cosine distance on the probability vectors that are obtained by removing the logit for ground-truth label in original probability distributions? Could we compute the distance on the probability vectors that contains the logit for ground-truth label? I hope the authors explain it in detail.\\n(3) As far as I know, there are some latest work on few-shot learning in 2019, especially the work \\u201cFew-shot Learning via Saliency-guided Hallucination of Samples\\u201d and \\u201cEdge-Labeling Graph Neural Network for Few-shot Learning\\u201d. I hope the authors can compare with these two methods to further demonstrate the effectiveness of the proposed model.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"In this paper, the authors address few-shot learning via a precise collaborative hallucinator. In particular, they follow the framework of (Wang et al., 2018), and introduce two kinds of training regularization. The soft precision-inducing loss follows the spirit of adversarial learning, by using knowledge distillation. Additionally, a collaborative objective is introduced as middle supervision to enhance the learning capacity of hallucinator.\", \"here_are_some_comments\": \"1. The novelty is relatively limited. The idea of hallucinating has been mainly introduced in (Wang et al., 2018). The soft precision-inducing loss is a straightforward extension of knowledge distillation (Hinton et al., 2015). \\n\\n2. It is not quite clear about how to use the collaborative objective on the hallucinator. Fig.2 and the text in ' Collaboration between hallucinator and learner ' (page 5) are not quite informative. Especially, how to perform the soft precision-inducing loss (l_hal^pre) for hallucinator?\"}" ] }

[ "Ying Da Wang", "Pawel Swietojanski", "Ryan T Armstrong", "Peyman Mostaghimi" ]

[ "Super Resolution Generative Adversarial Networks", "Perceptual Loss Functions" ]

{ "note_id": [ "2aPZHzwv7c", "HyguxlXIoS", "Hyl0aS9aYr", "SJlcwtH2tB", "HygS97ittS" ], "note_type": [ "decision", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730927, 1573429232038, 1571820997746, 1571735906338, 1571562380688 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1728/Authors" ], [ "ICLR.cc/2020/Conference/Paper1728/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1728/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1728/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes to use the grey level co-occurrence matrix method (GLCM) for both the performance evaluation metric and an auxiliary loss function for single image super resolution. Experiments are conducted on X-ray images of rock samples. Three reviewers provide comments. Two reviewers rated reject while one rated weak reject. The major concerns include the lack of clear and detailed description, low novelty, limited experiment on only one database, unconvincing improvement over the prior work, etc. The authors agree that the limited experiment on one database does not demonstrate the generalization capability of the proposed method. The AC agrees with the reviewers\\u2019 comments, and recommend rejection.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Novelty and Experiments\", \"comment\": \"Regarding the comments:\\n\\n> We agree that the method we propose should be further tested on other image types, such as medical images and natural images. These experiments are in progress\\n\\n> The novelty of this work is the summation of its parts and the performance it can obtain. As we state in the previous point, we agree that results would be more convincing and general with reported results on natural images and other X-ray images.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper adopts a loss metric called Grey Level Co-occurence Matrix (GLCM) as a new measurement of perceptual quality for single image super-resolution. The GLCM is particularly well suited for automatic perceptual/textural in-domain comparisons, which does not require time-consuming expert MOS evaluations. Experimental validation is carried on X-ray images of rock samples and promising results are achieved.\", \"my_main_concerns_are_as_follows\": [\"Novelty is quite limited. First, as the main contribution of this work, GLCM is proposed from Haralick et al.(1973), which is not novel. Second, the network structure used in this work is also based on SRGAN/EDSRGAN, which is also not novel.\", \"Experiments are not convincing. In super-resolution task, the differences in Tab. 2 are quite minor, which can also be regarded as the same or may be caused by randomness. Moreover, the authors should conduct experiments on generic image SR to demonstrate the effectiveness of GLCM. Further, more qualitative visualizations are also needed to demonstrate the effectiveness of GLCM.\", \"Page 5 only contains 1 figure, leaving a lot of space that is not fully used. Besides, Fig. 3 cannot reflect the advantage of GLCM, since SRGAN is much larger and also better than SRCNN.\"]}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper considers the problem of generating a high-resolution image from a low-resolution one. The paper introduces the Grey Level Co-occurrence Matrix Method (GLCM) for evaluating the performance of super-resolution techniques and as an auxiliary loss function for training neural networks to perform well for super-resolution. The GLCM was originally introduced in a 1973 paper and has been used in a few papers in the computer vision community. The paper trains and validates a super-resolution GAN (SRGAN) and a super-resolution CNN (SRCNN) on the DeepRock-SR dataset. Specifically, for the SRGAN, the paper uses the EDSRGAN network trained on loss function particular to the paper: The loss function consists of the addition of the L1 pixel-wise loss plus the VGG19 perceptual objective plus the proposed GLCM loss.\\nThe paper finds that SRCNN outperforms SRGAN in terms of the PSNR metric, but SRGAN performs better in terms of the spatial texture similarity. \\nNext, the paper shows that when trained with the mean L1-GMCM loss function, SRGAN performs best.\\n\\nIn summary, the paper proposes to use the GMCM loss for training and evaluation of the super-resolution methods. However, this metric is well known in the computer-vision community along with many others. Also, the idea to use this metric in training is only evaluated for one network (albeit a very sensible one) and only for one dataset (the DeepRock one). Since the novelty provided by the paper is small, I cannot recommend the acceptance of the paper.\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper addresses the super-resolution problem. The key is to use pixel co-occurrence-based loss metric. The idea is very straightforward. But the description could be clearer. For example, what is the spatial size of P (\\\\bar P)? How does it influence the optimization?\\n\\nEquation (2): There are four loss functions on the right hand. How are the loss defined?\\n\\nHow is the GAN used?\\n\\nIn experiments, there is no evidence showing the benefit from the pixel Co-occurrence\\n\\nThere is a lack of much details. Given the current presentation, I cannot judge if the quality reaches the ICLR bar.\"}" ] }

[ "Duygu Ataman", "Wilker Aziz", "Alexandra Birch" ]

[ "neural machine translation", "low-resource languages", "latent-variable models" ]

{ "note_id": [ "w8Ipi4yCb", "HkeTmSckqS", "S1lPyVNjFS", "rkgZFW2VYr", "SJl2KhcNtB" ], "note_type": [ "decision", "official_review", "official_review", "comment", "official_review" ], "note_created": [ 1576798730900, 1571951908850, 1571664862795, 1571238265005, 1571232900342 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1727/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1727/AnonReviewer1" ], [ "~Adithya_Renduchintala1" ], [ "ICLR.cc/2020/Conference/Paper1727/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Spotlight)\", \"comment\": \"This paper proposes a model for neural machine translation into morphologically rich languages by modeling word formation through a hierarchical latent variable model mimicking the process of morphological inflection. The model boils down to a VAE-like formulation with two latent representation: a continuous one (governed by a Gaussian) which captures lexical semantic aspects, and a discrete one (governed by the Kuma distribution) which captures the morphosyntactic function, shared among different surface forms. Even though the empirical improvements in terms of BLEU scores are fairly small, I find this a very elegant model which may foster interesting future research directions on latent models for NMT.\\n\\nThe reviewers had some concerns with some experimental details and model details that were properly addressed by the authors in their detailed response. In the discussion phase this alleviated the reviewers' concerns, which leads me to recommend acceptance. I urge the authors to follow the reviewer's recommendations to improve the final version of the paper.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes to incorporate a latent model (in the form of a variational auto-encoder) in the decoding process of neural machine translation. The motivation is to capture morphology. Experiments on three language-pairs (English to Arabic, Czech, and Turkish) show promising improvements in translation accuracy (BLEU).\\n\\nI am quite excited by this paper. So far there are not that many successful demonstrations of VAE in neural machine translation. The method is sound and interesting. The results show convincing improvements; some practioners may argue that reported BLEU gain (e.g. 0.8) is not impressive, but I think for a new model like this it is worthy.\", \"some_suggestions_or_questions\": [\"One alternative evaluation metric that might be interesting is to lemmatize both translation outputs and reference, then compute BLEU. This will help distinguish whether the proposed method is improving by getting the morphological inflections correct, or whether it is improving across the board on various word types.\", \"Table 3 is interesting. If there is space, I would suggest more analysis along those directions, i.e. investigating what morphology is learned, what is in the latent spaces.\", \"Do you think the results will vary depending on decoder layer depth? I wonder if different kinds of latent spaces will be learned with different depth.\", \"Also a related question is how about varying the source input BPE merge operation? Again, it seems like these design choices might affect results, especially when dealing with morphology.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proposes a method, Latent Morphology Model (LMM), for producing word representations for a (hierarchical) character-level decoder used in neural machine translation (NMT). The main motivation is to overcome vocabulary sparsity or highly inflectional languages such as Arabic, Czech, and Turkish (experimented in the paper). To model the morphological inflection, they decouple lemmas and inflection types into 2 latent variables (z and f) where f is enforced to be sparse (arguably mimic the process of the human). The literature review of NMT and the discussion on the potential advantage of morphology are concise. The proposed model is a variation of Luong & Manning (2016) and Schulz et al (2018) models, thus, their main contribution is an introduction of the latent morphological features to the decoder. The proposed LMM is trained by sampling z and f from prior directly, and the sparsity of the morphological features is encouraged by L0 of the feature vector (parameterized as independent Kumaraswamy variables). They perform the main empirical study using 3 languages by translating from English to justify the proposed LMM. Lastly, they provide a quantitative analysis of the perplexities of unseen words and a qualitative on words generated of a lemma with different feature vectors.\\n\\nOverall, this paper could provide a novel insight into the role of modeling morphology as latent variables. However, the experiments and analyses do not sufficiently support the claim of mimicking the process of the inflection (besides the gain in performance). Some clarification would be appreciated.\\n\\nIn section 3, the model, to my understanding, is agnostic to the morphological inflection, except the sparsity regularization (i.e., it could model any transformation including changing the lemma itself). In addition, there is nothing in the training particularly specific about the morphology. For example, z and f are generated from only the prior whereas we could get more accurate posteriors from observing the word itself. Some background on the language might help motivate the choice of model. Are morphological inflections ambiguous? Are the morphology labels hard to obtain? I think more discussion on previous attempts to model morphology (e.g., Vylomova et al 2017, Passban et al 2018) will be very helpful to the readers.\\n\\nFor the experiments (section 4), the model with LMM is shown to consistently outperform character and hierarchical models. The multi-domain performance also shows the model's ability to tackle a larger vocabulary set. However, we cannot certainly conclude that the gain comes from LMM successfully modeling the inflections. A contrast of performance with less morphological languages might reveal some insight (unfortunately I do not have enough knowledge to recommend languages). Finally, the feature variation analysis is interesting, but we only see one lemma from one language. Further discussion such as the consistency of features and the types of morphology, or the similarity of the lemma vector across context, will be helpful.\", \"minor_questions\": \"1. Given a word, how ambiguous it is to determine the stem and the morphological type in the subject languages?\\n2. How do you compute char-level PPL of the subword model?\\n3. In 4.4.4, did you obtain z of `go` from just translating `go` without any context?\"}", "{\"comment\": \"Thanks for posting on this interesting topic and the architecture proposed seems promising! Recently, some of our research has shown that for small datasets (such as TED) the BPE merge hyperparameter choice is very critical. We observe a 3-4 BLEU point difference simply by choosing a better BPE merge hyperparameter. Also, it seems like 32k for small data is generally a poor choice (see table 3 for details https://www.aclweb.org/anthology/W19-6620.pdf). Did you try other BPE choices?\", \"title\": \"Choice of subword merge rules\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper addresses the problem of translating into morphologically-rich languages, which suffers from the problem of sparse vocabularies and high numbers of rare and unseen words. In particular, current approaches such as subwords lack explicit notions of morphology and are obtained independently of the translation objective, while operating at the character-level renders the learning of long-distance dependencies more difficult.\\n\\nMore concretely, this paper models the generation of target words in a stochastic, hierarchical process, where the morphological features are modelled as latent variables. At each target word, the model samples a vector representing its lemma, followed by a k-dimensional latent inflectional features (e.g. nominative or accusative). To induce sparsity in the inflectional features, the paper uses a \\\"stretch-and-rectify\\\" distribution (Louizos et al., 2018) using the Kumaraswamy distribution. The paper further applies a sparsity-inducing regulariser to encourage the inflectional features to take discrete values of \\\"0\\\" or \\\"1\\\". Parameter estimation is done by optimising a lower-bound on the marginal log-likelihood. Experiments on translations into three morphologically-rich languages, English-{Arabic, Czech, Turkish}, indicate fairly small but consistent and statistically significant gains in BLEU. Further analysis indicates improved perplexity upper bound on rare words compared to subword-based baselines, along with somewhat interpretable latent inflectional features.\\n\\nOverall, this paper proposes an elegant solution to an important problem, and yields statistically significant BLEU improvements over the baselines. I have listed some pros and cons below, with a recommendation of \\\"Weak Accept\\\". I would be willing to raise my scores assuming my concerns are sufficiently addressed.\", \"pros\": \"1. The proposed approach is easy-to-follow and explained clearly, and the paper is overall well-written.\\n\\n2. Experiments are done on translation into three morphologically-rich languages with both concatenative and non-concatenative morphology, and results in small but mostly consistent BLEU improvements over the best subword, hierarchical, and character baselines for each language.\\n\\n3. The proposed approach is general and can potentially be applied on top of other model architectures (e.g. Transformers) and other language generation problems, such as language modelling, summarisation, etc, although the experiments in the paper focus on NMT with a GRU-based architecture.\", \"cons\": \"1. It is still unclear how much computational overhead is introduced by the approach (in terms of both training and prediction times), and how scalable the approach is when applied to language pairs that have much bigger training sets. In Tables 4 and 5, it seems that the largest dataset is English-Turkish multi-domain (434K sentence pairs), which is fairly small compared to other datasets.\\n\\n2. Regarding the feature variations result in section 4.4.4, it is hard to draw any conclusions about the latent inflectional features just based on varying the inflectional features at one position. I would suggest running the same experiment for multiple words at different positions, and see whether the same set of inflection features always results in similar inflections. This would better convince the reader that the latent inflectional features really are capturing useful morphological information, and that the target word generation process is appropriately controlled by the latent variable.\\n\\n3. Since some of the experiments did not use the same standard splits (e.g. the multi-domain Turkish experiments), it would be nice to report how well external models (e.g. OpenNMT) would do in the authors' dataset split, to make sure that the reported numbers here are at least comparable to that. This would help ensure the credibility of the findings.\\n\\n4. The effect of data size experiments (Section 4.4.2) for the character model seems somewhat counter-intuitive. Why would making the training set bigger (by incorporating data from a more diverse domain) make the character model worse? The explanation that \\\"[the character model]'s capacity cannot cope with the increased amount of sparsity\\\" does not seem satisfactory. Why does adding more data result in increased sparsity? Furthermore, one can simply use more hidden units or deeper layers to mitigate this problem in the character-based model. Assuming the authors are controlling for model capacity, then the character model's hidden state can be made bigger (since the vocabulary size is lower for character models, thus the embedding layer and the softmax layer are by definition smaller than the subword-based model).\", \"minor_suggestions\": \"1. Section 3.5 is not very clear; adding some figures or more explanation there would help understand how prediction is done.\\n\\n2. It would be interesting to explore other potential metric for measuring morphological generalisation. For instance, at evaluation time, does the model predict words that it has never seen before (e.g. predicting \\\"plays\\\", even though the model has only seen \\\"play\\\" at training time) at a higher frequency than character/subword/hierarchical baselines? If yes, this would provide more evidence that the latent inflectional features are properly capturing morphological information.\"}" ] }

[ "Erich Merrill", "Alan Fern" ]

[ "point cloud", "autoencoder" ]

{ "note_id": [ "szSFylL1gQ", "Hkx9xqR35B", "SkeaYr229H", "HkgILpeiFS" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730872, 1572821489833, 1572812165125, 1571650893779 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1726/AnonReviewer5" ], [ "ICLR.cc/2020/Conference/Paper1726/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1726/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The reviewers have raised several important concerns about the paper that the authors decided not to address.\", \"title\": \"Paper Decision\"}", "{\"rating\": \"1: Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #5\", \"review\": \"Summary\\n\\nThis paper studies the problem of 3D point cloud autoencoding in a deep learning setup, and in particular, the choice of the architecture of a 3D point cloud decoder. The main claim of the paper is the adaptation of the PointNet architecture (a set of fully-connected layers applied per point) from encoding to decoding purposes. To achieve that, the authors propose three similar mechanisms to inject noise in the bottleneck shape representation, which are used to sample per point feature vectors suitable for a PointNet-like decoder. These include appending a low dimensional gaussian noise sample to the bottleneck shape representation, adding individual gaussian noise sample to each bottleneck shape representation dimension or using trainable variances per bottleneck shape representation dimension. Resulting networks are optimized by minimizing Chamfer distance between the ground truth and output point clouds. The main motivation of their choice of these architectures is their ability to output an arbitrary size point clouds, compared to naive multilayered fully-connected network applied to a bottleneck shape representation and outputting a \\\"flattened\\\" fixed-size point cloud.\\n\\nTo verify their claims, the authors considered autoencoding task for ModelNet40 dataset. They compared a naive multilayered fully-connected decoder to PointNet-like decoders with all three proposed noise-injection mechanisms. This comparison was also performed for 5 variations of the proposed models differing is network size. To compare quantitatively, the authors present all results graphically, showing the dependency of resulting average Chamfer distance from the network size. This graph shows that using trainable per dimension variance for noise injection works consistently better, than other options; also, it shows inconsistent results for the baseline multilayered fully-connected decoder. In qualitative comparisons, the authors compare output point clouds visually and study the influence of sampling from various noise dimensions on the reconstructed points.\\n\\n\\nReview\\n\\nAlthough the authors of this paper touch a very interesting topic in generative modeling of point clouds, I do not believe this paper should be published in its current state.\", \"cons\": \"1. The main issue in this paper is the absence of comparisons to any external approaches. Even though comparison to their variation of the baseline model is provided, it is impossible to verify where their results stand compared to the related work. There is a whole stream of works in the same task using ShapeNet dataset, which the authors completely omit. Even if the authors do not want to perform experiments on ShapeNet, they, at least, could have provided comparison to AAE and PC-GAN results from PC-GAN paper [1] for generative modeling metrics on ModelNet40. As the authors stated, there are at least two works, PC-GAN and PointFlow, which allow sampling of the arbitrary size point clouds and comparison to these approaches is crucial (especially since the proposed approach is technically significantly simpler compared to hierarchical GANs from PC-GAN and continuous normalizing flows combined with VAEs from PointFlow papers).\\n\\n2. The paper misses numerous citations to related work in generative 3D shape modeling:\\n- voxel grid based, (since the authors mention them): [2, 3, 4, 5, 6, 7]\\n- point cloud based: [8, 9, 10]\\n- graph based: [11, 12]\\nAll these approaches propose various types of decoders for 3D shapes represented accordingly and some of them perform evaluations for pure generative, autoencoding task, which, again, can be used for comparison.\\n\\n3. I believe, there is a factual mistake and arguable diminishing of the current state of the art in the related work section:\\nThe authors state, that both PC-GAN and PointFlow models \\\"inhabit GAN settings\\\". However, PointFlow is, in fact, an approximate likelihood model and is considered as an instance of Variational Autoencoder with a shape prior and a point cloud decoder implemented as continuous normalizing flows (a special case of invertible neural networks). The authors also state, that it is \\\"unclear... if they are capable of outperforming the baseline MLP approach\\\" when they talk about these papers, while, in fact, both PC-GAN and PointFlow explicitly show that by comparing to [9], which use the multilayered fully-connected baseline decoder. It is also not unclear if these approaches need specific loss functions, contrary to what authors say. They were designed so they need their particular loss functions, which does not disqualify them from the comparison. These papers are the main competitors of this work and should be addressed properly instead of being diminished.\", \"pros\": \"1. Besides the cons, the paper is written clearly and is easy to follow.\\n\\n2. Rigorousness with respect to comparison of the proposed options (Table 1 and the fact that every model was retrained 15 times to obtain average result).\\n\\n3. Qualitative experiments with control over noise dimensions are novel and interesting.\\n\\nOverall, this could have been a paper about a nice and simple alternative to PC-GAN and PointFlow for arbitrary size point cloud modeling, but, unfortunately, it missed the opportunity to properly address the current state-of-the-art in its current version.\", \"small_remarks_and_questions_for_authors\": \"1. Citation of Valsesia at al. in the related work should be reorganized to be consistent with other citations. Omitting full names there and in case of PC-GAN will also be more consistent.\\n\\n2. All figures in the paper should be updated for better visibility of both graphics and text. Figures 1-4, 7 - text is barely visible, the scheme blocks and lines are too thin and overall lack readability. Figures 5-6, 8 - point clouds are almost invisible even with zooming, remove coordinate grid in the background, zoom in on the objects, so they take as much space in the images as possible, use bigger point sizes for better visibility.\\n\\n3. Maybe I do not follow something, why have you ignored applications of your point cloud decoders, and particularly generative modeling task and its evaluation protocols, presented for example in [9]?\\n\\n\\n[1] C.-L. Li, M. Zaheer, Y. Zhang, B. P\\u00f3czos and R. Salakhutdinov. Point Cloud GAN. In ICLR Workshop on Deep Generative Models for Highly Structured Data, 2019\\n[2] C. Choy, D. Xu, J.-Y. Gwak, K. Chen, and S. Savarese. 3D-R2N2: A unified approach for single and multi-view 3d object reconstruction. In ECCV, 2016.\\n[3] R. Girdhar, D. Fouhey, M. Rodriguez, and A. Gupta. Learning a predictable and gen- erative vector representation for objects. In ECCV, 2016.\\n[4] J. Wu, C. Zhang, T. Xue, W. Freeman, and J. Tenenbaum. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In NeurIPS, 2016.\\n[5] M. Tatarchenko, A. Dosovitskiy, and T. Brox. Octree generating networks: Efficient\\nconvolutional architectures for high-resolution 3D outputs. In ICCV, 2017.\\n[6] S. R. Richter and S. Roth. Matryoshka Networks: Predicting 3d geometry via nested shape layers. In CVPR, 2018.\\n[7] R. Klokov, J. Verbeek, and E. Boyer. Probabilistic Reconstruction Networks for 3D Shape Inference from a Single Image. In BMVC, 2019\\n[8] T. Groueix, M. Fisher, V. Kim, B. Russell, and M. Aubry. AtlasNet: A papier-m\\u00e2ch\\u00e9 approach to learning 3D surface generation. In CVPR, 2018.\\n[9] P. Achlioptas, O. Diamanti, I. Mitliagkas, and L. Guibas. Learning Representations and Generative Models for 3D Point Clouds. In ICML, 2018.\\n[10] P. Mandikal, K. Navaneet, M. Agarwal, and R. Babu. 3D-LMNet: Latent embedding matching for accurate and diverse 3D point cloud reconstruction from a single image. In BMVC, 2018.\\n[11] N. Wang, Y. Zhang, Z. Li, Y. Fu, W. Liu, and Y. Jiang. Pixel2Mesh: Generating 3D\\nmesh models from single RGB images. In ECCV, 2018.\\n[12] C. Wen, Y. Zhang, Z. Li, and Y. Fu. Pixel2Mesh++: Multi-View 3D Mesh Generation via Deformation. In ICCV, 2019.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I made a quick assessment of this paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"The paper deals with a task of learning latent representations for 3D point clouds by introducing a series of particular point cloud decoders. The advantage of the proposed architectures is their ability of producing variable-size output 3D point cloud. While the work on point-based architectures is relevant to the current agenda in the vision community, the novelty of this paper is somewhat limited, as the proposed architecture is constrained to one class (PointNets) only, while the decoding network could just as easily be used with other encoders (e.g., DGCNN [1]).\", \"major_comments\": \"One major shortcoming of the paper is its very limited experimental evaluation, that does not really allow to judge whether the proposed decoder architecture brings the real utility for most problems. For instance, authors only provide results for two distinct experiments: building an auto-encoder and understanding how each architecture uses noise. Much more tasks, such as shape denoising, upsampling, completion, or retrieval could have been explored, too. Only one (simple) dataset is used, i.e. ModelNet, whereas more complex datasets (e.g., ShapeNet, ABC, or datasets related to human shapes) could have been considered.\\n\\nI wonder why authors didn't compare to (or even cite) [1], which seems to be one of the earliest papers in the area. \\n\\nOther possible research questions are as follows. What is the effect of the proposed decoder networks on the learned latent representations in the entire architecture? Will the interpolatory and semantic properties of the learned latent spaces be preserved when using the autoencoders? \\n\\nHaving said that, I believe the paper cannot be accepted in the present form, however, I would change my decision, should more experiments validating the architecture be provided.\", \"minor_comments\": \"* in definition of a point cloud, both n and N are used to indicate its size\\n\\n\\n[1] Wang, Y., Sun, Y., Liu, Z., Sarma, S. E., Bronstein, M. M., & Solomon, J. M. (2019). Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5), 146.\\n[2] Achlioptas, P., Diamanti, O., Mitliagkas, I., & Guibas, L. (2018, July). Learning Representations and Generative Models for 3D Point Clouds. In International Conference on Machine Learning (pp. 40-49).\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper is about decoder networks for 3D point cloud data, i.e., given a latent vector that encodes shape information, the decoder network outputs a 3D point cloud. Previous approaches like PU-Net, or TopNet have used an MLP with a fixed number of output points for this task. The approach in this paper takes inspiration from the work of Li et al. and extends a sample based method for the decoder network. Specifically, the core contribution is to sample from a Gaussian distribution with the latent vector as mean and a variance computed by a small network conditioned on the latent vector. This sampled vector is then feed to an MLP to generate a 3D point estimate. The sampling approach is compared to the method of Li et al. (concatenating noise values to the latent value), and to directly adding Gaussian noise to the latent vector on auto-encoding shapes from the ModelNet40 dataset. Further studies on the impact of the noise on the reconstructed shapes are presented.\\n\\nI argue for a weak reject of this paper (in its current form) because (1) the presented method seems to be too incremental, and (2) I am missing experiments that compare to other state-of-the-art methods (e.g. on the task of 3D reconstruction from single images).\\n\\nRegarding the first point, let's start with the related work that is missing some very relevant paper. I would like to see the relation and comparison to implicit surface models (e.g., Park et al. \\\"DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation\\\", or Mescheder et al. \\\"Occupancy Networks: Learning 3D Reconstruction in Function Space\\\") that sample not based on a random vector, but on the 3D position itself. Further, I see Fan et al. \\\"A point set generation network for 3d object reconstruction from a single image\\\" in the references, but not where it is discussed in the paper. Further, Yang et al. \\\"FoldingNet: Point Cloud Auto-encoder via Deep Grid Deformation\\\" and Groueix et al. \\\"AtlasNet: A Papier-M\\u00e2ch\\u00e9 Approach to Learning 3D Surface Generation\\\" seem to be quite relevant.\\nFurther, the core contribution is to add a noise vector to the latent vector, instead of appending it to the latent vector as in Li et al. The presented evaluations in Fig. 3 and Fig. 4 show that this might be slightly beneficial, but I am not convinced how this performance increase would translate to a other tasks.\", \"this_is_also_my_second_major_point_of_criticism\": \"The presented evaluations are nice, but also seem like a toy problem. What would have been interesting is to compare the presented decoder network to state-of-the-art methods on other tasks presented in the literature. One example might be 3D reconstruction from a single image (cf. Fan et al.), or shape completion (cf. Park et al.). Even for the auto-encoding setting it would be possible to compare to other methods (cf. Yang et al., PointFlow).\\n\\nOne part that might lower the performance on the above mentioned task is the inferior encoder network (PointNet) that takes no neighbourhood information into account. There are quite a few stronger methods that can be considered, e.g. PointNet++, or any of the recently proposed point convolution methods.\\n\\nIn Fig. 3 and 4 the loss seems to be quite large, given that a Chamfer distance is used and the shapes are scaled to fit into a unit sphere. What function is used for dist? And are there severe outliers in the generated point clouds?\\n\\nOn page 5 it is stated that the Chamfer distance is used due to its relative speed. Relative to what?\\n\\nOn page 7 it is stated that the MLP results appear to be more evenly distributed. The last row in Fig. 5 suggests otherwise on the back of the chair. Also, the NoiseAdd and NoiseAppend seem to miss quite some details in the second row (or are just more noisy in general), which one is the proposed NoiseLearn?\", \"things_to_improve_the_paper_that_dis_not_impact_the_score\": [\"On the second pager, first line, there is an inconsistency in the citing style (use of parentheses)\", \"What is meant with a \\\"parameterized ReLU\\\" on page 4?\", \"The point cloud visualizations can be greatly improved. The grid is not necessary and the images are very small, making it hard to see details.\"]}" ] }

[ "Kai Lagemann", "Gregor Roering", "Christoph Henke", "Rene Vossen", "Frank Hees" ]

[ "model-free reinforcement learning", "model-based reinforcement learning", "Baysian neural network", "deep learning", "reinforcement learning" ]

{ "note_id": [ "ZRTkeVkYK", "BklIkSDXqH", "S1l3VHlRKS", "SJeE8rx6YH" ], "note_type": [ "decision", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730844, 1572201693643, 1571845428206, 1571779915592 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1725/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1725/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1725/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors propose a hybrid model-free/model-based policy gradient method that attempts to reduce sample complexity without degrading asymptotic performance. They evaluate their approach on a collection of benchmark tests.\\n\\nThe reviewers raised concerns about limited novelty of the proposed approach and flaws in the evaluation. The authors need to compare to more baselines and ensure that the baseline algorithms are performing as previously reported. Even then, the reported improvements were small.\\n\\nGiven the issues raised by the reviewers, this paper is not ready for publication at ICLR.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper presents a new model-free + model-based algorithm, MBPGE, that trains a policy using a policy gradient algorithm on top of the learned models. Contrary to previous approaches, they use a true Bayesian distribution by means of the randomized anchorized MAP. Furthermore, they combine rollouts from the real environment and from the learned dynamics directly for the policy training, instead of relying just on the ones of learned dynamics, which induces a larger distributional shift.\\n\\nThe paper readability could be improved. First of all, the introduction spans over 2 pages. The readability could be significantly improved by splitting the introduction into introduction (containing the motivation) and related work. Secondly, the preliminary section spans more than their approach, which hints that the original contribution is too incremental. All of the actual method is just combining section 2.1 and 2.2. Lastly, section 3.1 is hard to parse. The authors extend the paper to 9 pages when there was no need for it.\\n\\nIn terms of contribution, it seems that the main contribution is to use the method presented [1] and replace the ensemble with the ensemble of [2] and use PPO [3] instead of TRPO [4]. There is no further insight in the paper, but the fact that the ensemble of [2] works better than the classical ensemble (fact shown in [2]) and that PPO works better than TRPO (fact shown in [3]). The other contribution is to combine samples from the real environment and the from the learned models to train the policy. This however has actually been done in the context of model-based reinforcement learning, see [5]. It would be interesting that the authors compared against different heuristics to choose the ratio between imagined and real rollouts (i.e., constant, annealed, fraction of improvement [1], etc\\u2026). Overall, I don\\u2019t think that the contributions of this paper are enough for publication. \\n\\nRegarding the experiment section I strongly believe that the comparison is flawed. The results they report on MB-MPO do not match with the ones in [6, 7]. For instance, half-cheetah does not learn at all while it is an easier environment than Walker and Hopper (and there\\u2019s learning in those). Furthermore, given this they should also compare against ME-TRPO and ME-TRPO switching the TRPO with PPO, since this would be a fairer comparison. This section lacks also of a proper ablation analysis to identify how much each element of the proposed algorithm affects the performance the choice of Bayesian ensemble and heuristic to determine the amount of real samples and imagined ones.\\n\\nAt this stage, there is not enough contribution in terms of novelty nor delta in performance.\\n\\n\\n\\n[1] Thanard Kurutach, Ignasi Clavera, Yan Duan, Aviv Tamar, Pieter Abbeel. Model-Ensemble Trust-Region Policy Optimization.\\n[2] Tim Pearce, Felix Leibfried, Alexandra Brintrup, Mohamed Zaki, Andy Neely. Uncertainty in Neural Networks: Approximately Bayesian Ensembling.\\n[3] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, Oleg Klimov. Proximal Policy Optimization Algorithms.\\n[4] John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, Pieter Abbeel. Trust Region Policy Optimization\\n[5] Michael Janner, Justin Fu, Marvin Zhang, Sergey Levine. When to Trust Your Model: Model-Based Policy Optimization\\n[6] Ignasi Clavera, Jonas Rothfuss, John Schulman, Yasuhiro Fujita, Tamim Asfour, Pieter Abbeel. Model-Based Reinforcement Learning via Meta-Policy Optimization\\n[7] Tingwu Wang, Xuchan Bao, Ignasi Clavera, Jerrick Hoang, Yeming Wen, Eric Langlois, Shunshi Zhang, Guodong Zhang, Pieter Abbeel, Jimmy Ba. Benchmarking Model-Based Reinforcement Learning.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"### Summary ###\\n\\nThis paper focuses on model based reinforcement learning (RL). Specifically, the authors consider the setting of combining model based and model free RL algorithms by using the learned dynamics model to generate new data for training the model free algorithm. In order to capture the uncertainty of the environment and the model, the author applied Baysian neural network to learn the dynamics of the environment. The authors approximated the true Bayesian inference process by keeping an anchored ensemble of neural networks, where the prior and posterior of network weights are approximately Gaussian. The ensemble of dynamics model is then used to generate data to train a PPO[1] based agent. In order to prevent the agent from exploiting the learned dynamics model, the authors propose a heuristic way of balancing the amount of real data and model generated data by comparing the rewards.\\n\\nThe authors evaluate the proposed algorithm on simulated robotic locomotion environments in MuJoCo, and the results show that the proposed method has better same efficiency compared to baseline methods in some environments.\\n\\n\\n### Review ###\\n\\nOverall I think this paper presents an interesting idea in combing model based and model free RL algorithms. The idea is very well presented and authors include empirical evidence to support the proposed method. However I do find a number of shortcomings that need to be addressed.\", \"pro\": \"1. The idea for this paper is really well presented. The structure of the paper is well organized and the experiment results are easy to interpret.\\n\\n2. The authors provide a detailed description of the configurations and the hyperparameters for each experiments. Such description would be very helpful if the results in this paper are to be reproduced.\", \"con\": \"1. I\\u2019m not convinced about the magnitude of novelty in this paper. The proposed method seems very similar to ME-TRPO[2] and it seems to me that the novelty comes from the application of Bayesian ensemble techniques and the generated data ratio tuning heuristics. While these variations might be important for the final performance of the proposed method, the paper does not include any ablation study to further justify the importance of these variations.\\n\\n2. I\\u2019m not convinced about some of the performance of some of the baseline methods presented in this paper. In this paper, MB-MPO[3] does not improve at all during training on Half-Cheetah environment. However, in the original MB-MPO paper, the algorithm does improve and the performance seems to be comparable to that of the proposed method in this paper.\\n\\n3. The experiment results are not very strong for the proposed method. In 3 of 4 environments, the proposed algorithm does not show much advantage over the baseline algorithms. The only environment in which the proposed method shows significant improvement is Half-Cheetah, and I believe that the baseline algorithms might not be properly tuned in this environment.\\n\\n4. The paper lacks certain baseline comparisons. There are many other model based RL algorithms developed recently, and it would be important to compare to these methods. Some examples would be ME-TRPO[2], SLBO[4] and MBPO[5].\\n\\n\\nThe idea in the paper is well presented and carefully investigated. However, I am still not convinced about the novelty of the proposed idea and the magnitude of performance improvement. Therefore, I would not recommend acceptance before these problems are addressed. \\n\\n\\n\\nReferences\\n\\n[1] Schulman, John, et al. \\\"Proximal policy optimization algorithms.\\\" arXiv preprint arXiv:1707.06347 (2017).\\n\\n[2] Kurutach, Thanard, et al. \\\"Model-ensemble trust-region policy optimization.\\\" arXiv preprint arXiv:1802.10592 (2018).\\n\\n[3] Clavera, Ignasi, et al. \\\"Model-based reinforcement learning via meta-policy optimization.\\\" arXiv preprint arXiv:1809.05214 (2018).\\n\\n[4] Luo, Yuping, et al. \\\"Algorithmic framework for model-based deep reinforcement learning with theoretical guarantees.\\\" arXiv preprint arXiv:1807.03858 (2018).\\n\\n[5] Janner, Michael, et al. \\\"When to Trust Your Model: Model-Based Policy Optimization.\\\" arXiv preprint arXiv:1906.08253 (2019).\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"### Summary\\n\\nThis paper described a model-based RL method which uses learned dynamics models to augment the replay buffer used when training a PPO agent. \\nSpecifically, the agent learns an ensemble of dynamics models, and then performs a PPO updates using a mixture of trajectories sampled from the true environment and trajectories sampled from teh learned dynamics models. \\nThe fictitious trajectories are sampled from different dynamics models in the ensemble, which helps prevent the policy from exploiting the errors of a single model.\\nThe proportion of real/fake trajectories are adapted using the ratio of the real vs. predicted rewards: if the average returns are similar, then the proportions will be approximately equal, but if they are very different then more trajectories from the real environment will be used. \\n\\nOverall, I think there are some interesting elements to this paper but it is currently not ready for publication. I recommend weak reject for three reasons: limited novelty, the experiments as they stand are not very convincing and the writing needs work. \\n\\n\\n#### Novelty\", \"the_two_stated_contributions_are\": \"1. the use of principled uncertainty estimates in model-based RL (an anchored ensemble)\\n2. the routine which adjusts the proportion of real vs. simulated trajectories for learning the policy. \\n\\nIn my opinion, 1. is not much of a contribution since a number of works have already made use of different types of model uncertainty in the context of model-based RL. Here they use a slightly different version (the anchored ensemble, which is similar to the randomized priors used in DQNs) but this is more of a choice of implementation than a contribution in itself (a number of methods for estimating uncertainty exist, such as regular ensembles, dropout masks, BNNs, etc). The manner in which they use the uncertainty is to perform simulations with several different models to avoid overfitting to one, and this was already proposed in the ME-TRPO paper. \\n\\nThe idea of 2. is interesting and in general, adapting between model-based and model-free RL regimes during execution seems to have potential. However the current method seems currently under-explored and heuristic. The fact that it is sensitive to the \\\\alpha hyperparameter (Figure 2) is a bit troubling. Is there a principled way which does not require tuning this hyperparameter? Or reformulating things so that it is less sensitive?\\n\\n\\n#### Experiments\\n\\nAlthough the experiments show some improvement, they are not that convincing. The improvements do not seem statistically significant for Swimmer and Walker, and the asymptotic performance is worse than PPO for Hopper. There is only half-cheetah where the algorithm shows a clear improvment. \\n\\nSince one of the claimed contributions is the use of the anchored ensemble, there should be an ablation experiment showing this gives an improvement over a standard ensemble, but this is not included. \\n\\n\\n\\n#### Writing:\\n\\nOne issue with the paper is that it spends *much* too long discussing related work and preliminaries. \\nThe first 5 pages are devoted to this and the proposed method is only introduced on page 6!\", \"some_detailed_comments\": [\"Paragraph 1 in the intro should be drastically cut. The sentence \\\"Recently...challenging tasks\\\" could be followed by references and then a sentence discussing the sample inefficiency and then moving on to the next paragraph. Most readers will be familiar with DQN, actor critic etc.\", \"Same for paragraphs 2 and 3. A short discussion of the general idea behind model-based RL and its improved sample efficiency, the issue of the policy exploiting model errors, and some references are sufficient.\", \"Similarly, section 2.1 is mostly unnecessary. It isn't necessary to detail the updates for standard algorithms such as PPO/TRPO unless they are useful for the proposed algorithm.\", \"Section 2.2 is also too long. They main idea is very simple and is summarized in Equation 13.\"]}" ] }

[ "Muyuan Fang", "Qiang Wang", "Jian Zhang", "Zhao Zhong" ]

[ "neural architecture search", "weight sharing", "auto machine learning", "deep learning", "CNN" ]

{ "note_id": [ "6n5hzSh9ma", "rkxrlpTisS", "BkxkWIvUiS", "rygU3SP8jH", "BkeEGBDIiH", "SJxVImW65r", "BkxCrcjmcH", "rkg_5UjRYr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730816, 1573801197156, 1573447158950, 1573447086178, 1573446923631, 1572832075617, 1572219462451, 1571890831624 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1724/Authors" ], [ "ICLR.cc/2020/Conference/Paper1724/Authors" ], [ "ICLR.cc/2020/Conference/Paper1724/Authors" ], [ "ICLR.cc/2020/Conference/Paper1724/Authors" ], [ "ICLR.cc/2020/Conference/Paper1724/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1724/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1724/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes a neural architecture search method that uses balanced sampling of architectures from the one-shot model and drops operators whose importance drops below a certain weight.\\n\\nThe reviewers agreed that the paper's approach is intuitive, but main points of criticism were:\\n- Lack of good baselines\\n- Potentially unfair comparison, not using the same training pipeline\\n- Lack of available code and thus of reproducibility. (The authors promised code in response, which is much appreciated. If the open-sourcing process has completed in time for the next version of the paper, I encourage the authors to include an anonymized version of the code in the submission to avoid this criticism.)\\n\\nThe reviewers appreciated the authors' rebuttal, but it did not suffice for them to change their ratings.\\nI agree with the reviewers that this work may be a solid contribution, but that additional evaluation is needed to demonstrate this. I therefore recommend rejection and encourage resubmission to a different venue after addressing the issues pointed out by the reviewers.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Additional Response to Reviewer #2\", \"comment\": \"We're sorry to say that the code and models could not be open sourced until the approaval from our organization. We have filed the application for releasing code and models with the legal affairs department. This procedure might cost several weeks. Our training curves and experiments settings for all reported models have been uploaded on Github( https://github.com/BetanasICLR2020/BetaNAS ) in details. Other materials will be updated once they are permitted. We believe that our method is easy to follow and re-implement. If anyone has some questions on the method and experiments, we will try our best to help other researchers to re-implement the method.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"Thanks for your instructive comments and suggestion.\\n1)\\nWe have updated our paper with evaluation experiments according to \\\"Evaluating the Search Phase of Neural Architecture Search\\\". More experiments with different random seeds are conducted to compare our methods with randomly sampled architectures from the same search space. Our method outperforms the randomly sampled ones and other competing methods, as shown in Section 4.3.2 and Fig.6. \\n\\n2) The suggestion is meaningful and instructive. We have made a discussion on recently weight sharing methods in Tab(3) and Section 4.3.4, to compare their training approaches, searching strategies, efficiency, etc. These papers and \\\"Evaluating ...\\\" are introduced to related works and analyzed in Experiments and Discussion sections.\\n\\nTable 3. Comparison with the other weight sharing methods on their strategies.\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\\n| Method | supernet optimization | search policy | sample sub-net | mutual interference |\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\\n| ENAS | imbalanced | RL | \\u221a | from all candidates |\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\\n| Darts | imbalanced | gradient | x | from all candidates |\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\\n| SinglePath | balanced | EA | \\u221a | from all candidates |\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\\n| ProxyLess | imbalanced | gradient/RL | \\u221a | from all candidates |\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\\n| Progressive| imbalanced | gradient | x | from fewer candidates|\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\\n| Darts+ | imbalanced | gradient | x | from all candidates |\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\\n| HM-NAS | imbalanced | gradient+mask| x | from all candidates |\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\\n| Ours | balanced | gradient | \\u221a | from fewer candidates |\\n+----------------+--------------------------------+--------------------+-----------------------+--------------------------------+\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"Thanks for your meaningful comments and suggestion. The model architectures, weights and training hyper-parameters of the reported BetaNet-A and BetaNet-B will be open sourced in Google Drive in 2 ~3 days. We believe that our method is easy to follow and re-implement.\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"Thanks for your helpful comments.\\n1)\\\"Matthew Effect\\\" means that the rich get richer and the poor get poorer (wiki), originally. We use this description as an analogy to describe the unfairly training strategy in some weight sharing methods. In those methods, paths with better performance at early searching phases can obtain more training opportunities than other candidates, and further enhance their performance. \\n2) alpha is calculated according to Eq (4). We first normalize alpha to its softmax form p, after that a path will be dropped once its corresponding p is lower than th_\\\\alpha. In our experiments, th_\\\\alpha is set to a constant 0.5. We found that overall performance is largely robust (insensitive) to that parameter (th_\\\\alpha), with less than 0.3% top1 variation when varying 0.4\\u2264 th_\\\\alpha \\u2264 0.6. On one hand, when th_\\\\alpha is set too high, some promising paths might be dropped at early stages, leading to degraded performance. On the other hand, when th_\\\\alpha is set too low, paths will be dropped slow and we need more searching steps to keep the amount of finally remaining paths small enough. th_alpha is set to 0.5 to ensure that when trained to convergence, there are 2 or 3 paths remaining in each choice block. \\n3) Thanks very much for your constructive and detailed comments. We will fix the typos.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper relates to automatic neural architecture search techniques. Current methods have certain drawbacks: Some train all network paths to convergence, which wastes computational efforts in unpromising paths, whereas some don't train all the branches uniformly, which can lead to unfair comparisons.\\n\\nThe authors propose a model to balance the two issues mentioned above. Their aim is to produced balanced training while trying to reduce conflicts between the different potential network paths. So their algorithm Has two phases, one where it randomly builds a block in the network and another one where it discards vertices form the layer that are below a certain threshold. \\n\\nThe experimental results seem sound to me, and I think this is a reasonable approach.\", \"here_some_general_comments_that_would_help_with_clarity\": \"1) I think the authors should explain more clearly what \\\"Matthew Effect\\\" in the introduction\\n\\n2) It's not very clear to me how th_\\\\alpha is computed. Could this please me made more specific. Section 4.4 says that \\\"only the operators with performance much lower than the average of others will be dropped.\\\" Is this approach conservative? Did they try different thresholds?\\n\\nThere are several minor typos that the authors might want to correct.\\n\\n1) There is a couple of spaces missing like between \\\"2018)have shown\\\" in page 1\\n - The word \\\"probability\\\" and p_1, p_2 in (1)\\n - Eq6 on page 5\\n\\n2) It is customary to use commas before and after \\\\ldots if one is listing a sequence. The authors don't do this in any of their lists, and this is very strange. \\n\\n3) In the description of Algorithm 1, I'd change \\\"S_max is denoted as\\\" for \\\"S_max denotes\\\"\\n\\n4) In (7) I think there is a \\\"{\\\" and a\\\"}\\\" missing before and after the o_{l,m_i}. It's set notation. \\n\\n5) In the discussion, they write the word \\\"differently.\\\" Would it be better to write \\\"by contrast\\\"?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"In this paper, the authors proposed a new training strategy in achieving better balance between training efficiency and evaluation accuracy with weight sharing-based NAS algorithms. It is consisted of two phrases: in phrase 1, all path are uniformly trained to avoid bias, in phrase 2, less competitive options are pruned to save cost. The proposed method achieved the SOTA on IN mobile setting.\\n\\n\\nOverall I found the idea proposed in the paper intuitive and convincing. Especially I appreciate the ablation study that identified one of the option is encouraged too much in the early stage will can lead to worse final performance. From the methodology perspective, I think this is a solid incremental contribution. However, the highlight of this paper, in my opinion, is the SOTA results on IN. My main concern is that the authors did not indicate that the code/model will be open sourced, which will help verification as well as reproducibility.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper introduces a better searching strategy in the context of automatic neural architecture search (NAS). Especially, they focus on improving the search strategy for previously proposed computationally effective weight sharing methods for NAS. Current search strategies for the weight sharing NAS methods either focus on uniformly training all the network paths or selectively train different network paths with different frequency, where both have their own issues like wasting resources for unpromising candidates and unfair comparison among network paths. To this end, this paper proposes a balanced training strategy with \\u201cselective drop mechanism\\u201d. Further, they validate their approach by showing leading performance on ImageNet under mobile settings.\\n\\nOverall, I appreciate the effort on exploring new strategies for better search algorithm for NAS in the context of weight sharing methods. However, the analysis of the current approach is limited and some relevant papers are missing. More details below.\", \"arguments\": \"1) This paper\\u2019s main focus is on developing a better search strategy for NAS based weight sharing methods. The validation of their approach is supported via showing improvement in the accuracy on ImageNet task. However, my main argument is that you have to validate the proposed approach with better analysis on the search space along with improvement on the tasks. For example, \\u201cEvaluating the Search Phase of Neural Architecture Search\\u201d paper which studies the effectiveness of current search strategies is not referenced in this paper. I would suggest to do analysis of you search strategy by following the analysis experiments conducted in the above paper. \\n\\n2) The related work doesn\\u2019t discuss the latest papers which suggest why and in what ways the current search strategies are based for weight-sharing based NAS approaches. As this paper is trying to address this problem, I would naturally assume that this is well discussed, but it\\u2019s missing!\"}" ] }

[ "Bowen Tan", "Zhiting Hu", "Zichao Yang", "Ruslan Salakhutdinov", "Eric Xing" ]

[ "Sequence generation", "sequence prediction", "reinforcement learning" ]

{ "note_id": [ "0hzK1f5oJF", "HJgYDw2osr", "rJgImv3sor", "Bkev-vnsoB", "S1xROF9CFr", "HylY7KLAKB", "S1xxM8j8Fr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730787, 1573795680976, 1573795613951, 1573795583326, 1571887478378, 1571871008522, 1571366407931 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1723/Authors" ], [ "ICLR.cc/2020/Conference/Paper1723/Authors" ], [ "ICLR.cc/2020/Conference/Paper1723/Authors" ], [ "ICLR.cc/2020/Conference/Paper1723/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1723/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1723/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors construct a weighted objective that subsumes many of the existing approaches for sequence prediction, such as MLE, RAML, and entropy regularized policy optimization. By dynamically tuning the weights in the objective, they show improved performance across several tasks.\\n\\nAlthough there were no major issues with the paper, reviewers generally felt that the technical contribution is fairly incremental and the empirical improvements are limited. Given the large number of high-quality submissions this year, I am recommending rejection for this submission.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Official Blind Review #1723\", \"comment\": \"Thanks for the comments! We\\u2019d like to clarify that this paper aims to reformulate the various algorithms and distill them into a single common formulation. The common formulation is governed by the reward function and two weight hyperparameters, and thus defines a *family* of sequence prediction algorithms. Changing the specifications of the three factors (i.e., reward and weights) leads to different specific algorithms. It\\u2019s indeed novel and non-trivial to reformulate these apparently distinct algorithms and discover the common underlying formulation. We will update the presentation (as also suggested by R#1) accordingly to make this contribution clearer.\\n\\nThe interpolation algorithm is an immediate production of the discovered common formulation, as it\\u2019s a natural idea, once we see the common formulation, to anneal the three governing factors to \\u201cinterpolate\\u201d between the specific algorithms in the family. We add the discussion of the annealing settings in the main paper. It\\u2019s also worth noting that the interpolation algorithm is just one of the many possible ways of taking advantage of the common formulation. For example, another natural idea would be to find (e.g., through hyperparameter optimization) the best configuration of the three governing factors in the common formulation, which is equivalent to finding the best optimization algorithm in the whole family and use it to learn sequence prediction. This shows the advantages of having the common formulation. We will make this clearer in the revised version.\\n\\nThis paper is the first to discuss the extensive set of algorithms (MLE, RAML, Data Noising, policy gradient, etc) jointly and find their common denominator. The resulting interpolation algorithm can also be seen as a generalization of previous approaches, such as MIXER, that \\u201cuses MLE and RL collaboratively\\u201d. (As discussed in the paper, MIXER can be seen as using a restricted annealing strategy in the proposed interpolation algorithm). The generalized interpolation algorithm also outperforms MIXER.\\n\\nWe will fix all typos and do proofread. Thanks for pointing this out.\"}", "{\"title\": \"Response to Official Blind Review #1\", \"comment\": \"Thanks for the great suggestion of adjusting the order of presentation! We agree that it can be clearer and more analytical by reaching the final formulation after visiting each individual algorithms. We will update the presentation in the revised version accordingly.\\n\\nBesides MT, the improvement in game imitation learning is indeed reasonably significant, especially when the number of expert demonstrations is small:\\n ** On HalfCheetah-v2 with 4 demonstrations, our approach achieves *3000 higher* reward than GAIL; On Ant-v2, our approach achieves *800 higher* reward on average.\\n ** The improvement level over GAIL is comparable or stronger than that in other papers (e.g., Fig.2 of NeurIPS-2018 https://arxiv.org/pdf/1805.08336.pdf; Fig.1 of ICML-2019 https://arxiv.org/pdf/1901.09387.pdf) which proposed to improve GAIL in different aspects orthogonal to ours.\\n\\nThe improvement on summarization is relatively moderate, partially because the output sequences of the task are short (8.2 tokens on average), while the proposed approach is designed for generating longer sequences (e.g., in the MT dataset, the output sequences contain 18.5 tokens on average).\"}", "{\"title\": \"Response to Official Blind Review #2\", \"comment\": \"Thanks for the valuable and encouraging comments.\\n\\n(1) Thanks for the great question and suggestion! The average output length in the machine translation dataset is 18.5. That is, each target sequence contains 18.5 tokens on average. In contrast, the length is 8.2 in the text summarization dataset. We speculate these different output lengths lead to the different variances and improvement --- the proposed approach by design brings more benefits when the output sequences are longer (e.g., MT) by combating the compounding error. The approach thus gets relatively significant improvement and lower variance compared to the other methods on MT; while on summarization, the approach achieves moderate improvement and does not reduce the variance. We will add more analysis and discussions in the revised version.\\n\\n(2) Thanks for the suggestion! As above, the improvement on the summarization task is moderate due to the short output sequences in the dataset. We will conduct experiments on image captioning and add the results and analysis.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper presents a formalism of entropy regularized policy optimization. They also show that various policy gradients algorithms can be reformulated as special instances of the presented novel formalism. The only difference between them being the reward function and two weight hyperparameters. Further, the paper proposes an interpolation algorithm, which, as training proceeds, gradually expands the exploration of space by annealing the reward function and the weight hyperparameters. Experiments on text generation tasks and game imitation learning show superior performance over previous methods.\\n\\nOverall, the paper is well written and the derivations and intuitions sound good. I appreciate the overall effort of the paper and the thorough experiments to validate the proposed interpolation algorithm, results seem not significant for text summarization. Hence, I suggest a week accept for this paper.\", \"arguments\": \"1) From Table 1 and Table 2, the proposed approach has the lowest variance on machine translation and the quiet opposite on the text summarization (i.e., it has high variance). Any thoughts on this? This also suggests to conduct experiments on ablating the variance in the training for various policy gradient approaches include the proposed one. \\n\\n2) Results seem not significant on the summarization tasks. Any thoughts on choosing this particular task? Why not try image captioning where most of these policy gradient approaches have been applied.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This submission belongs to the field of sequence modelling. In particular, this submission presents a unified view on a range of training algorithms including maximum likelihood (ML) and reinforcement learning (RL). The unified view presented I believe is interesting and could be of interest to a large community. Unfortunately this submission has two issues 1) presentation and 2) experimental validation.\\n\\nI find it peculiar that an objective function that features ML and variants of RL as special cases called ERPO is proposed by statement. I find it more likely that it came out by analysing ML, the variants of RL and other commonly used objective functions, noticing similarities between them and then formulating a function that would render all above as special cases. Had the order been different this submission would have been much more analytical and interesting to read. \\n\\nI find experimental results a bit limited and not entirely conclusive as it seem that MT provides the only strong experimental evidence. I find quite hard to interpret the significance of difference, for instance, between 36.72 and 36.59 in ROUGE-1.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1723\", \"review\": \"This paper claims to propose a general entropy regularized policy optimization paradigm. MLE and RL are special cases of this training paradigm. Paper is well written, and the experimental results are convincing enough.\\nHowever, there are still some minor problems in the paper. For the optimization framework ERPO (shown in Equation 1), it consists of three parts, a cross-entropy term (Shannon entropy), a $p,q$ KL divergence term, and a reinforcement learning reward loss item. From the framework point of view, it is not like the author claim that is supposed to present a general optimization framework, including various optimization algorithms. Instead, it is just a combined loss through weight control and the selection of corresponding functions. It may not really theoretically work to unify various types of optimization algorithms for general cases, let alone claiming that this is a general optimization algorithm framework. \\n\\nFor the interpolation algorithm (I regard this is the true technical contribution of this paper), the authors used an annealing mechanism to use different weights and functions at different stages of training. The essence is that after MLE pre-training, different optimization algorithms are used in different stages, and this should be the focus of the article. The annealing settings used is only introduced in the appendix simply. Without more comparison experiments, we cannot clearly get the conditions for the annealing algorithm to be effective and ineffective. \\n\\nFor the title of connecting the dots between MLE and RL, this paper did not do so, MLE and RL are only used collaboratively, and this has also been mentioned in previous work.\\n\\ntypo\\nPage 6 Paragraph \\u201cOther Algorithms & Discussions\\u201d: We We show in the appendix\\u2026 -> We show in the appendix\\u2026\"}" ] }

[ "Maximilian Baader", "Matthew Mirman", "Martin Vechev" ]

[ "adversarial robustness", "universal approximation", "certified network", "interval bound propagation" ]

{ "note_id": [ "kncwFMK75M", "Hyl3mPnOsB", "BygpXxpDjH", "BkefWiKwir", "ByehldKDjS", "r1gcEwFvsS", "Hyg928YvoH", "S1lbHrKwiB", "HJgjoVkWcr", "Hkg_xk_TFH", "SkenkkVpFB", "SyefKVndur" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment" ], "note_created": [ 1576798730758, 1573599012436, 1573535780856, 1573522170189, 1573521396283, 1573521201675, 1573521074500, 1573520697474, 1572037794670, 1571811055755, 1571794660487, 1570452601532 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1722/Authors" ], [ "ICLR.cc/2020/Conference/Paper1722/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1722/Authors" ], [ "ICLR.cc/2020/Conference/Paper1722/Authors" ], [ "ICLR.cc/2020/Conference/Paper1722/Authors" ], [ "ICLR.cc/2020/Conference/Paper1722/Authors" ], [ "ICLR.cc/2020/Conference/Paper1722/Authors" ], [ "ICLR.cc/2020/Conference/Paper1722/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1722/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1722/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1722/Authors" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This work shows that there exist neural networks that can be certified by interval bound propagation. It provides interesting and surprising theoretical insights, although analysis requires the networks to be impractically large and hence does not directly yield practical advances.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Answer\", \"comment\": \"Thank you for the questions.\", \"q\": \"Is there really a meaningful distinction between considering continuous and Lipschitz-continuous functions as on any compact set Lipschitz functions are dense in the continuous functions?\", \"a\": \"We believe it is a meaningful distinction to make. The reason is that our theorem handles continuous functions directly and has the advantage of certifying arbitrary precise networks with arbitrary small error $\\\\delta$. To see the advantage, consider function $\\\\sqrt{x}$ for $x \\\\in [0,1]$ (continuous but not Lipschitz continuous) approximated with Lipschitz functions. Here, the Lipschitz constant of the approximation tends to infinity as the approximation improves. This implies that the error $\\\\delta$ gets arbitrarily large (in the language of [2], certifiable regions get arbitrarily small).\\n\\nFinally, we note that we focus on a different certification methodology than [1] and [2]: [1] and [2] focus on certification via Lipschitz constants, while we focus on certification via linear relaxations. We are happy to clarify this point better in the paper.\"}", "{\"title\": \"On network size\", \"comment\": \"Thank you for your response. Can you clarify what is the best upper bound on network size that can be given? You say \\\"This drastically reduces the number of neurons\\\", but it is not clear to me what the new result should be. It sounds from the rest of this response that it should still be exponential?\\n\\nA related question would be what network size you believe this approach should give (independently of what you can currently prove).\\n\\n> we consider any continuous function, while [1] and [2] considers the more restricted class of Lipshitz-continuous functions. \\n\\nIs this really a meaningful distinction? On any compact set Lipschitz functions are dense in the continuous functions. The fact that [1] and [2] are also exponential seems like the more relevant point here.\"}", "{\"title\": \"Answer\", \"comment\": \"We thank Reviewer 2 for their comments, suggestions and in particular for carefully checking the derivations. We reframed the Abstract and Introduction as suggested and added a small paragraph on what the theorem means in the case of a given dataset.\", \"q\": \"Do you provide an implementation following the construction of your theorem?\\nYes, we provide a Python implementation here [ https://glot.io/snippets/fhrl5ouae3 ], and a Haskell implementation here [ https://glot.io/snippets/fhrl8jt8pp ].\"}", "{\"title\": \"Answer\", \"comment\": \"We thank Reviewer 1 for their comments. We would like to clarify, that the network needed to represent a bump function consists of $6m-1$ ReLUs grouped in $2 + \\\\lceil \\\\log_2(2m+1) \\\\rceil $ layers. Thus, the number of neurons needed to represent a bump function does not grow exponentially with the input dimension $m$. Nevertheless, Reviewer 1 is right that the worst case scaling of the size network with its current construction is exponential. In the comment for Reviewer 3, we provide a (straightforward) improvement. However, in the worst case, the number of bump functions still grows exponentially with the input dimension. Further discussion is given in the comment to Reviewer 3 [ https://openreview.net/forum?id=B1gX8kBtPr&noteId=Hyg928YvoH ].\"}", "{\"title\": \"Answer (2)\", \"comment\": \"*On Randomised Smoothing vs. Linear Relaxation Guarantees*\\n\\nWe updated the paper to clarify the relationship to randomised smoothing in the paper. In randomised smoothing, the classifier is a smoothed version of a base classifier (e.g. a neural network). It is infeasible to evaluate the smoothed classifier exactly. An approximated version of the smoothed classifier is constructed using Monte Carlo estimation. Thus the classifier is *not* a neural network, but a randomly sampled neural network. The certification is not given for a neural network, but for a randomly sampled one. \\n\\nFurther, contrary to neural networks, the classification is probabilistic (with high confidence). Similarly, certification is also probabilistic (with high confidence). While we think that randomised smoothing is a promising idea, the guarantees provided by the two methods differ.\\n\\nFor further discussion on comparing the guarantees provided by smoothing with linear relaxations, please refer to a recent comment [4] by Jeremy M. Cohen (author of the randomized smoothing paper [3]). \\n\\n*Regarding state-of-the-art results on MNIST and CIFAR10:*\\n\\nConsidering CIFAR10 with $l_\\\\infty$ perturbations up to 2/255, Gowal et. al. [8] report an accuracy of 70.1% with certified robustness of 50% while Wong et. al. [9] report an accuracy of 68.3% with reported accuracy of 53.9%. Considering $l_\\\\infty$ perturbations up to 8/255, Mirman et. al. report an accuracy of 46.2% with certified robustness of 27.2% (see [6]) while Zhang et al. [5] report an accuracy of 41.3% with certified robustness of 28.8%. \\nWhile the work of Raghunathan et. al. [7] is very interesting, it does not scale to networks of the size needed to achieve state-of-the-art certification and accuracy.\\n\\n[1] Anil et. al.: https://arxiv.org/abs/1811.05381\\n[2] Jeremy E. Cohen: https://arxiv.org/abs/1904.04861\\n[3] Jeremy M. Cohen et. al.: http://proceedings.mlr.press/v97/cohen19c.html\\n[4] Jeremy M. Cohen: https://openreview.net/forum?id=SJxSDxrKDr&noteId=HJlVRSuADr\\n[5] Zhang et. al.: arXiv preprint arXiv:1906.06316\\n[6] https://github.com/eth-sri/diffai\\n[7] Raghunathan et al.: http://papers.nips.cc/paper/8285-semidefinite-relaxations-for-certifying-robustness-to-adversarial-examples\\n[8] Gowal et. al.: https://arxiv.org/abs/1810.12715\\n[9] Wong et. al.: http://papers.nips.cc/paper/8060-scaling-provable-adversarial-defenses\"}", "{\"title\": \"Answer (1)\", \"comment\": \"We thank Reviewer 3 for their comments, which we address below.\\n\\n*On network size from our construction:*\\n\\nThe focus of this work was to present a proof which establishes the existence of interval certifiable networks; our focus was not on an efficient construction. \\n\\nNevertheless, we do not sum over all boxes in the Grid, as Reviewer 3 claims; we sum over all boxes fulfilling $f(B) \\\\geq \\\\xi_{k+1}+ \\\\delta/2$. One can instantly obtain a more efficient construction by modifying the construction of the set $\\\\Delta_k$ slightly: Instead of adding all boxes $B$ fulfilling $f(B) \\\\geq \\\\xi_{k+1}+ \\\\delta/2$, we only add the largest box $B\\u2019$ containing $B$, fulfilling $f(B\\u2019) \\\\geq \\\\xi_{k+1}+ \\\\delta/2$ (if there are multiple largest boxes, pick any one). This drastically reduces the number of neurons. We leave the refinement of this heuristic as future work. \\n\\n*Related work [1] and [2] consider more restricted functions than us and are either non-constructive or are worst-case exponential in the dimension:*\\n\\nReviewer 3 claims works [1] and [2] are related and produce networks of more reasonable sizes than our work. \\n\\nFirst, our work considers a more general class of functions than [1] and [2]: we consider any continuous function, while [1] and [2] considers the more restricted class of Lipshitz-continuous functions. \\n\\nSecond, despite the fact [1] and [2] consider more restricted functions, the proof of [1] seems to actually by non-constructive, while the proof of [2] produces networks of size that is exponential in the worst-case, like we do.\", \"to_elaborate_more\": \"- The proof given in [1] does not seem to be constructive: [1] shows that the set of Lipschitz networks with the GroupSort activation function are dense in the space of Lipschitz functions. \\n- While the construction given in [2] is constructive, its worst case complexity is also exponential in the input dimension $m$:\\n - Consider the piecewise linear function $f \\\\colon [0,1]^m -> [0,1]$, mapping the center point of $[0,1]^m$ to 1, from \\n which $f$ decreases in a piecewise linear fashion to 0 at the boundary of $[0,1]^m$. It can be easily seen, that $f$ is \\n made up out of $2m$ affine functions in the sense of Lemma 1 in [2]. \\n - Next, we define the function $g \\\\colon [0,d]^m -> [0,1]$, mapping the center points of the cubes defined by the integer \\n grid to 1, while $g$ decreases in a piecewise linear fashion to 0 at the boundary of every integer cube. We see that $g$ \\n is made up out of $2m * d^m$ affine functions in the sense of Lemma 1 in [2].\\n - The width of the first hidden layer of the 3 layer FullSort network in Equation (11) in [2] is equal to the number of affine \\n functions, that is in our case for $g$, the width is $2m * d^m$ which is exponential in $m$ (input dimension).\\n\\nFurther, when it comes to network depth, the two methods are not comparable. That is, if we compare the depth of the network in [2] with $3 + \\\\lceil log_2(2m+1) \\\\rceil$ from our work, we observe that the depth is non-comparable, because [2] uses 2 FullSort layers while we use ReLU layers. Using the FullSort activation function, we could immediately reduce the number of layers to 1 FullSort layer followed by 2 ReLU layers by replacing the $nmin_{2m}$ and $R_[*,1]$ in the construction of the local bump functions with the FullSort activation. \\n\\nPlease note that there are cases where our construction yields significantly smaller networks compared to the one constructed in [2]. Considering the function $g$ from above and $\\\\delta = 2$, our construction would return the network mapping all inputs to 0, while the construction of [2] would return a network of width $2m * d^m$. \\n\\n[1] Anil et. al.: https://arxiv.org/abs/1811.05381\\n[2] Jeremy E. Cohen: https://arxiv.org/abs/1904.04861\\n[3] Jeremy M. Cohen et. al.: http://proceedings.mlr.press/v97/cohen19c.html\\n[4] Jeremy M. Cohen: https://openreview.net/forum?id=SJxSDxrKDr&noteId=HJlVRSuADr\\n[5] Zhang et. al.: arXiv preprint arXiv:1906.06316\\n[6] https://github.com/eth-sri/diffai\\n[7] Raghunathan et al.: http://papers.nips.cc/paper/8285-semidefinite-relaxations-for-certifying-robustness-to-adversarial-examples\\n[8] Gowal et. al.: https://arxiv.org/abs/1810.12715\\n[9] Wong et. al.: http://papers.nips.cc/paper/8060-scaling-provable-adversarial-defenses\"}", "{\"title\": \"General comment on the updated paper:\", \"comment\": [\"We would like to thank the reviewers for the feedback and comments. We incorporated the suggested changes and uploaded the updated version of the paper. In particular, we changed Abstract, Introduction and Related Work to improve overall presentation and clarity. We\", \"added clarifications and reframed the abstract and intro as suggested by Reviewer 2,\", \"provide both, Haskell [ https://glot.io/snippets/fhrl8jt8pp ] and Python [ https://glot.io/snippets/fhrl5ouae3 ] code of the construction which can be run directly,\", \"added the clarifications and citations suggested by Reviewer 3.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper proves the universal approximation property of interval-certified ReLU networks. Such a network achieves the best certified accuracy against l_infty adversaries at large epsilons on MNIST and CIFAR10, but the certified accuracy is still far from satisfaction. With the results from this paper, we can at least make sure the interval-certified ReLU networks is able to fit the robust classifier with arbitrarily small error if it exists.\\n\\nI did not check the details of the proofs, but I think such a theorem is a great reference for future work. \\n\\nHowever, we cannot conclude whether a network with a moderate size exists to fit the robust classifier well from this paper. From the construction of local bumps (nmin), the number of neurons needed to represent a bump function grows exponentially with the input dimension, which makes the size of the network vacuously large when the input is images. We are not getting any closer to improving the certified robustness with this paper.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary:\\nThis paper proves a theoretical result, that for every continuous function f, there exists a ReLU network approximating f well, and for which interval propagation will result in tight bounds.\", \"the_proof_is_by_construction\": \"the function to approximate is decomposed into a sum of functions (slices) with bounded range, which will be approximable to the correct precision. Each slices is approximated as a sum of local bumps, who essentially approximates an indicator function over a grid.\", \"comments\": [\"Equation (1) of Theorem 1.1, I assume that the left inclusion could be replaced by [l, u], given that interval propagation generate valid bounds? Or is the case that due to the approximating network being only an approximation, interval propagation on it might not give actual bounds on the real function?\"], \"opinion\": \"The paper is very technical and takes some effort to parse to understand the construction, but the authors do a good job of providing examples and illustrations to make it at least intuitive. The content is clearly novel and the result is far from trivial.\\n\\nThe only thing that i'm slightly wary is the significance of what is proven. The paper shows that approximating a continuous function with a network such that bound propagation works well on it is feasible.\\nOn the other hand, the motivation given is \\\"given a dataset and a specification, is there a network that is both certified and accurate with respect to these\\\", which is a slightly different problem. If this was proven to be yes, then we would know that when our robust training fails, this is due to optimization not being good enough or our not network not having enough capacity. The result proven here does not suffice, because it pre-suppose the existence of a correct function, that would match the dataset and specification, which is not guaranteed (see for example the proofs that robustness and accuracy are at odds, or you can just imagine verification of robustness for too-large radius which would lead to infeasible specification). I still like the paper and I think it might just be a matter of framing the abstract and introduction.\\n\\nDo the authors have an implementation, even for only toy-size examples, of the generation of a Neural Network approximating a given function or is this purely theoretical? Especially when the proof is constructive like this, this can strengthen the confidence that no subtle mistakes have been made. If so, this should be mentioned.\", \"potential_typos\": \"\", \"page_4\": \"\\\"Further\\\" -> \\\"Furthermore\\\"?\", \"page_5\": \"\\\"Increasing the fines\\\" -> \\\"Increasing the finesse?\\\"\\nPage 6, just before Lemma 4.3, missing a mathcal before the N?\", \"disclaimer\": \"I carefully checked the derivations in the paper but only assessed the sensibility of the proofs in the appendix.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The paper aims to show that there exist neural networks that can be certified by interval bound propagation. It claims to do this by showing that for any function f, there is a neural network g (close to f) such that interval bound propagation on g obtains bounds that are almost as good as one would get by applying interval bound propagation to f.\\n\\nThere are a couple issues with this paper. The first is that the main theorem does not imply the claimed high-level take-away in the paper in any practically relevant regime. This is because while the paper does show that there is a network that approximates f well enough, the size of the network in the construction is exponential (for instance, one of the neural network components in the construction involves summing over all hyperrectangles lying within a grid). The result is only plausibly practically relevant if there is a polynomially-sized network for which the bound propagation works.\\n\\nThe second issue is the comparison to related work, which omits both several key techniques in the literature and two related papers on universal approximation.\\n\\nOn universal approximation papers, see these two: https://arxiv.org/abs/1904.04861 and https://arxiv.org/abs/1811.05381. They consider a different proof strategy but more plausibly yield networks of reasonable size.\", \"on_key_techniques_in_the_literature\": \"\\\"typically by employing methods based on mixed integerlinear programming, SMT solvers and bound propagation\\\" ignores several major techniques in the field: convex relaxations (SDP + LP), and randomized smoothing. Similarly, \\\"specific training methods have beenrecently developed which aim to produce networks that are certifiably robust\\\" again ignores entire families of techniques; Raghunathan et al. train SDP relaxations to perform well, while Cohen et al. train randomized smoothing to work well. Importantly, randomized smoothing is not about creating an over-approximation of the network but explicitly constructs networks that are smooth.\\n\\n\\\"some of the best results achieved on the popular MNIST (LeCun et al., 1998) and CIFAR10(Krizhevsky, 2009) datasets have been obtained with the simple Interval approximation (Gowalet al., 2018; Mirman et al., 2019)\\\"\\n-Is this actually true? I just looked at Mirman et al. and it doesn't seem to explicitly compare to any existing bounds. My impression is that randomized smoothing (Cohen et al.) currently gets the best numbers, if we're allowed to train the network. If not allowed to train the network, Raghunathan et al. (Neurips 2018) performs well and is not compared to in the Mirman paper.\\n\\nThese omissions must be addressed as the introduction and related work is misleading in its current state.\\n\\nFinally, some writing issues:\\n>>> While the evidence suggests \\\"no\\\", we prove that for realisticdatasets and specifications, such a network does exist and its certification can beestablished by propagating lower and upper bounds of each neuron through thenetwork\\n\\nOne cannot \\\"prove\\\" something for \\\"realistic datasets\\\", since \\\"realistic datasets\\\" is not a formal assumption. Please fix this.\\n\\n>>> \\\"the most relaxed yet computationally efficient convex relaxation\\\"\\n\\nWhat does most relaxed mean? Most relaxed would most straightforwardly mean outputting the entire space as the set of possibilities at each point, which is clearly not intended.\"}", "{\"comment\": \"To be fully clear, we note that B := [a,b] for Theorem 1.1.\", \"title\": \"Clarification\"}" ] }

[ "Nikaash Puri", "Sukriti Verma", "Piyush Gupta", "Dhruv Kayastha", "Shripad Deshmukh", "Balaji Krishnamurthy", "Sameer Singh" ]

[ "Deep Reinforcement Learning", "Saliency maps", "Chess", "Go", "Atari", "Interpretable AI", "Explainable AI" ]

{ "note_id": [ "K0df5XHrRTd", "7_AytLUgQXr", "cyuuQfZ2kn", "SkekPRl2sS", "BylD_6l3sS", "BJeXh9ehsH", "Hyeu8gpssS", "HJe916UjsH", "SJg5oTWtsB", "Hyxvkl6uoH", "S1xBnuYNir", "Byx0WCDViS", "rJxZZaPEjS", "Bkeeh2vEoB", "Skx402YnYB", "rJx9cWb3YS", "rJlbK2tLOB" ], "note_type": [ "official_comment", "comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1589545807950, 1588731269302, 1576798730730, 1573813846942, 1573813614870, 1573812907386, 1573797968184, 1573772513788, 1573621154014, 1573601247463, 1573324973393, 1573318150410, 1573317880678, 1573317799847, 1571753163971, 1571717522103, 1570311288882 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Paper1721/Authors" ], [ "~Akanksha_Atrey2" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1721/Authors" ], [ "ICLR.cc/2020/Conference/Paper1721/Authors" ], [ "ICLR.cc/2020/Conference/Paper1721/Authors" ], [ "ICLR.cc/2020/Conference/Paper1721/Authors" ], [ "ICLR.cc/2020/Conference/Paper1721/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1721/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1721/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1721/Authors" ], [ "ICLR.cc/2020/Conference/Paper1721/Authors" ], [ "ICLR.cc/2020/Conference/Paper1721/Authors" ], [ "ICLR.cc/2020/Conference/Paper1721/Authors" ], [ "ICLR.cc/2020/Conference/Paper1721/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1721/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1721/AnonReviewer1" ] ], "structured_content_str": [ "{\"title\": \"Thanks for the comment\", \"comment\": \"Hey, thanks for the comment. To generate their maps, we directly used the code they had made available with their paper. We observed that for some frames in Atari their approach generates focused saliency maps, while for others, it generates noisy saliency maps. In contrast, SARFA generates focused saliency maps for most frames. We selected illustrative frames from the gameplay to show this distinction. These look different from the ones in their paper because they chose some other illustrative frames. The full videos for the gameplay showing SARFA Saliency are uploaded to our repository along with the full videos generated by Greydanus et al.\"}", "{\"title\": \"Discrepancy in SOTA Saliency Map Method\", \"comment\": \"Interesting work! It is important to develop saliency map methods that capture relevant and specific features. This type of taxonomy is also useful for generating and evaluating explanations in the future. We have a similar paper in ICLR '20 on saliency maps demonstrating this need along with the need to evaluate saliency maps appropriately (https://openreview.net/pdf?id=rkl3m1BFDB). I do have one question. If you compare figure 2 in your paper and Figure 1, 3 and 4 in Greydanus et al. (https://arxiv.org/pdf/1711.00138.pdf), the saliency maps differ by quite a bit despite being on the same environment. I am wondering if there are some hyperparameters that have not been set accordingly which may have lead to these differing maps and could have possibly impacted the user study where visual output of saliency maps is evaluated.\"}", "{\"decision\": \"Accept (Poster)\", \"comment\": \"A new method of calculating saliency maps for deep networks trained through RL (for example to play games) is presented. The method is aimed at explaining why moves were taken by showing which salient features influenced the move, and seems to work well based on experiments with Chess, Go, and several Atari games.\\n\\nReviewer 2 had a number of questions related to the performance of the method under various conditions, and these were answered satisfactorily by the reviewers.\\n\\nThis is a solid paper with good reasoning and results, though perhaps not super novel, as the basic idea of explaining policies with saliency is not new. It should be accepted for poster presentation.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Additional revision\", \"comment\": [\"We have uploaded another revision to the paper, here are the salient changes from the last revision:\", \"Added saliency maps for deep RL agents for Chess in Appendix C.\", \"Perturbed non-salient parts of the board to observe the effect on the actions, as an approach to evaluate the quality of the saliency map (Section 3.5).\", \"Added more discussion in Section 5.\", \"Minor additions to the description of the properties in Section 2.\"]}", "{\"title\": \"Additional changes\", \"comment\": \"Thank you for your comments. We have made some additional changes in the latest revision.\\n\\n> the measure does particularly well for generating saliency maps for Stockfish\\n> might be quite different \\u2026 for deep neural network RL agents. \\n> v) The saliency method might be particularly suited for Stockfish (whose action-value estimates might be strongly influenced by human saliency and chess theory). \\n> See whether the method still produces good results for other chess agents\\n\\nAdded saliency maps for a deep learning agent for Chess to Appendix C.\\n\\n> iii) Show that the saliency map stays roughly constant for seemingly irrelevant perturbations. \\n> I would still like to encourage the authors to try and come up with some perturbations that should have very predictable effects on the saliency\\n\\nAdded results in Section 3.5\"}", "{\"title\": \"Additional desired properties\", \"comment\": \"Thank you for your comment. Based on our initial evaluation, it is very challenging to derive a scalable approach purely based on using specificity and relevance as the \\u201caxioms\\u201d, due to the need to evaluate exponentially many subsets of features, made much worse due to the fact that many such perturbations will be invalid or not meaningful. However, this direction is definitely useful, and we will explore this in future work.\\n\\nIn regards to whether specificity and relevance are sufficient, and whether there are additional desired properties for saliency, our extensive experiments with humans for puzzle-solving and on the chess saliency dataset show that our approach (specificity + relevance) performs significantly better than existing approaches. There may be other additional properties though, especially in light of the discussion of the limitations in Section 5, that we will explore in future work.\"}", "{\"title\": \"Perturbation and Gradient-Based methods\", \"comment\": \"Agreed, both perturbation and gradient methods are useful for different use cases; we mention the difficulty in identifying meaningful perturbations in Section 5.\"}", "{\"title\": \"Thank you for the thorough response and the improvements to the manuscript\", \"comment\": \"Most of my issues have been addressed to a satisfying degree during rebuttal. I will therefore raise my score accordingly. As I stated before, the main strength of the paper to me is the effort that the authors made to evaluate the quality of the produced explanations. Evaluations of this kind (and rigor) are severely lacking from large bodies of work on interpretability/saliency.\\n\\nI would still like to encourage the authors to try and come up with some perturbations that should have very predictable effects on the saliency (such as adding non-relevant pieces in non-relevant positions, or moving all pieces on the board by two rows/columns, or \\\"mirroring\\\" a scenario along the vertical axis as well as swapping colors). While this is tricky to do in full generality, I am sure that it is possible for particular situations. In the best case it would make the results even stronger, otherwise it would add some critical reflection and current shortcomings to be addressed in future work.\"}", "{\"title\": \"A small question\", \"comment\": \"I would like to thank the authors for the reply.\\nI have a small question if there is any desired properties other than Specificity and Relevance.\\nSome studies (e.g. SHAP, IntegratedGrad) considered these desired properties as \\\"axioms\\\", and developed saliency methods as the ones that satisfy the axioms.\\nI am interested in to see if a similar axiomatic approach is applicable for this study.\"}", "{\"title\": \"Perturbations\", \"comment\": \"Thank you for adding this line of related work. Your criticisms are perfectly valid. Though, for what it's worth, having working out a \\\"meaningful\\\" perturbation may be difficult depending on the domain, so this is both a pro and a con.\"}", "{\"title\": \"Thank you for the positive feedback\", \"comment\": \"We would like to thank the reviewer for the positive feedback. We will work on further improving the quality of the paper over subsequent versions.\"}", "{\"title\": \"Response to the comments\", \"comment\": \"We would like to thank the reviewer for their comments, many of the comments are very useful and will address them in the following revisions.\\n\\n> saliency maps are rarely evaluated rigorously. \\n\\nIndeed, evaluating saliency maps is a very challenging problem. Thank you for appreciating our attempt at evaluation; note that it is fairly comprehensive compared to most of the published approaches. We included the lack of generalization in Section 5.\\n\\n> The main goal of explainability/interpretability methods must be to come up with testable hypotheses that tell us something about how the artificial system makes its decisions in novel/unseen situations.\\n\\nOne of the goals of interpretability is indeed to describe the extrapolation/generalization of the agent, however, the goals of interpretability are multifold. Just knowing the reasons behind a single prediction/action can be quite useful, as we show. Note that this is aligned with the existing literature in \\u201clocal explanations\\u201d, that has been shown to be useful for interpreting, debugging, developing trust, and for the users learning the task. However, we agree with the limitations, and include it in Section 5.\\n\\n> iv) Apply non-relevant perturbations to the salient pieces. I.e. take the same puzzle/scene and move it across the board, does the saliency-map move in (roughly) the same way.\\n\\nUnfortunately, it is not easy to identify non-relevant perturbations, since moving the relevant pieces across the board, for example, may completely change the underlying strategy that is applicable, making different pieces salient or not salient.\\n\\n> vi) Add a section that discusses current shortcomings and caveats with the method, and saliency maps for explainability in general.\\n\\nWe have added such a section (Section 5).\\n\\n> vii) (Experimental details) For each domain, please explain the perturbations that you used (removing pieces in Chess/Go, blurring pixels in Go, anything else e.g. replacing pieces?). \\n> e) Please provide all experimental details (in the appendix) that are necessary to reproduce the experiments. Referring to a code-repository is not a replacement for describing the methods in detail.\\n\\nWe have added such a section (Appendix A).\\n\\n> In all experiments, was it always the action with the highest Q-value that was being explained?\\n\\nYes, we only explain the actions with the highest Q-value, since greedy policies are often used at test time. However, we have also included examples of saliency maps for other actions, in Appendix B.\\n\\n> b) Chess saliency dataset. Are the expert saliency ratings binary? Why not have multiple degrees of saliency?\\n\\nYes, the saliency ratings are binary. Different degrees of saliency is harder to obtain. For instance, in a Chess position, it is much easier for experts to label whether pieces are important than to estimate (or agree on) the degree of importance. \\n\\n> c) Would the Greydanus et al. 2018 approach deliver similar results when using a threshold to cut off low-saliency inputs?\\n\\nWe feel that the \\u201cthresholding\\u201d should be part of the explanation technique; they should provide a sparse representation of which features are relevant (all of the baselines do already zero out irrelevant features). It is difficult to evaluate different thresholding amongst the features that they have identified as salient, since the threshold itself may be crucially dependent on the state or the value function, and it is not possible to search the different threshold values at each state. Nonetheless, we do provide an evaluation that is independent of the effect of the threshold (using ROC curves) but instead depends on the ranking of the features.\\n\\n> d) Why is the saliency map in 3.4 binary (pieces are either salient or not)? How was the binarization threshold chosen? What would the non-binary saliency maps look like?\\n\\nWe have updated the saliency maps in Section 3.4 to the original non-binary saliency maps output by our approach.\"}", "{\"title\": \"Comparison to gradient-based approaches\", \"comment\": \"We have expanded the review of these approaches in the revised Section 4, including the connection of the axioms of attribution to perturbation-based techniques. More empirically speaking, Greydanus et al compared their approach against a gradient-based approach, and they noted in their paper that such saliency maps \\u201ccan be difficult to interpret\\u201c, \\u201cchoose perturbations which lack physical meaning\\u201d, and \\u201cmay move it off from the manifold of realistic input images\\u201d. In our paper, we show that our explanations are more useful than Greydanus et al. Note that the gradient-based techniques not only need white-box access to the agent, but even with white-box access, some agents may not be fully differentiable, such as Stockfish for chess.\"}", "{\"title\": \"Revised version with additional content\", \"comment\": [\"Thank you for the helpful and constructive comments on our paper. In order to address some of the comments by the reviewers, we have uploaded a revision of the paper, addressing the following:\", \"Added more background information for readers unfamiliar with the Chess notation as a footnote in the Introduction when we first introduce Figure 1.\", \"Updated the saliency maps in Figure 7 and Section 3.4 to show the relative saliency (instead of the binarized map as in the original version).\", \"Expanded the text on gradient-based approaches to saliency maps (non-RL) in Section 4 and their connection to our proposed work.\", \"Added a discussion section describing the shortcomings of our approach, and of saliency maps (Section 5).\", \"Added more details about the experiment setup, in particular, the perturbation technique for each of the evaluations (Appendix A).\", \"Added examples of the saliency maps for actions that were not the highest Q (in Appendix B), in particular, for the top three moves in a particular chess position.\"], \"we_are_currently_carrying_out_a_few_more_experiments_for_the_next_revision\": [\"Evaluating and analyzing the saliency maps for deep RL agents for Chess.\", \"Perturbing non-salient pieces for Chess, and evaluating how much the saliency map changes.\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"[score raised from weak reject to weak accept after rebuttal - on a more fine-grained scale, I would rate this paper now an accept (7), but not a strong accept (8), however since this year's scale is quite coarse, I'll stick with a score of 6]\\n\\nSummary\\nThe paper proposes a new perturbation-based measure for computing input-saliency maps for deep RL agents. As reported in a large body of literature before, such saliency maps are supposed to help \\u201cexplain\\u201d why a deep RL system picked a certain action. The measure proposed in the paper aims at combining two aspects: specificity and relevance, which should ensure that the saliency map highlights inputs that are relevant for a particular action to be explained (specificity), and this particular action only (relevance). The paper shows illustrative examples of the proposed approach and two previously proposed alternatives on Chess, Go and three Atari games. Additionally the paper evaluates the method and the two previous alternatives on two interesting chess-tasks: chess-puzzles where human players were shown to be able to solve puzzles faster and with higher accuracy when given the proposed saliency map in addition, and evaluation against a curated dataset of human-expert saliency maps for 100 chess puzzles. \\n\\nContributions\\ni) Novel, perturbation-based saliency measure composed of two parts. Main idea of specificity is (similar to Iyer et al. 2018) to use State-Action value function (Q-function) with a specific action, instead of State-Value function only. Main idea of relevance is to \\u201cnormalize\\u201d by taking into account change in Q-value for all other actions as well. Both parts are combined in a heuristic fashion.\\n\\nii) More objective/reproducible assessment of how saliency maps produced by different methods overlap with human judgement of saliency of pieces in chess. To this end: two experiments with human chess players/experts (puzzle solving, and expert-designed saliency maps).\\n\\nQuality, Clarity, Novelty, Impact\\nThe paper is well written and most sections are easily understandable, though for some parts it might help if the reader is quite familiar with Chess/Go. The proposed saliency measure seems to address some shortcomings of previously proposed measures - my main criticism is that the construction of the measure seems rather ad-hoc and heuristic. It would have been great to formally define specificity and relevance (e.g. in an information-theoretic framework as Sparse Relevant Information) and then derive a suitable measure that is shown to satisfy/approximate the formal definitions. At least, there is one ablation study that justifies parts of the heuristic construction to some degree. \\nThe proposed measure seems to do reasonably well on board-game domains, in particular chess. However it might also be the case that the measure does particularly well for generating saliency maps for Stockfish (which is the agent that happens to be evaluated in the chess domain), which might be quite different from previously reported methods that have been designed for deep neural network RL agents. The illustrative examples on Atari and Go do not allow for a statistically significant judgement of the quality of the proposed method.\\n\\nOn a conceptual level, a bigger issue (of many saliency methods in general, but the criticism also applies to the paper) is that the \\u201cexplanations\\u201d drawn from saliency maps are rarely evaluated rigorously. The paper makes a nice attempt by trying to establish some \\u201cground-truth\\u201d saliency in chess using humans to increase the degree of objectivity, which I greatly appreciate. However, it remains unclear whether explanations that happen to coincide with human notions of saliency really are of higher quality for assessing how an artificial system makes its decisions. The main goal of explainability/interpretability methods must be to come up with testable hypotheses that tell us something about how the artificial system makes its decisions in novel/unseen situations. The goal is not to explain a decision mechanistically after the fact (which is trivial, given a deterministic, differentiable feed-forward system), but to come up with non-trivial explanations that extrapolate/generalize. Specificity and Relevance are probably important ingredients of such explanations, but I think it\\u2019s important to formalize them properly first. Currently I am in favor of suggesting a major revision of the work, but I am happy to reconsider my decision based on the rebuttal and other reviewers\\u2019 assessment. Having said that, I do want to reiterate that I think it is great that the authors included some ablation studies and measures of \\u201cusefulness\\u201d of the saliency method.\\n\\nImprovements / Major comments\\ni) Formally define specificity and relevance (e.g. as sparsity and compression?). Ideally derive a saliency measure based on these formal definitions.\\n\\nii) Show how saliency maps (of the same situation) change when producing explanations for different actions. I assume that currently the illustrative examples only show the action with the highest Q-value, what changes when using e.g. a less likely action?\\n\\niii) Show that the saliency map stays roughly constant for seemingly irrelevant perturbations. In particular, using the chess-dataset with expert annotations, apply various kinds of perturbations to non-salient pieces (e.g. removing them, swapping them for other pieces and potentially moving them in irrelevant ways) and see whether the AUC stays roughly constant.\\n\\niv) Apply non-relevant perturbations to the salient pieces. I.e. take the same puzzle/scene and move it across the board, does the saliency-map move in (roughly) the same way.\\n\\nv) The saliency method might be particularly suited for Stockfish (whose action-value estimates might be strongly influenced by human saliency and chess theory). See whether the method still produces good results for other chess agents (ideally trained without human heuristics or data). If this is hard to do for chess, think about a different application where this is easier.\\n\\nvi) Add a section that discusses current shortcomings and caveats with the method, and saliency maps for explainability in general.\\n\\nvii) (Experimental details) For each domain, please explain the perturbations that you used (removing pieces in Chess/Go, blurring pixels in Go, anything else e.g. replacing pieces?). In all experiments, was it always the action with the highest Q-value that was being explained?\\n\\n\\n\\nMinor comments\\n\\na) Table 1: (add error-bars) What is the variance across players? Are the results for the proposed method statistically significant?\\n\\nb) Chess saliency dataset. Are the expert saliency ratings binary? Why not have multiple degrees of saliency?\\n\\nc) Would the Greydanus et al. 2018 approach deliver similar results when using a threshold to cut off low-saliency inputs?\\n\\nd) Why is the saliency map in 3.4 binary (pieces are either salient or not)? How was the binarization threshold chosen? What would the non-binary saliency maps look like?\\n\\ne) Please provide all experimental details (in the appendix) that are necessary to reproduce the experiments. Referring to a code-repository is not a replacement for describing the methods in detail.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes an algorithm for explaining the move of the agents trained by reinforcement learning (RL) by generating a saliency map.\\nThe authors proposed two desired properties for the saliency map, specificity and relevance.\\nThe authors then pointed out that prior studies failed to capture one of the two properties.\\nTo combine the two components into a single saliency map, the authors proposed using the harmonic mean.\\n\\nThe experimental results demonstrated that the proposed saliency map successfully focused only on important parts while the other method tend to highlight some irrelevant parts also.\\nThe authors also did a great job for evaluating the goodness of the saliency maps, by preparing a human annotated chess puzzle dataset.\\n\\nI think the paper is well-written, and the basic idea look reasonable and promising.\\nThe experimental evaluations are well designed and the results look convincing.\\nSaliency map for RL is not yet mature, and I expect to see further improvements (especially, more theoretically principled ones) follow this study.\\n\\n\\n### Updated after author response ###\\nThe response from the authors seem to be reasonable to me. I therefore keep my score.\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #1\", \"review\": \"Interpreting the policies of RL agents is an important consideration if we would like to actually deploy them and understand their behaviours. Prior works have applied saliency methods to DRL agents, but the authors note that there are two properties - specificity and relevance - that these methods do not take into account, and therefore result in misleading saliency maps. The authors define (and provide examples) of these properties, propose simple ways to calculate these (and like prior methods, relying on perturbations and therefore applicable to almost black box agents), and combine them neatly using the harmonic mean to provide a new way to calculate saliency maps for agents with discrete action spaces. While the improvements on Atari are hard to quantify, the results on chess and go are more interpretable and hence more convincing. The study using saliency maps to aid human chess players is rather neat, and again adds evidence towards the usefulness of this technique. Finally, the chess saliency dataset is an exciting contribution that can actually be used to quantify saliency methods for DRL agents. The proposed method is relatively simple, but is well-motivated and demonstratively (quantitatively, which is great work for general interpretability methods) better than prior methods, and in addition the authors introduce a new dataset + quantitative measure for saliency methods, so I would give this paper an accept.\\n\\nAlthough the authors motivate their choice of perturbation-based saliency methods as opposed to gradient-based methods, they should expand their review of the latter. As the technique the authors introduced is very general, it would be useful to know how it compares to the current state of research in terms of identifying properties that attribution methods should meet - a good example of this is integrated gradients (Sundararajan et al., 2017), which similarly identify \\\"sensitivity\\\" and \\\"implementation invariance\\\" as \\\"axioms\\\" that their method satisfies.\"}" ] }

[ "Hae Beom Lee", "Hayeon Lee", "Donghyun Na", "Saehoon Kim", "Minseop Park", "Eunho Yang", "Sung Ju Hwang" ]

[ "meta-learning", "few-shot learning", "Bayesian neural network", "variational inference", "learning to learn", "imbalanced and out-of-distribution tasks for few-shot learning" ]

{ "note_id": [ "Vgm-PPd3W", "rked9EZ2oH", "SJxmPQW3jH", "rkgunfZ2sS", "B1xVgzZ3iS", "S1lLvaCTYr", "HygSWGg2tS", "rJep4Itotr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730701, 1573815440303, 1573815130654, 1573814960446, 1573814763586, 1571839326480, 1571713533484, 1571685940661 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1719/Authors" ], [ "ICLR.cc/2020/Conference/Paper1719/Authors" ], [ "ICLR.cc/2020/Conference/Paper1719/Authors" ], [ "ICLR.cc/2020/Conference/Paper1719/Authors" ], [ "ICLR.cc/2020/Conference/Paper1719/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1719/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1719/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Talk)\", \"comment\": \"The reviewers generally agreed that the paper presents a compelling method that addresses an important problem. This paper should clearly be accepted, and I would suggest for it to be considered for an oral presentation.\\n\\nI would encourage the authors to take into account the reviewers' suggestions (many of which were already addressed in the rebuttal period) and my own suggestion.\\n\\nThe main suggestion I would have in regard to improving the paper is to position it a bit more carefully in regard to prior work on Bayesian meta-learning. This is an active research field, with quite a number of papers. There are two that are especially close to the VI method that the authors are proposing: Gordon et al. and Finn et al. (2018). For example, the graphical model in Figure 2 looks nearly identical to the ones presented in these two prior papers, as does the variational inference procedure. There is nothing wrong with that, but it would be appropriate for the authors to discuss this prior work a bit more diligently -- currently the relationship to these prior works is not at all apparent from their discussion in the related work section. A more appropriate way to present this would be to begin Section 3.2 by stating that this framework follows prior work -- there is nothing wrong with building on prior work, and the significant and important contribution of this paper is no way diminished by being up-front about which parts are inspired by previous papers.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Summary of the updates in the revision\", \"comment\": [\"We thank all reviewers for their insightful and constructive comments. Based on the comments, we updated the paper by making the following changes:\", \"We included the comparison between Bayesian TAML and regularized deterministic TAML suggested by Reviewer #1 (R1), in Table 6.\", \"We included the dataset names in Table 4 and 5 as suggested by R1.\", \"We added in more details of the experimental setup for the experiments in Section 5.2, as suggested by R1.\", \"We revised the discussions on the metric-based meta-learning approaches in the Related Work section, as suggested by R1.\", \"We added in standard errors to all the results in Section 5.2 and 5.3 as suggested by Reviewer #3 (R3).\", \"We reported the values for the three task-specific balancing variables for the complete Bayesian TAML model in Appendix E, as suggested by R3.\", \"We have corrected the phrase \\\"obtains significantly improves over\\\" to \\\"significantly improves the performance\\\", as suggested by R3.\", \"We added the experiments for $\\\\theta_0$ initialization in Appendix F, as suggested by Reviewer # 2 (R2).\", \"We believe that our paper became significantly stronger with this revision with all the additional experimental results and corrections, thanks to the suggestions from the reviewers.\"]}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We really appreciate your constructive comments. We respond to each comment as follows.\\n\\n1. How was the gradient w.r.t. $\\\\theta$ was computed?\\n\\n- Basically, computation of gradients w.r.t. initial $\\\\theta$ in our model is exactly the same as the original MAML framework. Suppose that the number of inner-gradient steps is 5 and k-th inner-gradient update is denoted as a function $u_k$. Then, from the initial model parameter $\\\\theta_0$, we have\\n\\n$\\\\theta_1 = u_1(\\\\theta_0)$\\n$\\\\theta_2 = u_2(\\\\theta_1)$\\n\\u2026\\n$\\\\theta_5 = u_5(\\\\theta_4)$\\n$\\\\text{Loss_test} = \\\\mathcal{L}(\\\\theta_5, \\\\mathcal{D}_\\\\text{test})$\\n\\nThus it is trivial to compute the gradient of Loss_test w.r.t. $\\\\theta_0$ using the chain rule.\\n\\n\\n2. Why constrain the task-specific learning rate decaying factor to a value between 0 and 1, while using softplus() that does not have such a constraint for the class-specific scaling factor?\\n\\n- Note that the task-specific learning rate decays as follows:\\n\\t\\n$\\\\alpha$, $\\\\alpha f(\\\\gamma)$, $\\\\alpha f(\\\\gamma)^2$, $\\\\alpha f(\\\\gamma)^3$, \\u2026\\n\\t\\nWithout such restriction to force it between 0 and 1, the learning rate may largely diverge due to its exponential form.\\n\\nOn the other hand, since the class-specific scaling factor is simply a multiplicative coefficient, there is no such restriction, except for the positivity constraint. Thus we used softplus function to model the class-specific scaling factor. \\n\\n\\n3. Applying a linear interpolation between $\\\\theta$ and a randomly-initialized vector in Eq (2) considering zero-entries caused by $g(z^{\\\\tau})$.\\n\\n- Thanks you for your insightful suggestion. As you mentioned, applying $g(z)$ to the shared initial parameter $\\\\theta$ may set some of the initial parameters to smaller values. However, this is not problematic, but rather beneficial since the parameters in the shared $\\\\theta$ that are irrelevant for the given task need to be properly suppressed such that it does not distract the learning on the current task. \\n\\nHowever, the interpolation technique you mentioned, which learns to interpolate between a random weight matrix and $\\\\theta$ ($\\\\theta_0 = \\\\theta\\\\odot g(z) + (1-g(z)) \\\\odot \\\\theta_{rand}$, where $\\\\odot$ is element-wise multiplication and $\\\\theta_{rand}$ is a random weight matrix) may have a similar effect, and we performed experimental validation of it during the rebuttal period.\\n\\nModels\\t\\t\\t\\t\\t Omniglot\\t MNIST\\t\\t\\nMeta-SGD\\t\\t\\t\\t 98.38+-0.23\\t90.42+-0.28\\t\\nBayesian g(z) TAML + Interpolation\\t 98.83+-0.08\\t92.03+-0.19\\t\\nBayesian g(z) TAML\\t\\t\\t 98.92+-0.30\\t92.06+-0.32\\t\\n\\nThe results show that Bayesian TAML with the interpolation technique (Bayesian g(z) TAML + Interpolation) have a very similar performance as the original Bayesian TAML with the masking scheme. Thus the two techniques seem to have a similar effect, on deciding what to use from the meta-knowledge and what to forget.\\n\\nWe included these results in Appendix F.\"}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"We really appreciate your constructive comments. We respond to each comment as follows.\\n\\n1. Including standard errors for the results in Section 5.2.\\n\\n- We omitted them due to space limitation. As you suggested, we added in the standard errors to all the tables in the Experiment section, including the tables in Section 5.2 in the revision.\\n\\n2. It would be interesting to see the results of analyzing the class imbalance variables are statistically significant, specially in light of a recent work by Byrd and Liption [1]\\n\\n- Based on the standard errors we added into the paper (See Table 5), the effectiveness of class-imbalance variables are indeed statistically significant.\\n\\n- Thank you for letting us know a very interesting paper [1]. It is interesting that for DNNs of high capacity, the effect of importance weighting asymptotically disappears if no regularizers are applied. However in Figure 6, we can see that the per-class improvements are higher as the number of instances becomes smaller, which suggests that in our case, the class balancing variable $g(\\\\omega)$ is still effective in fixing class imbalance. \\n\\n- We conjecture that such correct behavior of $g(\\\\omega)$ is an effect of early-stopping, which is enforced to any MAML-variants by design [2]. Byrd and Lipton [1] mention that early-stopping prevents the effect of importance sampling from disappearing (Figure 5 and Figure 6 in [1]). Since all MAML-variants, including our model, only take few gradient steps for each task, our model does not get affected by such vanishing effect. This early-stopping behavior is actually further promoted in our model, as we have an additional learning-rate decaying factor $f(\\\\gamma)$ which reduces the total length of the learning trajectory. \\n\\n= Reference =\\n[1] Byrd and Lipton., What is the Effect of Importance Weighting in Deep Learning?, ICML, 2019.\\n[2] Grant et al., Recasting Gradient-Based Meta-Learning as Hierarchical Bayes, ICLR, 2018.\\n\\n\\n3. Can you report the values of the balancing variables of Section 5.1 to the Appendix?\\n\\n- We report the values of the balancing variables with the full model in Section E of the appendix. The trends on the three balancing variables are similar to what we report in the ablation study.\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We really appreciate your constructive comments. We respond to each comment as follows.\\n\\n1. Comparison between Bayesian TAML and regularized Deterministic TAML. \\n\\n- Thank you for your helpful suggestion. As you suggested, we applied L2 regularization on the three balancing variables, with coefficients {1e-1,1e-2,1e-3,1e-4}. Note that since the inference network is shared for three balancing variables, the regularized version of deterministic TAML only introduce a single hyperparameter to tune. We named the model as Deterministic TAML (L2 Reg.). The results are as follows.\\n\\nModels\\t\\t\\t\\t\\t Omniglot\\tMNIST\\nMAML\\t\\t\\t\\t\\t 98.32+-0.26\\t89.19+-0.31\\nDeterm. TAML (No Reg.)\\t\\t98.75+-0.34\\t90.77+-0.36\\nDeterm. TAML (L2 Reg., 1e-1) 98.98+-0.08 85.92+-0.31\\nDeterm. TAML (L2 Reg., 1e-2)\\t98.85+-0.08\\t90.97+-0.22\\nDeterm. TAML (L2 Reg., 1e-3)\\t98.93+-0.09\\t91.52+-0.21 \\nDeterm. TAML (L2 Reg., 1e-4)\\t98.58+-0.10\\t91.13+-0.22\\nBayesian TAML\\t\\t\\t 99.13+-0.31\\t92.38+-0.33\\n\\nWe observe that applying weight decay regularization to the balancing variables of Deterministic TAML improves the model performance. However, it may even underperform the original MAML if the hyperparameter $\\\\lambda$ is not tune with cross-validation (e.g. $\\\\lambda = 1e-1$). Bayesian TAML does not require any additional hyperparameter tuning, and still largely outperform this deterministic version of TAML. In the revision, we added in the performance of this regularized version of Deterministic TAML, in Table 6. \\n\\n2. Effectiveness of merging $\\\\gamma$ which handles task imbalance and $z$ which tackles out-of-distributions as a single set of parameters.\\n\\n- Thank you for your insightful suggestion. Learning of the coordinate-wise learning rate vector that you suggested may be effective if we can take large number of steps with sufficiently large learning rates. However, in meta-learning scenarios where the model takes only a few gradient steps at each episode, the initialization step becomes significantly more important, and we cannot obtain a similar effect without directly controlling the initialization via $g(z)$. Moreover, adding this balancing variable marginally increases the computational cost since it shares the inference network with other balancing variables. In an earlier version of our model, we had a model that is exactly the same as what you suggested (which learns a task-dependent learning vector), but we empirically found that having separate $\\\\gamma$ and $z$ performs better.\\n \\nDuring the rebuttal period, we performed experimental comparison of a modified version of our model (we omitted $\\\\omega$ to see the effect of $\\\\gamma$ and $z$ only), against the model with the task-dependent learning rate vector (denoted as Bayesian Task Dependent $\\\\alpha$) under the same experiment setup we used to obtain the results in Table 1:\\n\\nModels\\t\\t\\t\\t\\t Omniglot\\tMNIST\\t\\t\\nBayesian Task Dependent $\\\\alpha$ \\t94.19+-0.16\\t78.40+-0.25\\t\\nBayesian $\\\\gamma$ and $z$ TAML\\t 96.76+-0.12\\t81.13+-0.22\\t\\n\\nThe results show that our Bayesian TAML with $\\\\gamma$ and $z$ significantly outperforms the Bayesian MAML with task-dependent learning rate.\\n \\n\\n3. Experiment on the Meta-Dataset\\n\\n- Thank you for your helpful suggestion. We agree, and actually, the OVD dataset in Page 7 is a subset of meta-dataset, on which our model significantly outperforms the baselines (Table 2). While we could not conduct experiments on the full Meta Dataset due to lack of time, we are preparing to conduct this experiment and will include the results on the dataset in the final version of the paper. \\n\\n4. Dataset used in Tables 4 and 5.\\n\\n- The dataset used in Table 4 and 5 is Omniglot. While this is mentioned in the texts, we updated all tables in the Experiment section to clearly describe the dataset in the revision.\\n\\n\\n5. The meaning of \\u00d75 and \\u00d715 of Table 5 in section 5.2.\\n\\n- Your interpretation is correct, and this describes the degree of class imbalance. (\\u00d7N) means that the maximum number of shots is N times larger than the minimum number of shots within the given task. Note that the number of instances for each task is exactly the same to control the task imbalance in this experiment. In Section 5.2, we added more detailed descriptions of the experimental setup.\\n\\n\\n6. In the Related Work Section, the description for metric-based approaches.\\n\\n- Thank you for the correction. As you suggested, we revised the descriptions of the metric-based meta learning approaches in the Related Work section.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary\\n========\\nThis paper introduces a mechanism for gradient-based meta-learning models for few-shot classification to be able to adapt to diverse tasks that are imbalanced and heterogeneous. In particular, each encountered task may have varying numbers of shots (task imbalance) and even within each task, different classes may have different numbers of shots (class imbalance). Further, test tasks might come from a different distribution than the training tasks. They propose to handle this scenario by introducing three new types of variables which control different facets of the degree and type of task adaptation, allowing to decide how much to reuse meta-learned knowledge versus new knowledge acquired from the training set of the given task.\\n\\nSpecifically, their newly introduced variables are: 1) the factor for learning rate decay (a scalar) for the inner-loop task adaptation optimization which allows to not deviate too much from the global initialization when insufficient data is available, 2) a class-specific learning rate (one scalar per class) that allows to tune more for under-represented classes of the training set, 3) a set of weights on the global initialization (one scalar per dimension) that can down-weigh each component if it\\u2019s not useful for the task at hand (e.g. if test tasks have significantly different statistics than training tasks did).\", \"the_values_of_these_variables_are_predicted_based_on_the_training_set_of_the_task\": \"the support set is encoded via a hierarchical variant of a set encoding (where pooling is done using higher order statistics too instead of simply averaging). The resulting encoded support set is the input to the network that produces the values for the three sets of variables discussed above. Each new variable is treated in a Bayesian fashion: a prior is defined over it (Normal(0,1)), which is updated by conditioning on the training set to form a posterior for each given task. Specifically, each of the above variables is represented by a Gaussian whose mean and variance are the learnable parameters that are produced by the network described above.\\n\\nExperimentally, this method outperforms others on a setting of imbalanced tasks (the shot is sampled uniformly at random from a designated range). The gain over other methods is large in particular when evaluated on out-of-distribution tasks (coming from a different dataset) and when the imbalance is large.\\n\\nComments (in decreasing order of importance)\\n========================================\\nA) The Bayesian framework helps because it offers an elegant way to use a prior. In the deterministic version, was any effort made to resemble the effect of that prior? For example, one can define a regularizer that penalizes behaviors that ignore the meta-knowledge too much (e.g. too large values for \\\\gamma, or for the class-specific learning rates etc). Albeit more \\u2018hacky\\u2019, if these regularization coefficients are tuned properly, they might result in a similar effect to that of having a prior. A fair comparison to the deterministic variant should include this.\\n\\nB) I think that \\\\gamma and z can be merged into a single set of parameters? In particular, imagine a per-dimension-of-\\\\theta learning rate. This would then be large for a dimension when there is a larger need for adapting that dimension of \\\\theta. In the case of large training sets, this can be large for all dimensions, recovering the behavior of a large \\\\gamma. For the case of diverse datasets, this would behave as the current z (updates a lot the dimensions of \\\\theta that are irrelevant for the given task due to the dataset shift).\\n\\nC) Meta-Dataset (https://arxiv.org/abs/1903.03096) is a recent benchmark for few-shot classification that introduces both of what is referred to here as task imbalance and class imbalance and also is comprised of heterogeneous datasets and evaluates performance on some held-out datasets too. The current state-of-the-art on it (as far as I know) is CNAPs [1] which employs a flexible adaptation mechanism on a per-task basis but is fully amortized (performs no gradient-based adaptation to each task) and makes no explicit effort to tackle imbalanced tasks as is done here. I\\u2019m curious how this method would compete in that setup. It definitely seems to be a strong candidate for that benchmark!\\n\\nLess important\\n=============\\nD) which dataset is used in Tables 4 and 5? I assume it\\u2019s Omniglot (due to the numbers being in the 90s) but it would be good to say this explicitly.\\nE) In section 5.2, expressions such as x5 and x15 are used to characterize the degree of imbalance of a task. How exactly are these computed? Does x5 mean that the largest shot is 5 times larger than the smallest shot? It would be good to explicitly state this.\\nF) In the Related Work section, in the Meta-learning paragraph there is a sentence that\\u2019s not accurate: \\u201cMetric-based approaches learn a shared metric space [...] such that the instances are closer to their correct prototypes than to others\\u201d. This sentence does not describe all metric based approaches. It describes Prototypical Networks (Snell et al) but not, for example, Matching Networks (Vinyals et al) nor many others that like Matching Networks perform example-based comparisons and don\\u2019t aggregate a class\\u2019 examples into a prototype.\\n\\nIn a nutshell\\n===========\\nI think this work is a useful contribution for moving towards a more realistic setting in few-shot classification. It captures some desiderata of models that can operate in more realistic settings and outperforms previous models in those scenarios. My comments above are mostly suggesting improvements and clarifications but I am inclined to recommend acceptance of this paper.\\n\\nReferences\\n=========\\n[1] Fast and Flexible Multi-Task Classification Using Conditional Neural Adaptive Processes. Requeima et al. NeurIPS 2019.\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": [\"The paper proposes a Bayesian approach for meta learning in settings were the tasks might be OOD or have imbalanced class distribution. The proposed approach has 3 task-specific balancing variables with a prior and an inference network. Using an amortized inference scheme, the model unifies the meta-learning objective loss with the lower bound of probabilistic model marginal likelihood.\", \"The paper is well-written and well-motivated. I only have some minor comments and questions:\", \"Can you add the standard errors for the results in Section 5.2 (maybe at least in the Appendix)?\", \"Specifically it would be interesting to see if the results for analyzing the class imbalance variable are statistically significant; specially in light of the recent work on the effects of importance weighting in DL (see \\u201cWhat is the Effect of Importance Weighting in Deep Learning?\\u201d by Byrd and Lipton) which essentially question the value of importance weighting for handling class imbalance in various DL settings.\", \"For the experiments in Section 5.1 can you also report the values for the three task-specific balancing variables? (Maybe in the Appendix).\"], \"minor\": \"\\u201cobtains significantly improves over\\u201d -> \\\"significantly improves over\\\"\\n\\nOverall, I found the paper interesting and practically useful, although I believe some additions to the empirical evaluation can improve the impact of the paper.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary\\n-------------\\nThis paper proposed to improve existing meta learning algorithms in the presence of task imbalance, class imbalance, and out-of-distribution tasks. Starting from the model-agnostic meta-learning (MAML) algorithm (Finn et al. 2017), to tackle task imbalance, where the number of training examples of varies across different tasks, a task-dependent learning rate decaying factor was learned to be large for large tasks and small for small tasks. In this way, the small task can benefit more from the meta-knowledge and the large task can benefit more from task-specific training. To tackle class imbalance, a class-specific scaling factor was applied to the class-specific gradient. The scaling factor was large for small class and small for large class so that different classes can be treated equally. To tackle the out-of-distribution tasks, a task-dependent variables was learned to emphasize meta-knowledge for the test task similar to training tasks. Additional model parameters are learned through variational inference. Experimental results on benchmark datasets demonstrate the proposed approach outperformed its competing alternatives. Analysis of each component confirm they work as expected.\\n\\nComments\\n---------------\\nThis paper is well motivated and clearly written. The empirical evaluation also support major claims in the paper.\\n\\nCan the author provide more details on the inference of the model? In the likelihood term in Eq. (7), the task specific parameters \\\\theta^{\\\\tau} was parameterized by Eq. (3), which contains K iterative gradient updates. How was the gradient w.r.t. \\\\theta was computed in this setting?\\n\\nThe task-specific learning rate decaying factor was constrained to be between 0 and 1 using the function f(). The class-specific scaling factor made use of the SoftPlus() function, for the same purpose of scaling learning rate, why do these two different options of functions were applied?\\n\\nFor the scaling vector of the initial parameters g(z^{\\\\tau}), for its zero entries, the initialization of the corresponding entries in task-specific parameter \\\\theta would be zero. Would it be better to apply a linear interpolation between \\\\theta and a randomly-initialized vector in Eq (2)?\\n\\nEdits after reading the author's rebuttal\\n==================================\\nThe author's reply well addressed my questions. After reading other reviewers' positive comments and the author's thorough reply, I decide to increase my rating to 8: Accept.\"}" ] }

[ "Kane Zhang", "Jian Zhang", "Qiang Wang", "Zhao Zhong" ]

[ "CNNs", "dynamic convolution kernel" ]

{ "note_id": [ "CTe8MDNzmx1", "NMkwkr7Ean", "rylQ2j7njS", "BkecUvXhjH", "HyeZFOZ3jB", "rkltsVmsiH", "H1gKEIgwsr", "rye239JPsr", "HyxAifCmsr", "SkeG0bSxoB", "SJxn_9yy9H", "B1lC6hMCYS", "r1gRUl0TYB", "BJx7isQ6tB" ], "note_type": [ "comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "comment" ], "note_created": [ 1603857986935, 1576798730672, 1573825450802, 1573824338350, 1573816440978, 1573758113459, 1573484081310, 1573481140266, 1573278374146, 1573044682252, 1571908212489, 1571855558058, 1571835990000, 1571793819055 ], "note_signatures": [ [ "~HC_Xu1" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1718/Authors" ], [ "ICLR.cc/2020/Conference/Paper1718/Authors" ], [ "ICLR.cc/2020/Conference/Paper1718/Authors" ], [ "ICLR.cc/2020/Conference/Paper1718/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1718/Authors" ], [ "ICLR.cc/2020/Conference/Paper1718/Authors" ], [ "ICLR.cc/2020/Conference/Paper1718/Authors" ], [ "ICLR.cc/2020/Conference/Paper1718/Authors" ], [ "ICLR.cc/2020/Conference/Paper1718/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1718/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1718/AnonReviewer2" ], [ "~Brandon_Yang1" ] ], "structured_content_str": [ "{\"title\": \"Could you open your code?\", \"comment\": \"this work looks easier to use, will you release your code in the future?\"}", "{\"decision\": \"Reject\", \"comment\": \"The paper proposed the use of dynamic convolutional kernels as a way to reduce inference computation cost, which is a linear combination of static kernels and fused after training for inference to reduce computation cost. The authors evaluated the proposed methods on a variety models and shown good FLOPS reduction while maintaining accuracy.\\n\\nThe main concern for this paper is the limited novelty. There have been many works use dynamic convolutions as pointed out by all the reviewers. The most similar ones are SENet and soft conditional computation. Although the authors claim that soft conditional computation \\\"focus on using more parameters to make models to be more expressive while we focus on reducing redundant calculations\\\", the methods are pretty the same and moreover in the abstract of soft conditional computation they have \\\"CondConv improves the performance and inference cost trade-off\\\".\", \"title\": \"Paper Decision\"}", "{\"title\": \"The revision is submitted and the points you concern have been updated.\", \"comment\": \">>> Response to \\u201cThe proposed add-on requires many parameters but the number of parameters is not shown in this paper\\u201d:\\nThe proposed add-on indeed requires many parameters, we have added the number of parameters in Table4, Table5 and Table6 to illustrate this point.\\n Table 4\\n+------------------+---------+--------+----------------+\\n| Methods | MParams | MFLOPs | Top-1 err. (%) |\\n+------------------+---------+--------+----------------+\\n| Dy-mobile(1.0) | 7.36 | 135 | 28.38 |\\n+------------------+---------+--------+----------------+\\n| Dy-shuffle(1.5) | 11 | 180 | 27.48 |\\n+------------------+---------+--------+----------------+\\n| Fix-mobile(1.0) | 2.16 | 129 | 33.57 |\\n+------------------+---------+--------+----------------+\\n| Fix-shuffle(1.5) | 2.47 | 171 | 30.3 |\\n+------------------+---------+--------+----------------+\\n Table 5\\n+-----------------+---------+--------+----------------+\\n| Methods | MParams | MFLOPs | Top-1 err. (%) |\\n+-----------------+---------+--------+----------------+\\n| Fix-mobile(1.0) | 2.16 | 129 | 33.57 |\\n+-----------------+---------+--------+----------------+\\n| Dy-mobile(gt=2) | 3.58 | 131 | 29.43 |\\n+-----------------+---------+--------+----------------+\\n| Dy-mobile(gt=4) | 5.47 | 133 | 28.69 |\\n+-----------------+---------+--------+----------------+\\n| Dy-mobile(gt=6) | 7.36 | 135 | 28.38 |\\n+-----------------+---------+--------+----------------+\\t\\n Table 6\\n+---------------------+---------+--------+----------------+\\n| Methods | MParams | MFLOPs | Top-1 err. (%) |\\n+---------------------+---------+--------+----------------+\\n| Dy-mobile(1.0,gt=1) | 2.64 | 131 | 30.85 |\\n+---------------------+---------+--------+----------------+\\n| Dy-mobile(1.0,gt=6) | 7.36 | 135 | 28.27 |\\n+---------------------+---------+--------+----------------+\\n| Dy-ResNet18(gt=1) | 3.04 | 556 | 33.8 |\\n+---------------------+---------+--------+----------------+\\n| Dy-ResNet18(gt=6) | 16.6 | 567 | 31.01 |\\n+---------------------+---------+--------+----------------+\"}", "{\"title\": \"The revision is submitted and the points you concern have been updated.\", \"comment\": \">>> Response to \\u201cWhy larger g_t doesn't increase the flops in Table 5?\\u201d\\n\\tThe FLOPs of coefficient prediction module and dynamic generation module is taken into account in this version (Figure5, Table1, Table2, Table4, Table5, Table6). Thus the flops in Table 5 increases slightly as g_t becoming larger.\\n Table 5\\n+-----------------+---------+--------+----------------+\\n| Methods | MParams | MFLOPs | Top-1 err. (%) |\\n+-----------------+---------+--------+----------------+\\n| Fix-mobile(1.0) | 2.16 | 129 | 33.57 |\\n+-----------------+---------+--------+----------------+\\n| Dy-mobile(gt=2) | 3.58 | 131 | 29.43 |\\n+-----------------+---------+--------+----------------+\\n| Dy-mobile(gt=4) | 5.47 | 133 | 28.69 |\\n+-----------------+---------+--------+----------------+\\n| Dy-mobile(gt=6) | 7.36 | 135 | 28.38 |\\n+-----------------+---------+--------+----------------+\\t\\n>>> Response to \\u201cThe author may need to show the comparisons of the number of parameters.\\u201d:\\n\\tWe have added the number of parameters in Table4, Table5 and Table6.\\n Table 4\\n+------------------+---------+--------+----------------+\\n| Methods | MParams | MFLOPs | Top-1 err. (%) |\\n+------------------+---------+--------+----------------+\\n| Dy-mobile(1.0) | 7.36 | 135 | 28.38 |\\n+------------------+---------+--------+----------------+\\n| Dy-shuffle(1.5) | 11 | 180 | 27.48 |\\n+------------------+---------+--------+----------------+\\n| Fix-mobile(1.0) | 2.16 | 129 | 33.57 |\\n+------------------+---------+--------+----------------+\\n| Fix-shuffle(1.5) | 2.47 | 171 | 30.3 |\\n\\n Table 6\\n+---------------------+---------+--------+----------------+\\n| Methods | MParams | MFLOPs | Top-1 err. (%) |\\n+---------------------+---------+--------+----------------+\\n| Dy-mobile(1.0,gt=1) | 2.64 | 131 | 30.85 |\\n+---------------------+---------+--------+----------------+\\n| Dy-mobile(1.0,gt=6) | 7.36 | 135 | 28.27 |\\n+---------------------+---------+--------+----------------+\\n| Dy-ResNet18(gt=1) | 3.04 | 556 | 33.8 |\\n+---------------------+---------+--------+----------------+\\n| Dy-ResNet18(gt=6) | 16.6 | 567 | 31.01 |\\n+---------------------+---------+--------+----------------+\"}", "{\"title\": \"Thanks for your helpful reply and we have updated our paper.\", \"comment\": \">>> Response to \\\"the authors should simply state all metrics (inference/training latency, parameters, FLOPS (including the coefficient prediction module) for the benchmarked networks\\\":\\n\\nThanks for your suggestion. We have updated the experiment tables to show the training latency and amount of parameters. The FLOPs of coefficient prediction module and dynamic generation module is taken into account in this version (Figure5, Table1, Table2, Table4, Table5, Table6). Because it has been shown in the Table2-6 that the proposed method increases the number of parameters and training latency, we don't add the parameter amounts in Table 1 in order to highlight the advantage is reducing computation cost while maintaining the accuracy.\\n\\nIn Table 2, we add the training latency:\\n Table2\\n+------------------+------------------+-----------------+----------------+\\n| Method | Top-1 error. (%) | Inference Time | Training Time |\\n+------------------+------------------+-----------------+----------------+\\n| Dy-mobile(1.0) | 28.27 | 58.3ms | 250ms |\\n+------------------+------------------+-----------------+----------------+\\n| MobileNetV2(1.0) | 28 | 109.1ms | 173ms |\\n+------------------+------------------+-----------------+----------------+\\n| Dy-ResNet18 | 31.01 | 68.7ms | 213ms |\\n+------------------+------------------+-----------------+----------------+\\n| ResNet18 | 30.41 | 90.7ms | 170ms |\\n+------------------+------------------+-----------------+----------------+\\n| Dy-ResNet50 | 23.75 | 135.146ms | 510ms |\\n+------------------+------------------+-----------------+----------------+\\n| ResNet50 | 23.67 | 199.6ms | 308ms |\\n+------------------+------------------+-----------------+----------------+\\n\\n\\nIn table 4, table 5 and table 6, we add the amount of parameters and and take the FLOPs of coefficient prediction module and dynamic generation module into account:\\n\\n Table 4\\n+------------------+---------+--------+----------------+\\n| Methods | MParams | MFLOPs | Top-1 err. (%) |\\n+------------------+---------+--------+----------------+\\n| Dy-mobile(1.0) | 7.36 | 135 | 28.38 |\\n+------------------+---------+--------+----------------+\\n| Dy-shuffle(1.5) | 11 | 180 | 27.48 |\\n+------------------+---------+--------+----------------+\\n| Fix-mobile(1.0) | 2.16 | 129 | 33.57 |\\n+------------------+---------+--------+----------------+\\n| Fix-shuffle(1.5) | 2.47 | 171 | 30.3 |\\n+------------------+---------+--------+----------------+\\n Table 5\\n+-----------------+---------+--------+----------------+\\n| Methods | MParams | MFLOPs | Top-1 err. (%) |\\n+-----------------+---------+--------+----------------+\\n| Fix-mobile(1.0) | 2.16 | 129 | 33.57 |\\n+-----------------+---------+--------+----------------+\\n| Dy-mobile(gt=2) | 3.58 | 131 | 29.43 |\\n+-----------------+---------+--------+----------------+\\n| Dy-mobile(gt=4) | 5.47 | 133 | 28.69 |\\n+-----------------+---------+--------+----------------+\\n| Dy-mobile(gt=6) | 7.36 | 135 | 28.38 |\\n+-----------------+---------+--------+----------------+\\n Table 6\\n+---------------------+---------+--------+----------------+\\n| Methods | MParams | MFLOPs | Top-1 err. (%) |\\n+---------------------+---------+--------+----------------+\\n| Dy-mobile(1.0,gt=1) | 2.64 | 131 | 30.85 |\\n+---------------------+---------+--------+----------------+\\n| Dy-mobile(1.0,gt=6) | 7.36 | 135 | 28.27 |\\n+---------------------+---------+--------+----------------+\\n| Dy-ResNet18(gt=1) | 3.04 | 556 | 33.8 |\\n+---------------------+---------+--------+----------------+\\n| Dy-ResNet18(gt=6) | 16.6 | 567 | 31.01 |\\n+---------------------+---------+--------+----------------+\\n\\n>>> Response to \\\"the authors should clearly state the dynamic convolutions have been proposed before.\\u201d\\nWe have updated Sec 2.3 to compare with existing work and state that the idea of linearly combining static kernels using predicted coefficients has been proposed before.\\n\\n>>>Response to \\u201cFigure 5: how are the models constrained to have same FLOPS?\\u201d :\\nWe have updated Sec 4.3 and Figure 5 to illustrate this issue as follows:\\n\\u201cFurthermore, we conduct detailed experiments on MobileNetV2. We replace the conventional convolution with the proposed dynamic one and get Dy-MobileNetV2. The accuracy of classification for models with different number of channels are shown in Figure 5. It is observed that Dy-MobileNetV2 consistently outperforms MobileNetV2 but the ascendancy is weaken with the increase of number of channels.\\u201d\"}", "{\"title\": \"Reviewer reply\", \"comment\": \"I read the authors' replies to the reviews and changed my score to weak accept.\\nI'm still on the fence as there is no comparison to previous work.\\n\\nI think this could be a better paper if the authors take into account the reviewer's suggestions. Most importantly, the authors should simply state all metrics (inference/training latency, parameters, FLOPS (including the coefficient prediction module) for the benchmarked networks and clearly state the dynamic convolutions have been proposed before.\\n\\nNote that pay less attention with dynamic convolutions predicts the parameters of a depthwise convolution with some additional weight sharing across layers which reduces the \\\"makes the number of parameters of the linear layer unbearable\\\" issue.\"}", "{\"title\": \"Thanks for the detailed response. Please see comments below.\", \"comment\": \">>> Response to \\u201cThe proposed add-on requires many parameters but the number of parameters is not shown in this paper\\u201d:\\nThanks for your suggestion, we will add the number of parameters in experiment tables in the revision.\\n>>> Response to \\u201cThe experiment shows the ablation to the case g_t=1, but what if we set the parameters to other numbers?\\u201d:\\nThanks for your comments! Actually, we have shown it in Table 5. The value of g_t does not change the computation cost of convolution but affects the performance of networks. The performance will become better when g_t gets larger.\\n>>> Response to \\u201cThe author should clarify the difference and the strong points of the proposed block compared to SEnet.\\u201d\\nThanks for your meaningful question! \\nFrom the method's point of view, we fuse feature maps instead of fusing kernels in the training stage, it can be done via the combination of group point-wise convolution and SE mechanism. However we fuse kernels in the inference stage, it can`t be done via this combination. This is the core difference with SEnet and reduces the computation cost by gt times compared with fusing feature maps or the combination of group point-wise convolution and SE mechanism.\\nFrom the experiments' point of view, as discussed in Sec 4.5, our proposed method will be the same as SENet when g_t=1. However, compared with g_t=6, its performance drops a lot while the computation cost of convolution stay the same as shown in Table 6. \\n>>> Response to \\u201cTesting the ImageNet trained network of the proposed method into an object detection task (as the pre-trained backbone).\\u201d\\nThanks for your suggestion. We are trying to implement the baseline and add experiments for detection in the revision before the deadline.\"}", "{\"title\": \"Thanks for the detailed response. Please see comments below.\", \"comment\": \">>> Response to \\u201cFigure 5: how are the models constrained to have same FLOPS?\\u201d:\\nThanks for your valuable question which help us realize some unclear statements. By changing the number of channels of MobileNetV2, we can get MobileNetV2(0.35), MobileNetV2(0.5), MobileNetV2(0.75), MobileNetV2(1.0) and MobileNetV2(1.4). To constrain our models to have the same FLOPS with them, we keep the channels of each layer the same as the original one, while replacing the conventional convolution with dynamic convolution. If we ignore the additional FLOPs of coefficient prediction module and dynamic generation module, which is negligible, the FLOPs will stay the same. This result shows that our proposed dynamic convolution can be deployed as a plug-and-play unit to replace conventional convolution.\\nWe realize that it is not proper to call these models \\u201cDy-mobile\\u201d because they have different structures. In the updated version we will call them as Dy-MobileNetV2(0.35), Dy-MobileNetV2(0.5), Dy-MobileNetV2(0.75), Dy-MobileNetV2(1.0) and Dy-MobileNetV2(1.4).\\n>>> Response to \\u201cConsider adding training times and parameter counts.\\u201d:\\nThanks for your suggestion! We will add parameter counts in experiment tables and discuss the training time in Sec 4.3 in the revision.\\n>>> Response to \\u201cLack of novelty.\\u201d:\\nFor previous works on dynamic convolution, they all directly generate convolution kernels via a linear layer (including \\u201cpay less attention with dynamic convolutions\\u201d). In computer vision tasks, the parameter counts of convolution is large, which makes the number of parameters of the linear layer unbearable. In our proposed method, the linear layer is merely used to predict the coefficients for linearly combining static kernels. It can solve this problem and thus be used to achieve real speed up for CNN on hardware. \\nThe most related work is \\u201csoft conditional computation\\u201d. They focus on using more parameters to make models to be more expressive while we focus on reducing redundant calculations in convolution. According to the theoretical analysis in appendix A and correlation study in Figure 6, we find that correlations among convolutional kernels can be reduced via dynamically fusing several kernels. Thus different from \\u201csoft conditional computation\\u201d which replaces the conventional convolution directly, we recommend to reduce the channel numbers (we reduce half of the channels for convolution layers in Dy-ResNet18/50 and Dy-shuffle) and then replace conventional convolution with dynamic one. We think this may bring a larger improvement. For example, compared with directly replace the conventional convolution in ResNet18, Dy-ResNet50 reduce 37.9% FLOPs while improves 2.03% top-1 accuracy on ImageNet. Moreover, since this paper is submitted to NeurIPS2019 this May as well, it may be regarded as a concurrent work\\nWe will add a comparison against existing work in Sec 2.3.\"}", "{\"title\": \"Thanks for the detailed response. Please see comments below.\", \"comment\": \">>> Response to \\u201cThe only difference with SENet is the introduction of g_t where the output dimension is much larger than SENet.\\u201d:\\nThanks for your meaningful question! The motivation of our proposed approach is different from SENet. We focus on reducing the computation cost during inference stage, while SENet learns channel-wise importance to recalibrates feature maps. From the methods point of view, during training, indeed the only difference is that the output dimension is much larger than SENet. However, during inference, SENet recalibrate feature maps by learned weights, while we fuse kernels instead of fusing feature maps, which can reduce the output dimension and computation cost of convolution kernels.\\nMoreover, as discussed in Sec 4.5, our proposed method will be the same as SENet when g_t=1. However, compared with g_t=6, its performance drops a lot while the computation cost of convolution stay the same as shown in Table 6.\\n>>> Response to \\u201cHow can this method save computations compared to classical convolution?\\u201d:\\nThanks for your valuable comments! The goal of our proposed approach is to reduce the computation cost in the inference stage. Equation (2) only shows that fusing feature maps is mathematically equivalent with fusing kernels. During training, we firstly generate C_cout*g_t output feature maps and then get C_out feature maps by fusing each g_t ones. Thus we can not save computation cost in training procedure. However, we fuse kernels during inference stage instead of fusing feature maps. After fusing, the weight for one convolution layer with shape [C_out*g_t, C_in, k, k] will become [C_cout, C_in, k, k], thus the FLOPs is reduced by g_t times.\\n>>> Response to \\u201cWhy larger g_t doesn`'t increase the flops in Table 5?\\u201d:\\nCompared with g_t=1, the computation cost of convolution is 6 times larger when g_t=6 during training. However, we will fuse each g_t kernels to get dynamic convolutional weight with shape [C_cout, C_in, k, k] during inference as Equation (1), it makes the FLOPs of convolution invariable with g_t. As for the additional FLOPs of fusing kernels, it varies with g_t but is negligible. For Dy-mobile(1.0), the FLOPs of fusing kernels is merely 0.459*g_t M, when g_t=6 it is 2.75M while the FLOPs of convolution is 129 M.\\n>>> Response to \\u201cThe author may need to show the comparisons of the number of parameters.\\u201d:\\nThanks for the suggestion! The new module indeed increases the parameters a lot, because we mainly focus on the computation cost rather than the number of parameters. We will add parameter counts in experiment table in our updated version.\"}", "{\"title\": \"Thanks for bringing this interesting work to our attention.\", \"comment\": \"Thanks for your comments! We note that our work share similar idea with yours. Since this paper is submitted to NeurIPS2019 this May as well, it may be regarded as a concurrent work. We will mention and cite your work in our updated version of the paper.\\n\\nIn table5, we don`t take the MFLOPs of coefficient prediction module and dynamic generation module into account because it is negligible. For example, when g_t=6, the computation cost of coefficient prediction module and dynamic generation module is only 3.37 and 2.75 MFLOPs respectively, while the computation cost of convolution is 129 MFLOPs.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposed dynamic convolution (DyNet) to accelerating convolution networks. The new method is tested on the ImageNet dataset with three different backbones. It reduces the computation flops by a large margin while keeps similar classification accuracy. The additional segmentation experiment on the Cityscapes dataset also shows the new module can save computation a lot while maintaining similar segmentation accuracy.\", \"clarity\": \"The novelty of the paper is limited and the experimental results are weird for me.\\n1. The proposed module named dynamic convolution is detailed in Sec 3.2. As far as I can see, it is very similar to the former SENet especially in Figure (3) and Equation (2). The only difference is the introduction of g_t where the output dimension is much larger than SENet.\\n\\n2. As shown in Equation (2), the proposed method contains the normal computation of fixed kernels. How can this method save computations compared to classical convolution? Is the computation flops calculated in the right way?\\n\\n3. The results in Table 5 are strange to me. Larger g_t will increase the flops absolutely according to Equation (2).\\n\\n4. The author may need to show the comparisons of the number of parameters. In my opinion, the new module will increase the parameters a lot (the output dimension of the fully connected layer is as large as C_cout*g_t).\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"=== Summary ===\\nThe authors propose to use dynamic convolutional kernels as a means to reduce the computation cost in static CNNs while maintaining their performance. The dynamic kernels are obtained by a linear combination of static kernels where the weights of the linear combination are input-dependent (they are obtained similarly to the coefficients in squeeze-and-excite). The authors also include a theoretical and experimental study of the correlation.\\nThe authors conduct extensive experiments on image classification and segmentation and show that dynamic convolutional kernels with reduced number of channels lead to significant reduction in FLOPS and increase in inference speed (for batch size 1) compared to their static counterparts with higher number of channels.\\n\\n=== Recommendation ===\\nThe experimental setup is rigorous but the current draft lacks some metrics that should be reported (as training times, parameter counts, memory requirements at training/inference) since the focus is on making CNNs more efficient.\", \"the_presented_experimental_results_are_satisfactory_but_the_studied_networks_are_not_quite_sota\": [\"they are much more competitive alternatives to ResNet and MobileNetv2. The correlation study is interesting.\", \"My main issue with the paper is the lack of novelty. The use of dynamic convolutions is by no means a novel idea and has been studied in multiple previous works in vision (mixture of experts, soft conditional computation, pay less attention with dynamic convolutions, ...) which the authors fail to cite/compare against.\", \"However, most previous work focuses on leveraging dynamic kernels to use more parameters so the focus on accelerating CNNs is novel.\", \"Overall, I am on the fence with this paper but slightly leaning towards rejecting it for the above reasons.\", \"=== Questions/Comments ===\", \"Figure 5: how are the models constrained to have same FLOPS? Is it by changing the number of channels?\", \"Consider adding training times for more transparency\", \"Consider adding parameter counts in experiment tables\", \"The related work subsection 2.3 is rather poor compared to existing work.\", \"'While model compression based methods' -> 'On the other hand, model compression based methods'\", \"'computing efficient' -> 'compute efficient'\", \"'values distribute' -> 'values distributed'\", \"'DETAIL ANALYSIS OF OUR MOTIVATION' -> 'Detailed analysis of our motivation'\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1718\", \"review\": [\"Main contribution of the paper\", \"The paper proposes a dynamic convolution selection method can be applied to arbitrary classification networks based on the global average pooled (GAP) feature map info.\", \"The method obtained improvements over various networks (SuffleNet v2, MobileNet v2, ResNet 18) on ImageNet.\", \"Methods\", \"Given the set of fixed convolutional filters, the method dynamically selects the (weighted sum) kernels by given a kind of channel attention.\", \"The GAP of the features gives the channel attention on each stage, and the method applies the dynamic selection of the kernels.\", \"The number of channels in skip-connection shluld be the same because it should be elementwise multiplied with the channel attention acquired from GAP.\", \"The author slightly revises the baseline networks to set the networks integrated with the proposed method to have smaller Flops.\", \"Questions\", \"According to Figure 4, it seems that the proposed add-on requires many parameters because it would include a FC layer for each block. But we cannot find the number of parameters in this paper.\", \"The parameter $g_t$ is defined as 6. The experiment shows the ablation to the case $g_t$ =1, but what if we set the parameters to other numbers?\", \"Strong points\", \"The proposed model achieved improvement with fewer Flops on large scale image classification dataset.\", \"The method shows effectiveness when it is attached to various classification networks.\", \"Concerns\", \"The main concern of the reviewer is that the model shares the core contribution to the existing method; squeeze-and-excitation network (SEnet, Hu et.al.). The method also proposes the attention-based scaling of channels, where the attention comes from GAP, so the reviewer thinks that it is possible to explain this work as some variation of SEnet.\", \"The author should clarify the difference and the strong points of the proposed block compared to SEnet.\", \"Also, the reviewer cannot guarantee that the networks trained by the proposed method can transfer the knowledge to other tasks such as detection.\", \"The reviewer thinks that it is a critical part because one of the primal reasons for training the network is to use them as the pre-trained backbone for the other tasks.\", \"Regarding this, the baseline methods (MobileNet V2, Shufflenet v2, ResNet) are widely used as a pre-trained backbone for object detection, and the papers mention the CoCo object detection results using the pre-trained backbones from their method. The reviewer thinks that the experiment regarding this should be included.\", \"The other thing is that the parameter increases. As in the question, the reviewer thinks that the number of parameters would be increased. The reviewer agrees that some recent works focus more on Flops, but the number of parameters is also discussed in general, when telling about the 'model size'.\", \"Conclusion\", \"The author proposed a dynamic kernel selection method (add-on), which can enhance the classification accuracy of the baseline network.\", \"However, the reviewer cannot convince the novelty of the proposed approach and usefulness of the pre-trained backbone network from the proposed method when applying it to the other tasks (Object detection).\", \"Inquiries\", \"Clarifying the difference between SEnet.\", \"Testing the ImageNet trained network of the proposed method into an object detection task (as the pre-trained backbone).\", \"Discussing the number of the parameter as well.\"]}", "{\"comment\": \"Hi,\\n\\nI found this to be a very interesting analysis on using input-dependent convolutional kernels. It might be worth noting that our previous work also studies the benefits of using input-dependent convolutional kernels over static convolutional kernels in the image classification and detection settings: https://arxiv.org/abs/1904.04971 . I\\u2019d be very interested in better understanding the differences between our two approaches.\\n\\nI also had a clarifying question about the work. In Table 5, why does increasing the group size (g_t) not change the MFLOPs of the network? In my understanding, the coefficient prediction module and dynamic generation module would both require more computation with larger values of g_t.\\n\\nThanks!\", \"title\": \"Related work\"}" ] }

[ "Hui Shi", "Yang Zhang", "Xinyun Chen", "Yuandong Tian", "Jishen Zhao" ]

[ "hiss", "symbolic expressions", "methods", "equivalence", "deep symbolic superoptimization", "task", "deep learning methods", "approaches" ]

{ "note_id": [ "HjUZRSKUJ", "Bygshft2sH", "H1g5MR8iiB", "rkegH68ijr", "HkeqEhUsjH", "S1lTyjUjsH", "HJlylMRX5H", "Skl0FlWAtS", "HklviNintr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730643, 1573847731263, 1573772817916, 1573772600197, 1573772337651, 1573772004912, 1572229606708, 1571848326284, 1571759262772 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1717/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1717/Authors" ], [ "ICLR.cc/2020/Conference/Paper1717/Authors" ], [ "ICLR.cc/2020/Conference/Paper1717/Authors" ], [ "ICLR.cc/2020/Conference/Paper1717/Authors" ], [ "ICLR.cc/2020/Conference/Paper1717/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1717/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1717/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This work introduces a neural architecture and corresponding method for simplifying symbolic equations, which can be trained without requiring human input. This is an area somewhat outside most of our expertise, but the general consensus is that the paper is interesting and is an advance. The reviewer's concerns have been mostly resolved by the rebuttal, so I am recommending an accept.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Re: Response to Review #3\", \"comment\": \"The revisions strengthen the paper, and I have updated my score accordingly.\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"We have updated our paper. Please refer to the thread \\u201cRevised Paper Uploaded\\u201d for a summary of major updates in our revision. Following is a detailed response to Review #3 comments.\\n\\n1. Training details. \\nWe add Appendix A.4 to describe training computation complexity. The most time-consuming portion of training is stage-2, which requires two weeks to converge, using a single RTX 2080. \\n\\n2. Datasets. \\nOur revision Section 5.1 and Appendix A.1 & A.2 provide a detailed description of datasets. We use two datasets: the \\u201ctraverse equivalence dataset\\u201d is used for stage-I training; a benchmark dataset \\u201cHalide dataset\\u201d is used by the stage-II training. The former dataset mainly contains short expressions under a depth of four. Section 5.2 experiment is trained and tested on the traverse equivalence dataset; Section 5.3 experiment employs (a) the traverse equivalence dataset for stage-1 training and (b) the Halide dataset for stage-2 training and testing.\\n\\n3. What to glean from Section 5.1 (Now Section 5.2). \\nAlthough the expressions to be evaluated are below depth 4, they can be very complicated expressions already. For example, min(min(v2, v0 + v1), v0 + v3) is a depth-4 expression, and is included in the dataset. Since HISS applies a beam search algorithm, it is useful to demonstrate how HISS differs from conventional search algorithms and that the simplification problem cannot be trivially solved by these search algorithms.\\n\\n4. Sections 5.3 and 5.4. \\nWe move the original Sections 5.3 and 5.4 to the appendix. We keep them in the appendix to provide insights into how HISS works.\\n\\n5. Method without embedding similarity loss. \\nOur revision Appendix C provides a set of ablation studies. Specifically, Figure 5 shows how significantly performance drops if the subtree embedding similarity loss is removed. Appendix C.1 discusses why the embedding similarity loss is important. In short, this loss will provide a more concrete training signal to assist convergence, which is otherwise difficult due to REINFORCE. In addition to the embedding similarity loss, the ablation studies also investigate the significance of the subtree selector and the tree-LSTM architecture. \\n\\n6. Rewrite rules comparison. \\nSection 5.4 (especially Tables 1 and 2) of our revision includes our new results, which directly compare HISS rewrite rules with Halide rewrite rules. Our results show that HISS rules are more comprehensive. We further compare the simplification traces of HISS and Halide, demonstrating how the comprehensiveness of the rewrite rules is translated to HISS\\u2019s performance benefit.\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"We have updated our paper. Please refer to the thread \\u201cRevised Paper Uploaded\\u201d for a summary of major updates in our revision. Following is a detailed response to Review #2 comments.\\n\\n1) Ablation studies. \\nOur revised Appendix C introduces a set of ablation studies. Specifically, figure 5 shows that the performance drops significantly if the subtree embedding similarity loss is removed. In addition, appendix C.1 explores why the embedding similarity loss is important. In short, this loss will provide a more concrete training signal to assist convergence, which is otherwise difficult due to REINFORCE. In addition to the embedding similarity loss, the ablation studies also investigate the significance of the subtree selector and the tree-LSTM architecture. For more details, please kindly refer to appendix C.\\n\\n2) Hyperparameter settings. \\nOur hyperparameters, including the LSTM dimensions, follow the common setting in previous works, without any explicit tuning. Our revised Appendix A.3 describes our hyperparameter setting and our decision making in detail.\\n\\n3) How data generation affects performance. \\nOur revised Section 5.1 and Appendix A.1 & A.2 provide a detailed description of datasets preparation and use. We use two datasets: (a) the \\u201ctraverse equivalence dataset\\u201d (mentioned by Review#3) is only used for stage-I training; (b) the benchmark dataset \\u201cHalide dataset\\u201d is used by the stage-II training. The former dataset mainly contains short expressions under a depth of four; the goal is to assist the stage-II training to be performed on the latter dataset. Therefore, It does not affect the performance that the traverse space is reduced when generating the dataset for complexity reasons. Section 5.1 and Appendix A.1 & A.2 describes more details about the datasets.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"We have updated our paper. Please refer to the thread \\u201cRevised Paper Uploaded\\u201d for a summary of major updates in our revision. Following is a detailed response to Review #1 comments.\\n\\n1) Details of design choices. \\nOur revision Appendix A.3 describes details of our design choices and the rationale behind making them. These choices were based on reference to previous common settings and our reasoning, rather than tuning. For example, regarding why using -0.1 as the reward scaling constant: we observed that for the Monte Carlo sampling, the probability of hitting equivalent expressions is around 10%; therefore, we set the penalty to be 0.1 to balance the reward and penalty, such that the average reward is close to 0. Also note that this constant is not a crucial hyperparameter, because the baseline removal procedure of REINFORCE will take the place to balance the reward and penalty during training. Appendix A.3 has more details about various design choices.\\n\\n2) Ablation studies. \\nOur revision Appendix C presents our ablation studies. Figure 5 shows a quantitative breakdown of the contributions of each design component. In particular, the performance significantly drops if we remove either the subtree embedding similarity loss or the subtree selector. If a linear seq2seq model is used instead of the tree-LSTM, the algorithm almost completely fails. The following are more detailed explanations.\\n\\n- The contribution of the embedding similarity loss is to improve training convergence (discussed and evaluated in Appendix C.1). Without the concrete training signal provided by this loss, the training can be stuck at a poor local optimum due to REINFORCE. \\n\\n- The contribution of the subtree selector is to ensure a broad coverage of simplifiable subexpressions, by helping pinpoint the subtrees that can be simplified (discussed and evaluated in Appendix C.2 and Section 5.4). Without the subtree selector, the algorithm will tend to overlook certain small subexpressions that can be simplified.\\n\\n- Appendix C.3 explains why tree-LSTM is much more effective than linear seq2seq. In short, a linear seq2seq model will have a hard time learning the correct syntax. With seq2seq, the probability of hitting a syntactically correct expression is already low, not to mention the probability of hitting any equivalent expressions. \\n\\n- The contribution of curriculum learning is to ensure a sufficiently high probability of hitting the equivalent expressions, in order to provide a meaningful training signal in RL.\\n\\n3) Hyperparameters. \\nOur revision Appendix A.3 describes our hyperparameter setting and our decision making in detail. The original results reported are from a single run of the inference. To address the review concern about the potential deviation in results if more runs were performed, our revision performs a multiple-run evaluation on the experiment described in Section 5.2. The corresponding results are reported in Appendix B.3, demonstrating consistently promising results.\"}", "{\"title\": \"Revised Paper Uploaded\", \"comment\": \"We thank the reviewers for their valuable feedback. Major concerns include the need for a more detailed description of our design choices, more explanation on experiment results, additional experiments, and ablation studies. We uploaded a revised version according to review comments.\\n\\nFollowing is a summary of our major updates (less major updates are described in response to each reviewer).\\n\\n1. We add a detailed description of our experiments, including a clearer description of our dataset in Section 5.1, and additional details in Appendix A.\\n\\n2. We perform a set of extensive ablation studies on key model designs in Appendix C, including embedding similarity loss (C.1), subtree selector (C.2), and tree-LSTM structure (C.3). We not only compare the performance with and without these designs, but also provide intuitive and empirical explanations of why they work.\\n\\n3. We add new experiment results, including a back-to-back comparison of the simplification processes of HISS and Halide (Section 5.4), and results with various initializations (Appendix B.3). The highlight of the new experiments is in Tables 1 and 2. Due to page limitation, the original Sections 5.3 to 5.5 are moved to the Appendix B and C.1.\\n\\n4. We update the results in the experiment introduced in Section 5.3. Our original experiment did not train the model until convergence due to time constraints and the result reported was produced by the half-trained HISS. Now the result is updated with that produced by the fully-trained HISS. Yet, our observations and conclusion drawn from the updated result are in line with our original ones.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper presents a method for symbolic superoptimization \\u2014 the task of simplifying equations into equivalent expressions. The main goal is to design a method that does not rely on human input in defining equivalence classes, which should improve scalability of the simplification method to a larger set of expressions. The solution uses a reinforcement learning method for training a neural model that transforms an equation tree into a simpler but equivalent one. The model consists of (i) a tree encoder, a recursive LSTM that operates over the input equation tree, (ii) a sub-tree selector, a probability distribution over the nodes in the input equation tree, and (iii) a tree decoder, a two layer LSTM that includes a tree layer and a symbol generation layer. The RL reward uses an existing method for determining soft equivalence between the output tree and the input tree along with a positive score for compressing.\\n\\nThe main strengths of the paper are that (i) it targets the scalability problem in simplifying arithmetic expressions by reducing the amount of human effort involved in the process, (ii) it provides some evaluation comparing against methods that use pre-defined rules for transforming equations, (iii) it provides a baseline method against which future scalable models can be compared against.\\n\\nThe following are the main concerns I have with the paper\\n\\n1) The model description should include more details about the design choices. \\nFor instance, the specific reward function, the central component of the model, is left rather under-discussed. The function form is sensible but why does the negative case use -0.1 as the reward scaling constant? Why not some other number? How is this tuned? \\n\\n2) As far as I can see there is no real ablation analysis that shows which of the components are actually useful. Is the sub-tree selector necessary? Is the curriculum training necessary? How much does the sub-tree embedding similarity loss contribute to the results? Even if each of these actually add value it will be useful to know how much. What about other design choices? If we trained a direct seq2seq model with linearized expressions instead of the tree structured inputs, would it work just as well? These are empirical questions that need to be answered to justify that the proposed model indeed is useful. \\n\\n3) The experimental details are sparse. In particular, there is no mention of how hyper-parameters of the proposed method are tuned. Are the performance numbers averages over a set of random seeds or is it simply the best performing number that has been reported? This is especially troublesome for a RL based model. \\t\\n\\nOverall the paper presents a particular model and strategy for training but lacks appropriate experimentation to establish their utility.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors present a framework for symbolic superoptimization using methods from deep learning. A deep learning approach operating on the expression tree structures is proposed based on a combination of subtree embeddings, LSTM RNN structures, and an attention mechanism.\\n\\nThe approach avoids the exploitation of human-generated equivalence pairs thus avoiding human interaction and corresponding bias. Instead, the approach is trained using random generated data. It remains somewhat unclear how the corresponding random data generation influences general applicability w.r.t. other tasks, as the authors apply constraints on the generation process for complexity reasons. A corresponding discussion would be valuable here.\\n\\nIn Secs. 3 & 4, the authors present their specific modeling and learning approach. However, they do not report on modeling or learning alternatives. It would be interesting for the audience to understand, how the authors reached these specific choices, and how (some of) these choice influence performance and learning stability. For example, in Sec. 4.1, an additional loss term is introduced to further support the learning of embeddings. However, it might interesting to see comparative results quantitatively investigating the effect of this additional loss term. Also, as far as I can see, no information on the choice of hyperparameters (e.g. LSTM dimensions) are provided or analyzed w.r.t. their effect on the performance of the proposed approach.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper provides a novel approach to the problem of simplifying symbolic expressions without relying on human input and information. To achieve this, they apply a REINFORCE framework with a reward function involving the number of symbols in the final output together with a probabilistic testing scheme to determine equivalence. The model itself consists of a tree-LSTM-based encoder-decoder module with attention, together with a sub tree selector. Their main contribution is that this framework works entirely independent of human-labelled data. In their experiments, they show that this deep learning approach outperforms their provided human-independent baselines, while sharing similar performance with human-dependent ones.\\nOverall, the work in this paper has the potential to be a contribution to ICLR but lacks completeness and clarity. A number of experimental details are missing, making it difficult to understand the setup under which results were obtained. Moreover, the paper does not seem to have been revised with many grammatical issues that make it hard to read.\", \"the_following_are_major_issues_within_the_paper_and_should_be_addressed\": \"\\u2022\\tThe paper does not mention the amount of compute given to their model, nor the amount of time taken to train. As the REINFORCE framework is generally quite computation-heavy, these are significant details. Without assessing the amount of compute and time allotted for training HISS, the comparisons to previous baselines lose a fair amount of meaning. The paper alludes to processes being \\u2018extremely time consuming\\u2019, but then does not provide any numbers.\\n\\u2022\\tThey do not mention the data used to train the model weights. In the comparisons sections, some details on datasets are given, but these seem to refer to data for inference.\\n\\u2022\\tThere are many grammatical errors that likely could have been detected with \\t revision. A handful of such errors would not affect the score, but they are so numerous as to make the paper much more difficult to understand.\\nAdditionally, these are comments that slightly detract from the quality of the paper:\\n\\u2022\\tIt\\u2019s unclear what to glean from Section 5.1, as the dataset and baselines seem to be fairly trivial. If their claim is to have the first nontrivial human-independent approach to simplifying symbolic expressions, there is no need to compare to baselines that can only handle small expressions.\\n\\u2022\\tSections 5.3 and 5.4 contribute little to the paper. For 5.3, the model was trained to embed equivalent expressions close together, using L2 regularization. It is therefore unsurprising that equivalent expressions are then closer together than non-equivalent ones. The paper also does not provide a comparison to the method without this regularization, and so it\\u2019s unclear if this embedding similarity helps in any way. For 5.4, the section is extremely short and contains very little content. Moreover, just as many of the variables in their provided examples oppose their conjectures as support them.\\n\\u2022\\tThe most interesting figure provided is the rewrite rules discovered by the model. It would be even better if an additional column containing the rules discovered by Halide (the main baseline) were provided.\\nOverall, in my understanding, the primary point in favor of the paper is in being the first nontrivial human-independent approach to simplifying symbolic expression. That said, this is not my area of expertise, so I cannot judge novelty or importance as well as other reviewers.\"}" ] }

[ "Matthieu Kirchmeyer", "Patrick Gallinari", "Alain Rakotomamonjy", "Amin Mantrach" ]

[ "domain adaptation", "imputation", "missing data", "advertising" ]

{ "note_id": [ "maRpdTbYwv", "B1xhZRqYjS", "BJxzTp9tsB", "HkxWJa9KsH", "HJeBF25tjH", "r1gWD5cFiH", "HJlx9WDWcS", "ryxFdPLRYH", "Ske7FfJTFB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730613, 1573658115728, 1573658041944, 1573657816563, 1573657725161, 1573657177516, 1572069768084, 1571870576823, 1571775099256 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1716/Authors" ], [ "ICLR.cc/2020/Conference/Paper1716/Authors" ], [ "ICLR.cc/2020/Conference/Paper1716/Authors" ], [ "ICLR.cc/2020/Conference/Paper1716/Authors" ], [ "ICLR.cc/2020/Conference/Paper1716/Authors" ], [ "ICLR.cc/2020/Conference/Paper1716/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1716/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1716/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper addresses the problem of performing unsupervised domain adaptation when some target domain data is missing is a potentially non-stochastic way. The proposed solution consists of applying a version of domain adversarial learning for adaptation together with an MSE based imputation loss learned using complete source data. The method is evaluated on both the standard digit recognition datasets and a real-world advertising dataset.\\n\\nThe reviewers had mixed recommendations for this work, with two recommending weak reject and one recommending acceptance. The key positive point from R3 who recommended acceptance was that this work addresses a new problem statement which may be of practical importance. The other two reviewers expressed concerns over the contribution of the work and the validity of the problem setting. Namely, both R2 and R4 had significant confusion over the problem specification and/or under what conditions the proposed setting is valid. \\n\\nIt is a difficult decision for this paper as there is a core disagreement between the reviewers. All reviewers seem to agree that the proposed solution is a combination of prior methods in a new way to address the specific problem setting of this work. However, the reviewers differ in precisely whether they determine the proposed problem setting to be valid and justified. Due to this discrepancy, the AC does not recommend acceptance at this time. If the core contribution is to be an application of existing techniques to a new problem statement than that should be clarified and motivated further.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Answer Reviewer 2 - Clarified problem specifications and missing data assumptions (Part 1/2)\", \"comment\": \"Thank you for your review. We have taken note of your comments and try to provide detailed answers to the points you raised:\\n\\nA) Problem specification:\\n \\nAll the data considered in the paper are vectors. Let $x \\\\in \\\\mathbb{R}^n$ be a complete data vector and $m \\\\in \\\\{0,1\\\\}^n$ a binary mask indicating which entries of $x$ are missing (1 for missing and 0 for observed). Given a dataset $X = (x_{(i)})_{i \\\\in \\\\{1, ..., N\\\\}}$, we define the missingness indicator matrix as $M=(m_{(i)})_{i \\\\in \\\\{1, ..., N\\\\}}$ where $N$ is the number of instances. In the following, we remove index $i$ for clarity. The hypotheses are the following:\\n\\n1. Source domain data are fully observed.\\n\\n2. Target domain data are partially observed and the missingness pattern is fixed: the indicator mask $m$ is the same for all the target domain data. In the paper when one refers to \\\"non-stochastic missing data\\\" we mean that the mask pattern is fixed for all target domain data. For simplicity, we have used the notations $x_S=(x_{S_1},x_{S_2})$ (source) and $x_T=(x_{T_1},x_{T_2})$ (target) with $x_{S_1}$ and $x_{T_1}$ corresponding to the mask positions with 0 value, i.e. features observed on both domains, and $x_{S_2}$ and $x_{T_2}$ corresponding to mask positions with value 1, i.e. missing features for the target domain.\\nPlease see the discussion below on the classical terminology used for missing data (item \\\"Rubin's theory for missing data\\\").\\n\\n3. $x_{S_2}$ and $x_{T_2}$ contain some information not present in $x_{S_1}$ and $x_{T_1}$. This means that the feature values $x_{S_2}$ and $x_{T_2}$ cannot be predicted directly (e.g. through a regressor) from the features $x_{S_1}$ and $x_{T_1}$ respectively.\\n\\n4. The distribution of features $x_{T_2}$ conditioned on the features $x_{T_1}$ could be inferred provided that we have some supervision for training the conditional distribution model; this is typically the case if for some $x_{T_1}$ observation, one also observes $x_{T_2}$, however such a supervision is not available for our problem. This is where adaptation comes into play. For inferring the conditional probability, the model makes use of the source domain information for which this supervision is available while adapting to the target domain. Note that the model operates in a latent space and not in the original one. It does not attempt to reconstruct the true conditional distribution of $x_{T_2}$, but a conditional distribution in a latent space for a projection of $x_{T_2}$ denoted $\\\\widetilde{x}_{T_2} $ in the paper.\\n\\n5. Finally, we make the typical covariate shift assumption seen in several UDA papers, such as DANN [Ganin2015] to address an unsupervised classification adaptation setting.\\n\\nThese hypotheses map to several real world problems as mentioned in this paper and the methods introduced show important improvement on our datasets based on these specifications.\\n\\n* Rubin's theory for missing data: \\n\\nThe foundations of missing data theory were established by Rubin [Rubin1976] and his colleagues [Little2014]. We briefly introduce this formalism in order to put in evidence the specificity and originality of our problem. Rubin distinguishes between a missingness pattern, which describes which values are missing and observed in the data and the missingness mechanism, which represents the statistical relationship between the probability of missing data and the data variables. Let $m$ define as above a pattern of missing data; the missingness mechanism is characterized by the conditional distribution of $m$ given $x$, $p_{\\\\phi}(m | x) ~\\\\forall x \\\\in X$ where $\\\\phi$ denotes a vector of unknown parameters describing the relationship between the $m$ and $x$ variables. $\\\\phi$ is known as the mechanism of missing data and provides the basis for distinguishing between three categories of missing data problems: Missing Completely at Random - MCAR ($\\\\forall x \\\\in X, p_{\\\\phi}(m|x)=p_{\\\\phi}(m)$), Missing At Random - MAR ($\\\\forall x \\\\in X, p_{\\\\phi}(m|x)=p_{\\\\phi}\\\\left(m| x ^{\\\\mathrm{obs}}\\\\right)$ with $x^{obs}$ being the observed features), and Missing Not At Random - MNAR that covers all the other cases.\\n\\nOur missingness mechanism trivially corresponds to MCAR on the target domain. However, the key idea behind Rubin's theory is that missingness is a variable with a probability distribution. In our case, the missingness pattern is deterministic, not stochastic. The problem is then more difficult than classical MCAR problems and does not lead for example to classical maximum likelihood solutions exploited in the literature.\"}", "{\"title\": \"Answer Reviewer 2 - Clarified problem specifications and missing data assumptions (Part 2/2)\", \"comment\": \"B) Other points:\\n\\n* The original contribution is the joint adaptation-imputation model. In the paper we propose two instances of this model, one is based on adversarial training as you mention, the other one is based on distribution matching by optimal transport. Other mechanisms could have been used as well, but these two are representative of two state of the art families of adaptation methods. As indicated we found out that both methods lead to similar performance on our datasets, but each require a specific parameter tuning. \\n\\n* Concerning the validity of the method, we have no formal proof of consistency. However, the experiments show that under the hypotheses specified above, the method offers systematic improvement w.r.t. a selection of representative baselines and for different instances of the adaptation-imputation problem. The results have been obtained on datasets with extremely different characteristics, including real world data with all their intrinsic complexity, and validated by intensive ablation studies. Of course, we do not claim to cover all situations, but we believe that our experiments already bring a strong evidence for the potential of the proposed method.\\n\\n* The reference to [Ben-David2010] as an example of domain shift was indeed an error and was removed. The work of [Ben-David2010] has however been fundamental for motivating and justifying several approaches to the adaptation problem. [Ganin2015] for example builds on the ideas introduced in [Ben-David2010] for justifying their approach and for approximating the H-divergence quantity introduced in [Ben-David2010]. Their conclusions are also valid for our ADV model.\\n\\n* We use the term component in its classical mathematical or physical acceptation. All the data we deal with are represented as vectors and a component of a vector is a feature or a subset of features of this vector.\\n\\n* Regarding the notation for referring to the input data $x$, we have consistently used a dash \\\"-\\\" to refer either to the Source or the Target domain and then used 1 or 2 in subscript to S or T as a way of differentiating components 1 (always observed) and 2 (observed only for source data) in a S or T instance.\\n\\n\\n[Ganin2015] Ganin et al. Unsupervised Domain Adaptation by Backpropagation. ICML 2015\\n[Little2014] Little and Rubin. Statistical analysis with missing data, volume 333. John Wiley & Sons, 2014\\n[Rubin1976] Rubin. Inference and Missing Data. Biometrika, 63, 581-592. 1976\\n[Ben-David2010] Ben-David et al. A theory of learning from different domains. 2010\"}", "{\"title\": \"Answer Reviewer 4 - Clarified joint domain shift and missing data for ads experiments (Part 1/2)\", \"comment\": \"Thanks a lot for your feedback and your recommendations. We provide below a detailed response.\\n\\nA) Joint domain shift and missing data hypothesis for the ads experiments: \\n \\nWe do agree that this is not obvious. As mentioned in the paper, the ads problem was our initial motivation for this work and our formulation of the problem comes from preliminary exploratory data analyses performed on ads datasets. We have added in Appendix E in the new paper version, distribution plots (Figure 6) and mean values (Table 6) for the different observed features used in the ads-kaggle dataset for the source and target domains. This shows that there is indeed a domain shift between the two domains. The same conclusion holds for the ads-real dataset. More details are provided below.\\n \\nYour description of the problem in the detailed comments (*Questions that impacted rating.*) is basically right. We figured out however that we might not have been precise enough in the text and we provide below more details clarifying the experimental setting. \\n\\nThe source dataset is composed of all user-partner pairs for which the user visited the partner and the target dataset is composed of all the user-partner pairs for which the user never visited this partner. A key point here is that there are several partners (and of course users) per domain, and this was probably not clear enough from the text. Typically we could expect thousands of partners depending on the size of an ads company's partner portfolio. For the source domain, we have available complete data (mean statistics on all visited partners + traces on a specific ad partner for a user-partner pair) and for the target domain only partial data (mean statistics but no partner specific traces for a user-partner pair). \\nRegarding domain shift, in Figure 6 Appendix E, we plot the normalized feature distributions for the source (blue plots) and target (red plots) domains. While some features have a similar distribution for the source and target domains, many have completely different distributions indicating a clear domain shift. This is synthesized in Table 6 Appendix E, giving the mean values for all the features for the two domains. We notice that feature 5 is missing on the target.\\n\\nThis shift was initially a surprising finding for us too. Our hypothesis is that the source domain includes users with a higher overall activity both for visiting partner websites and for interacting with the websites. Target domain includes users that are probably less active. This is confirmed by the mean value of the features in the two domains in Table 6 Appendix E: feature distributions from users in the source domain tend to have higher mean values than features in the target domain. These features typically measure click, visit and sale activities which is consistent with the above hypothesis.\"}", "{\"title\": \"Answer Reviewer 4 - Clarified joint domain shift and missing data for ads experiments (Part 2/2)\", \"comment\": \"B) Other (*Additional notes.*)\\n\\n1. We will update the paper in a revised version with the reviewer's recommendations for a more direct explanation of the motivations. We will as suggested rely on some examples from the literature such as the ones shortly described below, that address imputation. These methods do not perform adaptation.\\n \\n[Cai2018]: presents an application of missing data for multi-modality clinical applications, such as tumor detection and brain disease diagnosis. On these problems, different modalities can usually provide complementary information, which commonly leads to improved performance. However, some modalities are often missing for some subjects due to various technical and practical reasons and thus multi-modal data might be incomplete. This paper, which is the closest to ours in the literature, attempts at reconstructing a fixed missing modality on target subjects with direct supervision from fully observed training instances.\\n \\n[Tran2017]: addresses an object recognition problem with multi-modal sensors data. It is common on these applications that the sensing equipment experiences unforeseeable malfunction or configuration issues, leading to corrupted data with missing modalities. Most existing multi-modal learning algorithms could not handle missing modalities, and would discard either all modalities with missing values or all corrupted data. This paper relies on a stochastic distribution of the missing modalities while we deal with fixed missing modalities.\\n\\n[Wang2018]: presents a model for rating prediction tasks with multi-modal data and draws the link between cold-start and missing modality considering the same setting as [Tran2017] in the context of recommendation.\\n\\n2. In the remaining non-missing 30% of the image there might still be important pixels (such as the top part of a \\\"7\\\" digit which could be mistaken for a \\\"1\\\" digit if not present) which can justify why the partial model is not performing better than the imputation model which could have imputed this missing information correctly. The difference with the missing model is probably due to suboptimal hyperparameters for the partial model as these were tuned for $p = 0.5$ and used as such for the other instances of $p$. Each of the three models should converge to the value of 77.6% (mentioned in Table 1 in the SVHN $\\\\rightarrow$ MNIST Adaptation-Full line) when $p \\\\rightarrow 0$.\\n\\n[Cai2018] Cai et al. Deep adversarial learning for multi-modality missing data completion. KDD 2018\\n[Wang2018] Wang et al. LRMM: Learning to Recommend with Missing Modalities. EMNLP 2018\\n[Tran2017] Tran et al. Missing Modalities Imputation via Cascaded Residual Autoencoder. CVPR 2017\"}", "{\"title\": \"Answer Reviewer 3 - Replied to comments\", \"comment\": \"We gratefully thank the reviewer for acknowledging the effort put in the experimental evaluation, for pointing out typos and advise on the structure of the paper.\\n\\n* Indeed as the reviewer notices, not all hyperparameters are used in practise in the experimental section. During initial tests, we experimented with different values of the hyperparameters and finally only tuned $\\\\lambda_{MSE}$ in the experiments. We kept the initial formulation to indicate that the other parameters could also be tuned further and might yield improved classification performance. We have added a sentence clarifying this in the revised version in Section 4.1.\\n\\n* OT vs ADV: in preliminary tests, we figured out that the NN architectures used for mapping data onto the latent space had usually a much higher influence on the performance than the choice of the alignment method itself. In order to provide a fair comparison and also because our goal was to show the effectiveness of the proposed mechanism rather than reaching the best possible performance, we decided to use NN architectures with similar complexity for both models. The OT models used in [Damodaran2018] indeed require an order of magnitude more parameters than the ADV models [Ganin2015] to reach similar performance. Concerning the approximation, the OT model in [Damodaran2018] uses an alignment on the joint (X,Y) distributions while we are only aligning the X distributions. Hence, our approach is equivalent to a primal Wasserstein version of the domain adaptation method proposed in [Shen 2018]. In our tests, aligning on the joint distribution did not perform better than aligning only on the marginal. Then, our most plausible explanation is still that the difference of performance is due to tuning.\\n\\n* The Naive model is referenced in the Baselines paragraph (in the revised version it figures in Section 5.1) . This model refers to the likelihood computed using as a prediction the mean CTR on the training set and is a typical baseline used on advertising problems.\\n\\n* We have moved the related works to Section 2 in the revised paper upon the reviewer's recommendation and corrected the typo.\\n\\n* Indeed there is in practise no guarantee that the target domain will be balanced and usually it is not. However, dealing with both domain shift and label shift is more complex and requires additional mechanisms. Most work up to now only considered one of the two hypothesis (either domain or label shift). Recent work such as [Zhao2019], [Wu2019] has started to examine more complex settings and it is of course of strong practical interest. We have also started analyzing joint domain shift and unequal class distributions. \\n\\n[Ganin2015] Yaroslav Ganin et al. Unsupervised Domain Adaptation by Backpropagation. 2015\\n[Damodaran2018] Bharath Bhushan Damodaran et al. DeepJDOT : Deep Joint Distribution Optimal Transport for Unsupervised Domain Adaptation. 2018\\n[Zhao2019] Han Zhao et al. On learning invariant representation for domain adaptation. ICML 2019\\n[Wu2019] Yifan Wu et al. Domain adaptation with asymmetrically-relaxed distribution alignment. ICML 2019\\n[Shen2018] Shen Ji et al. Wasserstein Distance Guided Representation Learning for Domain Adaptation, AAAI 2018\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper proposed to address a compound problem where missing data and distribution shift are both at play. The paper goes on to describe some heuristic methods that resemble the gradient reversal methods due to Ganin et al for handling both problems.\\n\\nThe novel part of the paper over DANNs is the joint, end-to-end training of latent representations for missing data, \\u00a0While it is sloppy with terminology, the paper is overall reasonably easy to follow although it might mislea a novice reader and sufficient details are provided to replicate their results. \\n\\nThe major problem here is the problem appears to be underspecified, and its not clear under what conditions if any the proposed methods are valid. Moreover it\\u2019s not clear to what extent the experimental results should ameliorate these concerns. \\n\\nIf the data is not missing at random then there is presumably confounding. The authors dance around this topic, just asserting that they are handling non-stochastic missing data but do not say precisely what is assumed about the relationship between the observed and missing data. \\n\\nIn short the paper addresses an under-specified problem with a heuristic technique based upon domain-adversarial nets which have recently been shown have a number of fundamental flaws. It's never made clear under what assumptions this proposed procedure is valid and the paper misrepresents the prior work on lable shift, including the theoretically sound work, e.g.:\\n\\n\\\"we assume covariate shift as in most UDA papers e.g. Ben-David et al. (2010); Ganin & Lempitsky (2015).\\u201d\\n>>> Ben-David 2010 is not about covariate shift \\u2026.\", \"some_minor_thoughts\": \"\\u201csome components of the target data are systematically absent\\u201d\\n>>> \\tNot clear what \\u201ccomponent\\u201d means at this point\\n\\n\\u201cWe propose a way to impute non-stochastic missing data\\u201d\\n>>> \\tWhat does this mean? Is non-stochastic, not missing at random? What is the pattern of missing-ness conditioned on? What assumption, if any, is made? \\n\\n\\u201cThis key property allows us to handle non-stochastic missing data,\\u201d \\n>>> \\tagain what precisely does this mean?\\n\\n\\u201cConsider that x has two components (x_1, x_2)\\u2026\\u201d\\n>>>\\tsloppy notation:\\n\\t\\u201cSource features\\u201d x_s = (x_S1, x_S2) are always available \\n\\n\\nI read the author's reply but do not believe that the responses are satisfactory. The authors do not address the primary concerns clearly and do not point to specific improvements in the draft that might cause me to change my mind.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"*Summary.* The paper presents and addresses the problem of performing domain adaptation when the target domain is systematically (i.e., not the result of a stochastic process) missing subsets of the data. The issue is motivated by applications where one modality of data becomes unavailable in the target domain (e.g., when deciding which ads to serve to new users, the predictor may have access to behavior across other websites but not on a specific merchant's website). The proposed method learns to map source and target data to a latent space where the representations for the source and target are aligned, the missing components of the target can be inferred, and classification can be performed successfully. These are achieved by adversarial/optimal transport loss on source and target features, a mean-squared error and adversarial loss on latent generation/imputation, and a cross entropy loss on source label prediction, respectively. Experiments are performed on digits and click-through rate (CTR) prediction and include a thorough set of baselines/oracles for comparison.\\n\\n*Review.* While the problem statement is novel, I am unconvinced that the advertising experiment includes both a domain adaptation and imputation problem. I describe this in detail below. For this reason, I am giving the paper a weak reject.\\n\\n*Questions that impacted rating.*\\n1. Ads experiment: From my understanding, the source domain is the traffic of users who have interacted with (clicked through to?) a specific partner and the target domain is the traffic of the users who have not interacted with that specific partner. The data that needs to be imputed is the click through rate for target users with that specific partner. In this case, it is not obvious to me why there is a domain shift between these two groups of users. This would imply that the traffic of source users and target users is different for other partners. I don't see why this would need to be true. Could the authors provide an explanation as to why this is the case (e.g., by showing that CTRs differ with other (partner, publisher) pairs between source and target). From my understanding, Table 5 only shows CTR averaged across all users in each domain, but does not show that the CTRs differ between source and target users for contexts/(partner, publisher) pairs (i.e., the results in table 5 could be due to the fact that the prior distribution over context is different for source and target users).\\n\\n*Additional notes. Immaterial to rating.*\\n1. I personally felt that the motivation for UDA vs imputation in the first paragraph was a bit muddled. I think sticking to one example would make the motivation more clear to the reader. E.g., explain the prediction problem for medical imaging (which I assume is disease diagnosis, but it is not stated explicitly), describe how some medical imaging may be missing for certain patients (imputation), then explain that there may be noise across different medical imaging systems (UDA), then list the other applications where this arises with citations (e.g., These phenomena have also been documented in advertising applications [1], ...).\\n2. I was surprised by the difference between Adaptation-Partial and the other two train/test conditions in Figure 2 when p=30%. Out of curiosity, do the authors have an explanation for this discrepancy? I would have predicted that, if most of the information necessary for prediction was available in the remaining 70% of the image that the performance of these cases would be very similar. I think it would be helpful to see the accuracy on the source domain and the labeled target domain to better understand that result.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"The submission describes an approach for unsupervised domain adaptation in a setting where some parts of the target data are missing.\\n\\nBoth UDA approaches as well as data completion approaches have a sizable research history, as laid out in the related work section (Section 5). The novelty here comes from the properties that a) domain adaptation and data imputation are handled in a joint manner, b) the missing data in the target domain is non-stochastic, and c) imputation is performed in a latent space. This maps to a fairly specific, but realistic enough set of real-world problems; the authors give an image recognition as well as an advertising prediction related problem as experimental examples.\\n\\nThe submission is overall well written and easy to understand. I'd rate the novelty as medium (smart combination of existing methods), but the exemplary experimental evaluation elevates it to more than a systems paper.\\n\\nThe method is described clearly in Section 3, and the joint training makes sense. I notice that not all hyperparameters ({lambda_adv, lambda_mse}, {lambda_1, lambda_2, lambda_3}) are truly needed. lambda_adv and lambda_1 could be canonically set to 1 for such a loss minimization problem, so why are the extraneous parameters included?\\n\\nIn addition to Section 3, the experimental evaluation on two very different data sets in Section 4 is highly detailed and describes the insights clearly, both qualitatively and quantitatively. I'm happy that mean standard deviations are reported on an acceptable experiment sample set size.\", \"regarding_the_different_approaches\": \"I'm wondering whether the higher performance of the ADV approach over OT (or the parameter hunger of OT over ADV) is only due to the tuning of the network architectures, or whether this is due to the approximations described in B.1.\\nThe ablation study in Section 4.4 is interesting w.r.t. the trade-off it shows between stable, consistent, \\\"average\\\" results from an MSE loss term, vs. high-variance (and on average better) results when a choice of mode is forced using an adversarial loss term.\", \"minor_comments\": [\"In Table 2, I am not sure what the first row ('Naive') refers to. As far as I can tell, it is not referenced in the text.\", \"I would move Section 5 (related work) to right after the introduction, as is common in conference papers and makes for smoother reading.\", \"Section 5.2: type \\\"impainting\\\" -> \\\"inpainting\\\"\", \"Appendix, section 'Pre-processing': It seems to me that there is a clear assumption made that the target set is balanced, since training happens with a balanced source set. Is this realistic in practical scenarios? There is work on DA with unequal class distributions between domains.\", \"In summary, I can clearly recommend this submission for publication.\"]}" ] }

[ "Pan Xu", "Felicia Gao", "Quanquan Gu" ]

[ "Policy Gradient", "Reinforcement Learning", "Sample Efficiency" ]

{ "note_id": [ "_guXzMQ6d6", "B1ljJcBFjS", "r1lWNm6Osr", "HyxX-XTusH", "HJgY3GadoS", "r1gEFzT_jr", "HJe3Ea2AYr", "r1xTId76KH", "rJxxBuwItB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730583, 1573636579137, 1573602088647, 1573602043445, 1573601968614, 1573601915729, 1571896627588, 1571792980562, 1571350584044 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1715/Area_Chair1" ], [ "ICLR.cc/2020/Conference/Paper1715/Authors" ], [ "ICLR.cc/2020/Conference/Paper1715/Authors" ], [ "ICLR.cc/2020/Conference/Paper1715/Authors" ], [ "ICLR.cc/2020/Conference/Paper1715/Authors" ], [ "ICLR.cc/2020/Conference/Paper1715/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1715/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1715/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"The paper introduces a policy gradient estimator that is based on stochastic recursive gradient estimator. It provides a sample complexity result of O(eps^{-3/2}) trajectories for estimating the gradient with the accuracy of eps.\\nThis paper generated a lot of discussions among reviewers. The discussions were around the novelty of this work in relation to SARAH (Nguyen et al., ICML2017), SPIDER (Fang et al., NeurIPS2018) and the work of Papini et al. (ICML 2018). SARAH/SPIDER are stochastic variance reduced gradient estimators for convex/non-convex problems and have been studied in the optimization literature.\\nTo bring it to the RL literature, some adjustments are needed, for example the use of importance sampling (IS) estimator. The work of Papini et al. uses IS, but does not use SARAH/SPIDEH, and it does not use step-wise IS.\\n\\nOverall, I believe that even though the key algorithmic components of this work have been around, it is still a valuable contribution to the RL literature.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Thanks for your reviews. Please take a look at the rebuttal.\", \"comment\": \"Dear reviewers,\\n\\nThank you very much for your efforts in reviewing this paper.\\n\\nThe authors have provided their rebuttal. It would be great if you take a look at them, and see whether it changes your opinion in anyway. If there is still any unclear point or a serious disagreement, please bring it up. Also if you are hoping to see a specific change or clarification in the paper before you update your score, please mention it.\\n\\nThe authors have only until November 15th to reply back.\\n\\nI also encourage you to take a look at each others\\u2019 reviews. There might be a remark in other reviews that changes your opinion.\\n\\nThank you,\\nArea Chair\"}", "{\"title\": \"Response to review #3\", \"comment\": \"Thank you for the helpful comments, which we address as follows.\", \"q1\": \"\\\"In section 3, it is not quite clear how the reference policy is defined, and the $\\\\theta^s$ is not clearly defined when $s\\\\geq1$.\\n\\\"\", \"a1\": \"Thanks for pointing out the missing definition. In the initialization, the reference policy is set to be the initial policy $\\\\tilde\\\\theta^0=\\\\theta_0$. For $s=0,1,\\\\ldots$, at the end of the $s$-th epoch, we update the reference policy as $\\\\tilde\\\\theta^{s+1}=\\\\theta_{m}^{s+1}$, where $\\\\theta_{m}^{s+1}$ is the last iterate of the $s$-th epoch. We have rewritten this part in Line 3 and Line 12 of Algorithm 1 in the revision. We did not use the notation $\\\\theta^s$ for $s\\\\geq 1$ in our paper. We guess you are referring to $\\\\theta_0^{s+1}$ in Line 3 of Algorithm 1 (the initial behavior policy at the beginning of each epoch), which is also set to equal to the reference policy $\\\\tilde\\\\theta^{s+1}$ (see Line 3 of Algorithm 1).\", \"q2\": \"\\\"In the main part of the paper, the discussion in Remark 4.6 and the following Corollary 4.7 is not quite clear.\\\"\", \"a2\": \"We have rephrased the discussion in Remark 4.6 and Corollary 4.7 in the revision. To summarize, in Remark 4.6, we compare our convergence rate with that in Papini et al. (2018) and show that our result does not have the additional additive term $1/B$ where $B$ is the batch size in inner loops. We also compare our results with that in Xu et al. (2019) and show that our batch size $B$ does not depend on the horizon length $H$, which is another big improvement.\\n\\nIn Corollary 4.7 and the discussion after it, we calculate the total number of samples (trajectories) that are required by our algorithm in order to find an $\\\\epsilon$-approximate stationary point. We show that our sample complexity is lower than that of Xu et al. (2019) by a factor of $O(1/\\\\epsilon^{1/6})$.\\n\\n\\nThank you for pointing out the typos. We have fixed them in the revision. We have removed the notation $T$ since it can be replaced by $T=Sm$ which makes the dependency clearer. We added the choice of $S=O(1/\\\\epsilon^{1/2})$ in Corollary 4.7.\"}", "{\"title\": \"Response to review #2\", \"comment\": \"Thank you for your constructive comments. We respond to your concerns point by point as follows.\", \"q1a\": \"\\\"The wording of theorem 4.5 and corollary 4.7 could be somewhat clarified. In particular, I did not see $\\\\Phi$ defined in the main text, and given its definition in the appendix, I believe theorem 4.5 could be stated simply in terms of J, avoiding any additional notation.\\\"\", \"a1a\": \"Thanks for the suggestion. We have removed $\\\\Phi$ in the statement of Theorem 4.5 and used $J$ and the indicator function according to your advice.\", \"q1b\": \"\\\"Similarly, corollary 4.7 could be stated somewhat more clearly, and in particular, the choice of S should be made explicit.\\\"\", \"a1b\": \"By Theorem 4.5 and Remark 4.6, we have $Sm=T=O(1/\\\\epsilon)$. Therefore, the choice of $S$ in Corollary 4.7 is $S=T/m=O(1/\\\\epsilon^{1/2})$ since we have set $m=O(1/\\\\epsilon^{1/2})$. We have added this in the statement of Corollary 4.7 in the revision.\", \"q1c\": \"\\\"In the appendix, I could not find a definition for $\\\\varphi$.\\\"\", \"a1c\": \"We are sorry for the typo in equation (B.3) in the appendix. The function $\\\\varphi_{\\\\mathbf{\\\\Theta}}$ should be the set indicator function $\\\\mathbf{1}_{\\\\mathbf{\\\\Theta}}$ over the set $\\\\mathbf{\\\\Theta}$. We have replaced $\\\\varphi_{\\\\mathbf{\\\\Theta}}$ by the set indicator function $\\\\mathbf{1}_{\\\\mathbf{\\\\Theta}}$ in equation (B.3) in the revision.\", \"q2a\": \"\\\"The empirical results presented are interesting, although I wish they were more comprehensive. In particular, it would be valuable to more exhaustively evaluate the impact of the hyper-parameters N, B and m.\\\"\", \"a2a\": \"Thank you for the suggestions. According to the proof of Corollary 4.7 in our paper (please see Appendix B), in order to achieve $\\\\epsilon$ precision, $N$ should be in the order of $1/\\\\epsilon$. Since we have $mB=N$ and $N=1/\\\\epsilon$, when we change the scale of $B$, $m$ will be changed accordingly as $m = N/B$. Therefore, we only need to evaluate the selection of hyper-parameter $B$ in our experiment, because once $B$ is chosen, $m$ is determined accordingly.\\\"\", \"q2b\": \"\\\"The authors should also clarify how the current values were chosen.\\\"\", \"a2b\": \"According to Corollary 4.7, the orders of the batch size parameters $N$ and $B$ and the epoch length $m$ should be chosen based on the user-defined precision parameter $\\\\epsilon$. In particular, we require $N=O(1/\\\\epsilon)$ and $mB=O(1/\\\\epsilon)$. Based on our theoretical result in Corollary 4.7 and our sensitivity study on the inner loop batch size $B$, we further choose $B=O(1/\\\\epsilon^{1/2})$ and thus $m=O(1/\\\\epsilon^{1/2})$. In our experiments, we set $N=C_0/\\\\epsilon$, $B=C_1/\\\\epsilon^{1/2}$ and $m=C_2/\\\\epsilon^{1/2}$ and tune the constant parameters $C_0, C_1, C_2$ using grid search.\", \"q2c\": \"\\\"Given that the theoretical results also apply to projected gradient descent, it would be interesting to see empirical results in that case.\\\"\", \"a2c\": \"The projection step is only employed in the theoretical analysis to cover cases where the optimal policy is constrained in certain convex set. In our experiments, we found that the proposed algorithm works well without the extra projection step for all the environments considered in our paper. Therefore, we did not use projection in our experiments.\"}", "{\"title\": \"Response to review #1 (Part 1)\", \"comment\": \"Thank you for the detailed comments. We address them point by point as follows.\", \"q1\": \"\\\"I think you may ignore the highly related work ... I would suggest the author(s) have some discussions to differentiate two papers\\\"\", \"a1\": \"Thank you for pointing out this related work. We were not aware of it at the time of submission. In the revision, we have commented on their paper in the \\\"Additional Related Work\\\" section. The differences between Yang and Zhang (2019) and our paper are summarized as follows:\\n\\n(a) The first version (released in June) of Yang and Zhang (2019) only achieved an $O(1/\\\\epsilon^2)$ sample complexity, which is the same as SVRPG in Papini et al. (2018). In contrast, our paper achieves $O(1/\\\\epsilon^{3/2})$ sample complexity, which is better than theirs. Their second version was recently updated *after* the ICLR submission deadline, which also achieves $O(1/\\\\epsilon^{3/2})$ sample complexity. Therefore, our work is concurrent and independent of their work. \\n\\n(b) Our algorithm is different from that of Yang and Zhang (2019). In particular, our algorithm uses the importance sampling weight to deal with the non-stationarity of the data distribution in reinforcement learning while the algorithm in Yang and Zhang (2019) does not use it. Due to our use of the step-wise importance sampling technique (equation (3.2) and Line 9 in Algorithm 1 in our paper), we prove that the smooth parameter $L$ of the performance function in our algorithm does not depend on the horizon length $H$, which is usually a very large number in practice and can even be infinity in some cases. As a result, the sample complexity of our algorithm does not depend on $H$. In contrast, the sample complexity of the algorithm proved in Theorem 2 in Yang and Zhang (2019) depends linearly on the smooth parameter $L$, which is in the order of $H^2$ (according to their Theorem 1) due to the lack of a proper importance weighting like in our paper. Therefore, the sample complexity of our algorithm is much lower than that of the algorithm in Yang and Zhang (2019).\", \"q2\": \"\\\"Could you please provide the reasons why you are only choosing two methods (GPOMDP and SVRPG) to compare? ... \\\"\", \"a2\": \"In this work, we proposed a new variance reduced policy gradient algorithm (SRVR-PG) based on GPOMDP and SVRPG. We proved that SRVR-PG converges with a lower sample complexity than GPOMDP and SVRPG. Since these two algorithms and our SRVR-PG share similar structures and are comparable in terms of theoretical performance, we chose to compare them in experiments as well. We would also like to emphasize that the A3C and A2C algorithms you pointed are actor-critic type algorithms, which learn the policy (using the actor, i.e., policy gradient) and the value function (using the critic) at the same time. However, GPOMDP, SVRPG and SRVR-PG proposed in our paper are only policy gradient algorithms and do not use any critic in the update. In addition, A3C and A2C are specifically focused on parallel training while our algorithm is a general policy gradient method that can be modified to be applicable in different scenarios. Therefore, due to the fact that our algorithm does not use the actor-critic scheme (policy gradient plus critic approximation) and it is not designed for the parallel setting, our algorithm is not directly comparable to the algorithms (i.e., A3C, A2C) you pointed out.\", \"q3\": \"\\\"I am not sure if you are using the right word novel to describe your method ...\\\"\", \"a3\": \"Although the SARAH/SPIDER estimator has been well studied in the optimization field, its application to reinforcement learning problems is non-trivial. In particular, in reinforcement learning, the data distribution is changing from iterate to iterate, thus the vanilla SPIDER estimator is not directly applicable in our setting. To overcome this challenge, we use importance sampling techniques (See equation (3.2) and the discussion around it in the revision) to correct the data sampled from one distribution to another distribution. Moreover, our importance weighted estimator is also different from Papini et al. (2018) and Xu et al. (2019) since we use the step-wise importance weight which leads to better sample complexity. Due to the use of the step-wise importance weighting, our result does not depend on the horizon length $H$, which is one of our main contributions. So we think our algorithm design is novel.\"}", "{\"title\": \"Response to review #1 (Part 2)\", \"comment\": \"Q4: \\\"There are some discussion on choosing a batch-size in the experimental part. However, I do not see the discussion on the learning rate. The choice of parameters for GPOMDP and SVRPG may also need to be discussed. \\\"\", \"a4\": \"Thank you for your suggestion. According to Theorem 4.5, the learning rate $\\\\eta\\\\leq 1/(4L)$, which can be chosen as a constant. In practice, we set $\\\\eta = C_0/L$, where $L=MR/(1-\\\\gamma)^2$ is a constant and $C_0$ is a tuning parameter. In our experiment, for simplicity, we use grid search to directly tune $\\\\eta$. For instance, we searched $\\\\eta$ for the Cartpole problem by evenly dividing the interval $[10^{-5},10^{-1}]$ into $20$ points in the log-space. The parameters of GPOMDP and SVRPG are tuned in a similar way. That is, we used the suggested values in Papini et al. (2018) and Xu et al. (2019) and multiplied them with some constants found by grid search. The detailed parameters of all methods in different environments are presented in Tables 2 and 3 in Appendix E. We have rephrased the description in Section 5 and Appendix E in the revision.\", \"q5\": \"\\\"In Papini et al., 2018, for the experiment part, they use a snapshot policy to sample in the inner loop, and use early stopping inner loop ... Do you apply any trick to your experiments? I wonder if your numerical experiments are totally followed on your theory.\\\"\", \"a5\": \"Following the source code provided by Papini et al. (2018), we normalized the gradient to avoid the gradient blowup. But we did not use any other tricks such as checking the variance to recover backup policy and the early stopping inner loop (varying epoch length) as Papini et al. (2018) did. As we explained in previous response (A.4), we set the order of batch size parameters as the theorem and corollary suggested and we tuned a constant parameter by grid search. The learning rates and other parameters are also tuned by grid search. So our numerical experiments followed our theory except the normalized gradient which is also used in Papini et al. (2018).\\n\\nThank you for pointing out the typo. We have fixed it in the revision.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The authors propose a new stochastic reduced variance policy gradient estimator, which combines a baseline GPOMDP estimator with a control variate integrating past gradients by importance re-weighting. The authors establish the sample complexity of gradient descent using the proposed estimator, and further demonstrate its effectiveness through some simple empirical results.\\n\\nI believe this paper is a good contribution for ICLR. The result is relevant and interesting, and extends recent ideas around reduced-variance policy gradient estimators. The paper is overall easy to read, and presents its ideas clearly. Some detailed comments:\\n\\n- The wording of theorem 4.5 and corollary 4.7 could be somewhat clarified. In particular, I did not see \\\\Phi defined in the main text, and given its definition in the appendix, I believe theorem 4.5 could be stated simply in terms of J, avoiding any additional notation. Similarly, corollary 4.7 could be stated somewhat more clearly, and in particular, the choice of S should be made explicit. In the appendix, I could not find a definition for \\\\phi.\\n\\n- The empirical results presented are interesting, although I wish they were more comprehensive. In particular, it would be valuable to more exhaustively evaluate the impact of the hyper-parameters N, B and m. The authors should also clarify how the current values were chosen. Given that the theoretical results also apply to projected gradient descent, it would be interesting to see empirical results in that case.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper studies the theory of sample efficiency in reinforcement learning, which is of great importance and has a potentially large audience.\", \"the_strong_points_of_the_paper\": \"1. This paper proposed a new algorithm stochastic variance reduced policy gradient algorithms. This paper establishes better sample complexity compared with existing work. The key part of the proposed algorithm for variance reduction is to have step-wise importance weights to deal with the inconsistency caused by varying trajectory distribution. \\n2. This paper provides experimental results verifies the efficiency and effectiveness of the proposed algorithm. \\n3. In addition, parameter-based exploration extension is discussed in the appendix, which enjoys the same order of sample complexity under mild assumptions and gives better empirical performance. \\n4. This paper is easy to follow. In particular, there are a lot of discussions comparing this work with existing work.\", \"the_weak_points_of_the_paper\": \"1. In section 3, it is not quite clear how the reference policy is defined, and the \\\\theta^s is not clearly defined when s >= 1.\\n2. In the main part of the paper, the discussion in Remark 4.6 and the following Corollary 4.7 is not quite clear.\", \"some_minor_comments_of_the_paper\": \"1. In introduce page 3, We note that a recent work by .... by a fator of H. --> Here H should be defined as the Horizon. \\n2. There is one additional parenthesis in Theorem 4.5. \\n3. In Corollary 4.7, T is not defined.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary:\\n\\nThe paper proposed a policy gradient method called SRVR-PG, which based on stochastic recursive gradient estimator. It shows that the complexity is better than that of SVRPG. Some experiments on standard environments are provided to show the efficiency of the algorithm over GPOMDP and SVRPG.\", \"comments\": \"1) I think you may ignore the highly-related work: Yang and Zhang, \\\"Policy Optimization with Stochastic Mirror Descent\\\", June 2019. Since both papers are highly-related, I would suggest the author(s) have some discussions to differentiate two papers. \\n\\n2) Could you please provide the reasons why you are only choosing two methods (GPOMDP and SVRPG) to compare? There are many policy gradient algorithms such as A3C, A2C, ... which have not mentioned or discussed in this paper. Are they completely different here? Is there no way to compare the performance among them with SRVR-PG? \\n\\n3) I am not sure if you are using the right word \\\"novel\\\" to describe your method. Basically, you adopt the existing estimator based on SARAH/SPIDER in optimization algorithms into RL problems. I do not think the word \\\"novel\\\" is proper here since it is not something total new. Notice that, the complexity result achieved in this paper is also matched the one for SARAH/SPIDER/SpiderBoost in nonconvex optimization. \\n\\n4) There are some discussion on choosing a batch-size in the experimental part. However, I do not see the discussion on the learning rate. The choice of parameters for GPOMDP and SVRPG may also need to be discussed. \\n\\n5) In Papini et al., 2018, for the experiment part, they use a snapshot policy to sample in the inner loop, and use early stopping inner loop. Moreover, they also check variance to recover the backup policy when it is blowup. Do you apply any trick to your experiments? I wonder if your numerical experiments are totally followed on your theory.\", \"minor_comments\": [\"Redundancy \\\")\\\" in \\\\eta*S*m in Theorem 4\"]}" ] }

[ "Gregor N. C. Simm", "José Miguel Hernández-Lobato" ]

[ "graph neural networks", "variational autoencoders", "distance geometry", "molecular conformation" ]

{ "note_id": [ "JoWxpSOYU9", "H1gRN1LPiB", "rJe4IsrvsB", "H1eLIqBvsr", "rkgKMOSwjB", "HJgLgvHvjH", "S1l4hHBvjB", "SkeK8VrDsB", "rkeWJZSPjB", "HyxOGi_D5H", "HyxVxQy0FH", "SkgzPOBuYB", "HJxAaKHQ_B" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment" ], "note_created": [ 1576798730552, 1573506869851, 1573505868334, 1573505613976, 1573505040938, 1573504750371, 1573504427949, 1573504081043, 1573503193275, 1572469520330, 1571840748327, 1571473497606, 1570097605737 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1714/Authors" ], [ "ICLR.cc/2020/Conference/Paper1714/Authors" ], [ "ICLR.cc/2020/Conference/Paper1714/Authors" ], [ "ICLR.cc/2020/Conference/Paper1714/Authors" ], [ "ICLR.cc/2020/Conference/Paper1714/Authors" ], [ "ICLR.cc/2020/Conference/Paper1714/Authors" ], [ "ICLR.cc/2020/Conference/Paper1714/Authors" ], [ "ICLR.cc/2020/Conference/Paper1714/Authors" ], [ "ICLR.cc/2020/Conference/Paper1714/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1714/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1714/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1714/Authors" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper presents a solution to generating molecule with three dimensional structure by learning a low-dimensional manifold that preserves the geometry of local atomic neighborhoods based on Euclidean distance geometry.\\n\\nThe application is interesting and the proposed solution is reasonable. The authors did a good job at addressing most concerns raised in the reviews and updating the draft.\", \"two_main_concerns_were_left_unresolved\": \"one is the lack of novelty in the proposed model, and the other is that some arguments in the paper are not fully supported. The paper could benefit from one more round of revision before being ready for publication.\", \"title\": \"Paper Decision\"}", "{\"title\": \"General Response to Reviews\", \"comment\": [\"We would like to thank all the reviewers for their thorough and constructive feedback. Below, we list major changes we made to the paper based on the comments from all reviewers. In the responses to the reviewers, we address their individual comments.\", \"1. We improved the overall readability of the paper, especially for readers who are not familiar with the field of molecular modeling.\", \"We reworked the introduction to clarify the concept of a conformation.\", \"We replaced Fig. 1 in the introduction with a new figure that clearly shows how a molecule's conformations are related to its graph representation.\", \"2. In Section 2.1, we expanded on the explanation of the molecular graph and how we obtain the distances between atoms in a conformation.\", \"We now define all node and edge features formally.\", \"We modified Fig. 2 to illustrate how the set of distances $\\\\mathbf{d}$ are extracted from the conformation $\\\\mathbf{x}$.\", \"3. In Section 2.2, we significantly improved the clarity of the explanation of the generative model.\", \"With additional equations and implementation details, the reader is now guided through the model and the overall data flow step by step.\", \"We added a new figure (Fig. 3) that shows the different components of the variational autoencoder and how the graph data flows through the model.\", \"We added additional information on how message-passing neural networks take graphs of arbitrary size and shape as input.\", \"All hyperparameters are now reported in the Appendix (the code is available online).\", \"4. We simplified Section 2.4 to improve the readability and clarity of the proposed importance sampling scheme.\", \"5. We now propose a new benchmark (Conf17) instead of a new dataset.\", \"6. We ran additional experiments:\", \"We studied the effect of the latent space size and report our results.\", \"We ran simulations to provide information on the computational effort of all methods.\"]}", "{\"title\": \"Response to Review #3 - continued 3\", \"comment\": \"13. \\\"The paper should clarify the hierarchy of molecular configurations:\\nfor a given composition (%C, %O, %H, etc) (here, fixed throughout the paper at C7,H10,O2)\\nthere are several structures possible (graphs or isomers),\\nfor a given graph (isomer), there can be a couple of chiral variants (mirroring of (pieces of) the molecule,\\nfor each chiral variant, there are conformations, which correspond to rather weak variations of the Cartesian coordinates, due to thermal fluctuations. Some of these variations can be important, e.g. when a bond can freely rotate.\\nSome correspond to bending or torsion, etc, of the molecule.\\\"\\n\\nBased on this suggestion, we rewrote the relevant section in the introduction.\\nOver the last few years, many highly-effective deep learning methods generating small molecules with desired properties (e.g., novel drugs) have emerged. These methods operate using graph representations of molecules in which nodes and edges represent atoms and bonds, respectively. A representation that is closer to the physical system is one in which a molecule is described by its geometry or conformation. A conformation $\\\\mathbf{x}$ of a molecule is defined by a set of atoms $\\\\{ (\\\\epsilon_i, \\\\mathbf{r}_i)\\\\}_{i=1}^{N_v}$, where $N_v$ is the number of atoms in the molecule, $\\\\epsilon_i \\\\in \\\\{\\\\text{H}, \\\\text{C}, \\\\text{O},... \\\\}$ is the chemical element of the atom $i$, and $\\\\mathbf{r}_i \\\\in \\\\mathbb{R}^3$ is its position in Cartesian coordinates. Importantly, the relative positions of the atoms are restricted by the bonds in the molecule and the angles between them. Due to thermal fluctuations resulting in stretching of and rotations around bonds, there exist infinitely many conformations of a molecule. A molecule's graph representation and a set of its conformations are shown in Fig. 1.\\n\\nHowever, we would like to say that we don't explicitly state the relation between composition and conformations as it could confuse the reader. In addition, the concept of the composition is not strictly necessary to understand how our model works. Furthermore, the model is (as explained above) transferable to different stoichiometries. Nonetheless, and as mentioned before, we replaced Fig. 1 to clarify how conformations and structural formulae relate.\\n\\n\\n14. \\\"Add a comment on why (in your opinion) the distributions are not exactly correctly sampled, which after all is the goal (since the initial graph structure is given as input here).\\\"\\n\\nThe reason why the distributions are not sampled exactly is that we only report performance metrics on the test set, i.e., graphs the model has not seen before. Since the benchmark is challenging (e.g., it contains many constrained rings), our model does not always predict the right distances. Nonetheless, it does significantly better than previous methods (see, for example, Table 1).\\n\\nIn the manuscript, we now stress that the results are for graphs from the test set only.\\n\\\"For each train-test split and each molecule G in the test set, we compute the MMD of the joint distribution of distances between C and O atoms (H atoms are usually ignored)...\\\"\\n\\nFurther, in the caption of Table 1 it now says:\\n\\\"Reported are the results for molecules in the test set from three train-test splits.\\\"\\n\\n\\n15. \\\"In section 2.3: only a single molecule output is taken from each distance matrix.\\nCould you comment on the possibility to sample several Cartesian coordinates, for each distance matrix ? It seems like several solutions should (often) exist. Sampling continuously from that distribution would surely provide a more continuous (rich) output (am I correct)?\\\"\\n\\nIndeed, there exist (in general) multiple solutions for the same distance matrix. However, since the upper and lower bounds (d_min, d_max) are pretty close together, the different solutions are very similar. Therefore, we decided to take only one set of coordinates per set of bounds.\\n\\nIn the latest version of the manuscript, we now write in a footnote:\\n\\\"Often there exist multiple solutions for the same set of bounds. As the bounds are generally tight, the solutions are very similar. Therefore, we only generate one set of coordinates per set of bounds.\\\"\\n\\n\\nAgain, thank you very much for your extensive review. We would appreciate it if you could let us know if you are willing to accept the paper now. If not, could you please let us know what we need to clarify further.\"}", "{\"title\": \"Response to Review #3 - continued 2\", \"comment\": \"9. \\\"As there was no mention of speed of computation in the paper (although we can guess it is faster), this is unsupported. However it is easy to remedy, simply give an order of magnitude of sampling time (for thousands of conformations of 1 isomer).\\\"\\n\\nBased on this comment we ran some additional experiments. In the results section we now write:\\n\\\"For each molecule in the test set, we generated 50 conformations with each method. This took DL4Chem, RDKit, and GraphDG on average around hundreds of milliseconds per molecule. In contrast, a single conformation in the ISO17 dataset takes around a minute to compute.\\\"\\n\\nWe also state the hardware configuration of the computer these computations were carried out on.\\n\\\"All simulations were carried out on a computer equipped with an i7-3820 CPU and a GeForce GTX 1080 Ti GPU.\\\"\\n\\n\\n10. \\\"This suggests that the method would also work for dense packings. However this is not shown in the experiments.\\nProbably I misunderstood the claim, and it only refers to the ability to tackle general molecular compounds, not just linear ones? However proteins are not always linear, but can present cycles, and other branch-style patterns, so the claim is not very clear to me.\\\"\\n\\nWe thank the reviewer for reading our submission so carefully. We rephrased our statement to eliminate that confusion:\\n\\\"Our principled representation based on pair-wise distances does not restrict our approach to any particular molecular structure.\\\"\\n\\nWith this statement, we aimed to highlight that many previous approaches are heavily tailored towards proteins (as they rely on their [linear] amino acid sequence) and cannot be directly applied to molecules that often feature cycles and many branches.\\n\\n\\n11. \\\"There is no discussion of hyper-parameter tuning, or any detail of the intermediate GNN that constitute the encoder and decoder in the CVAE. So it was either not done at all, and then it should be mentioned; or it was done and it should be discussed quickly.\\\"\\n\\nIn the latest version of the manuscript, we detail how the individual hyperparameters were tuned with a Bayesian optimizer. In the Appendix, we list all hyperparameters and the values that were identified by the Bayesian optimization procedure.\\n\\n\\n12. \\\"In particular, in section 2.2 there is no discussion on the impact of the latent space' dimension, here set to Nv, i.e. 7+2+10=19 (correct?). I guess there is a balance between accuracy and generalizability, a short discussion on this point would be interesting.\\\"\\n\\nYes, that is correct, the latent space size is 19. We ran multiple experiments with latent space sizes of Nv, 2 Nv, and 3 Nv to study the effect of the latent space size\\non the model's performance. Our experiments showed that a latent space of 2 Nv performs slightly worse than Nv. 3 Nv results in overfitting.\", \"we_now_mention_our_results_and_justify_our_choice\": \"\\\"A conformation has, in general, $3N_v - 6$ spatial degrees of freedom (dofs): one dof per spacial dimension per atom minus three translational and three rotational dofs. Therefore, the latent space should be proportional to the number of atoms in the molecule. In addition, the latent space should be smaller than $3N_v$ as it is the role of the encoder to project the conformation into a lower-dimensional space.\\nAs a result, we set $k=1$ to avoid overfitting.\\\"\\n\\nIn a footnote, we then write:\\n\\\"Experiments showed that our model performs similarly with a latent space of R^{2N_v} and overfits with latent spaces of R^{3N_v}.\\\"\"}", "{\"title\": \"Response to Review #3 - continued\", \"comment\": \"3. \\\"However sometimes the presentation rushes a bit towards results, in the sense that it is not easy to figure out over which ensemble the mean/median/max/min has been computed.\\\"\\n\\nWe expanded on the experimental setup in Section 5.1 to address this issue and hope that it is clear now how we arrived at our results.\\n\\n4. \\\"A deeper comment: in the field of conformation sampling, there are two issues. One is to correctly sample states in each local minimum of the free energy. This is addressed in the paper. The second issue is to hunt for all possible local minima of the isomer, i.e. sample even unlikely but crucial conformations. This is not addressed here. ...\\\"\\n\\nWe highly appreciate your general comment on this topic. We agree that hunting for local minima is an outstanding challenge in conformational sampling. While many advanced techniques (e.g., umbrella sampling) exist, none of them can assure that all local minima have been sampled sufficiently for complex molecules. For larger molecules, it is practically impossible to guarantee that all local minima have been visited at least once. All methods we compare our method with, including RDKit and the simulation used to generate the ISO17 dataset, suffer from this shortcoming.\\n\\nHowever, the aim of this method is not to only generate conformations in local minima but to sample from the distribution $p(\\\\mathbf{x} | \\\\mathcal{G})$ in a one-shot fashion. While minimum-energy conformations will be sampled more frequently than out-of-equilibrium conformations our model is not restricted to those. Similarly, an MD or MCMC simulation does not only visit minimum energy conformations. However, we agree that minimum-energy conformations strongly contribute to average properties (e.g., dipole moment). Nonetheless, as can be seen from the results in Table 2, even highly optimized algorithms like the one in RDKit struggle to correctly sample local minima of highly constrained molecules.\\n\\nTo clarify this matter, we wrote in the benchmark section:\\n\\\"Most molecules feature highly-strained, complex 3D structures such as rings which are typical of drug-like molecules. It is thus the structural complexity of the molecules, not their number of degrees of freedom, that makes this benchmark challenging.\\\"\\n\\nFurther, we now address that issue by writing in the Conclusions:\\n\\\"Further, methods have to be devised to estimate how many conformations need to be generated to ensure all important conformations have been sampled.\\\"\\n\\nFinally, we stress the important distinction between conformers (minimum-energy structures) and conformations:\\n\\\"Datasets such as the one published by Kanal et al. only include conformers, i.e., the most stable conformations of a molecule, and not a distribution over conformations.\\\"\\n\\n\\n5. \\\"However, the experiments have been performed only on isomers of C7,H10,O2.\\nFurthermore, there is no discussion on the computational cost of the method at all, so its extendibility to larger shapes is not obvious.\\\"\", \"we_clarify_in_the_model_definition_why_our_method_can_be_applied_to_molecules_of_arbitrary_shape\": \"\\\"At the center of this model are message-passing neural networks (MPNNs) (Gilmer 2017) with multi-head attention (Velickovic 2018). In short, an MPNN is a convolutional neural network that allows end-to-end learning of prediction pipelines whose inputs are graphs of arbitrary size and shape. In a convolution, neighboring nodes exchange so-called messages between neighbors to update their attributes. Edges update their attributes with the features of the nodes they are connecting. It is a well-studied technique that achieves state-of-the-art performance in representation learning for molecules.\\\"\\n\\nWhile it is true that the benchmark only comprises molecules with the C7H10O2 stoichiometry, MPNNs have been reported to work very well for multiple regression task on different-sized molecules in the original publication by Duvenaud et al. Therefore, we have reasons to believe that this should hold for this task as well.\\n\\n\\n6. We addressed your remarks regarding claim 3 (the dataset / benchmark) in our answer Nr. 2.\\n\\n\\n7. \\\"However the paper's methodology doesn't address the issue of ''hunting for new local minima''.\\n\\nIn response Nr. 4, we discuss this issue of local minima.\\n\\n\\n8. \\\"The data set could be enriched also with a labeling of ''which conformation is in which local minimum of free energy''.\\\"\\n\\nThe ISO17 dataset was generated by an MD simulation at 500 K. Therefore, it is a difficult task to assign a snapshot of this simulation unambiguously to a particular minimum-energy conformation, especially when one considers the simulation temperature.\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"We would like to thank the reviewer for this extensive review and helpful feedback. We have incorporated all the reviewer's remarks and suggestions into the latest version of the manuscript. We also followed his advice on how to further improve the paper by running additional experiments.\", \"general_comments\": \"================\\n\\nWe would like to thank all the reviewers for their thorough and constructive feedback.\", \"we_made_the_following_major_changes_to_the_paper_based_on_the_comments_from_all_reviewers\": [\"1. We improved the overall readability of the paper, especially for readers who are not familiar with the field of molecular modeling.\", \"We reworked the introduction to clarify the concept of a conformation.\", \"We replaced Fig. 1 in the introduction with a new figure that clearly shows how a molecule's conformations are related to its graph representation.\", \"2. In Section 2.1, we expanded on the explanation of the molecular graph and the distances between atoms in a conformation.\", \"We now define all node and edge features formally.\", \"We modified Fig. 2 to illustrate how the set of distances $\\\\mathbf{d}$ are extracted from the conformation $\\\\mathbf{x}$.\", \"3. In Section 2.2, we significantly improved the clarity of the explanation of the generative model.\", \"With additional equations and implementation details, the reader is now guided through the model and the overall data flow step by step.\", \"We added a new figure (Fig. 3) that shows the different components of the variational autoencoder and how the graph data flows through the model.\", \"We added additional information on how message-passing neural networks take graphs of arbitrary size and shape as input.\", \"All hyperparameters are now reported in the Appendix.\", \"4. We simplified Section 2.4 on the proposed importance sampling scheme to improve readability and clarity.\", \"5. We now propose a new benchmark (Conf17) instead of a new dataset.\", \"6. We ran additional experiments:\", \"We studied the effect of the latent space size and report our results.\", \"We ran simulations to add information on the computational effort of all methods.\"], \"individual_response\": \"====================\\n\\n1. \\\"The precise combination of architectures is explained in section 2.2. This explanation is a little fast for me (but I am new to this field). In that section, it could be nice to precise that the job of the CVAE is to learn (among others) the parameters mu,sigma of q_phi(z|d,G), using a batch of conformations of a given isomer (and then generalize to other isomers, each time guessing the correct mu,sigma parameters, even for a new isomer). This is not obvious for the inexperienced reader (I hope I \\\"guessed\\\" correctly).\\\"\\n\\nWe acknowledge that the model definition in Section 2 was not detailed enough, and therefore, completely reworked this section. In the latest version of the manuscript, we explicitly state the role (and dimensionality) of every function and variable that appears in the implementation of the CVAE. Further, we added equations to define more formally every part of the model. Further, we replaced Fig. 2 B) with an illustration in which we show how the distances are extracted from the conformations. Finally, we added another figure (Fig. 3) that shows exactly how the graph data is flowing through the model. We hope that it is more apparent now how the CVAE learns parameters mu and sigma of q_phi(z|d,G).\\n\\n\\n2. \\\"Indeed, their sharing of this augmented version of ISO17 is absolutely necessary: I would simply put it differently, stating from the beginning that they share an augmented version of a pre-existing data set.\\\"\\n\\nWe thank you for this constructive feedback on this matter. We agree with you and now stress that this is not a new dataset but a benchmark using an existing dataset and rephrased this matter throughout our manuscript.\", \"in_the_introduction_of_the_latest_version_of_the_paper_it_now_reads\": \"\\\"We create a new, challenging benchmark for conformation generation, which is made publicly available.\\\"\\n\\nIn the spirit of reproducibility and since arriving at the data required for running this benchmark is non-trivial, we decided to publish the data we employed in this study. Further, we updated the presentation of this benchmark in Section 4 accordingly to reflect that.\"}", "{\"title\": \"Response to Review #2 - continued\", \"comment\": \"4. \\\"I think that the definition of the model in Sec 2.2. could have been better. As before, I believe that it is hard to understand explicitly what the functions input and output is, since the dimensionalities are not defined. In my opinion, it would have been better if instead of this high level text you could provide explicitly all the steps of the model e.g. the concatenations, the appends, etc.\\\"\\n\\nWe rewrote the section on the model definition (Section 2.2) to improve readability and precision. We now give details on the individual functions, their inputs, the dimensions of the output, and trainable parameters. As stated above, we added Fig. 3, a figure that captures the details of our model and the general data flow. Further, we now provide all the hyperparameters of the model in the Appendix.\\n\\n\\n5. \\\"I could not understand if and how the model is able to handle graphs with different number of nodes, and different number of distances among them.\\\"\\n\\nThe model can handle graphs with different numbers of nodes and edges because we are employing message passing graph neural networks (MPNN). In an MPNN, neighboring nodes exchange so-called messages over the edges connecting them to update their attributes. This network architecture is (in theory) applicable to any graph size or shape and has been applied to molecular graphs many times in the recent literature.\", \"see_for_example\": \"Duvenaud, D. K.; Maclaurin, D.; Iparraguirre, J.; Bombarell, R.; Hirzel, T.; Aspuru-Guzik, A.; Adams, R. P. \\\"Convolutional Networks on Graphs for Learning Molecular Fingerprints\\\". In Advances in Neural Information Processing Systems 28; 2015; pp 2224\\u20132232.\\n\\nIn addition to the many changes we made to Section 2, we now also give a short overview of what an MPNN is:\\n\\\"In short, an MPNN is a convolutional neural network that allows end-to-end learning of prediction pipelines whose inputs are graphs of arbitrary size and shape. In a convolution, neighboring nodes exchange so-called messages between neighbors to update their attributes. Edges update their attributes with the features of the nodes they are connecting. The MPNN is a well-studied technique that achieves state-of-the-art performance in representation learning for molecules.\\\"\\n\\n\\nWe have updated our paper according to your review. We would appreciate it if you could let us know if you are willing to accept the paper now, and if not, let us know what we need to address/clarify further.\\nThank you very much.\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"We thank you for your extensive review, list of comments, and constructive feedback. We have incorporated all your remarks in the latest revision of the manuscript.\", \"general_response\": \"================\", \"we_made_the_following_major_changes_to_the_paper_based_on_the_comments_from_all_reviewers\": [\"1. We improved the overall readability of the paper, especially for readers who are not familiar with the field of molecular modeling.\", \"We reworked the introduction to clarify the concept of a conformation.\", \"We replaced Fig. 1 in the introduction with a new figure that clearly shows how a molecule's conformations are related to its graph representation.\", \"2. In Section 2.1, we expanded on the explanation of the molecular graph and the distances between atoms in a conformation.\", \"We now define all node and edge features formally.\", \"We modified Fig. 2 to illustrate how the set of distances $\\\\mathbf{d}$ are extracted from the conformation $\\\\mathbf{x}$.\", \"3. In Section 2.2, we significantly improved the clarity of the explanation of the generative model.\", \"With additional equations and implementation details, the reader is now guided through the model and the overall data flow step by step.\", \"We added a new figure (Fig. 3) that shows the different components of the variational autoencoder and how the graph data flows through the model.\", \"We added additional information on how message-passing neural networks take graphs of arbitrary size and shape as input.\", \"All hyperparameters are now reported in the Appendix.\", \"4. We simplified Section 2.4 on the proposed importance sampling scheme to improve readability and clarity.\", \"5. We now propose a new benchmark (Conf17) instead of a new dataset.\", \"6. We ran additional experiments:\", \"We studied the effect of the latent space size and report our results.\", \"We ran simulations to add information on the computational effort of all methods.\"], \"individual_response\": \"==================\\n\\n1. \\\"I think that authors should specify the dimensionalities of each variable and function in the paper. Unfortunately, none of them is clearly defined, which makes the technical part a bit hard to follow.\\\"\\n\\nWe recognize that the first revision lacked precision in the introduction and Section 2. In the latest version, we now define all variables and functions (including the dimensions) explicitly. Please, see our general remarks and Section 2 in the manuscript.\\n\\n\\n2. \\\"As a non-expert of the filed, I would like to know what exactly is a conformation x? Does it refer to the Euclidean coordinates of each node (atom) v of the graph?\\\"\\n\\nIn the introduction, we now define a conformation. We also replaced Fig. 1 with a new figure to help illustrate this difficult concept, especially for readers who are not familiar with chemical modeling.\\n\\nIn the introduction, we now write:\\n\\\"Over the last few years, many highly-effective deep learning methods generating small molecules with desired properties (e.g., novel drugs) have emerged. These methods operate using graph representations of molecules in which nodes and edges represent atoms and bonds, respectively. A representation that is closer to the physical system is one in which a molecule is described by its geometry or conformation. A conformation $\\\\mathbf{x}$ of a molecule is defined by a set of atoms $\\\\{ (\\\\epsilon_i, \\\\mathbf{r}_i)\\\\}_{i=1}^{N_v}$, where $N_v$ is the number of atoms in the molecule, $\\\\epsilon_i \\\\in \\\\{\\\\text{H}, \\\\text{C}, \\\\text{O},... \\\\}$ is the chemical element of the atom $i$, and $\\\\mathbf{r}_i \\\\in \\\\mathbb{R}^3$ is its position in Cartesian coordinates. Importantly, the relative positions of the atoms are restricted by the bonds in the molecule and the angles between them. Due to thermal fluctuations resulting in stretching of and rotations around bonds, there exist infinitely many conformations of a molecule. A molecule's graph representation and a set of its conformations are shown in Fig. 1.\\\"\\n\\n\\n3. \\\"For each G, there are multiple possible conformations x_i, where each x_i specifies the Euclidean coordinates of all the nodes (atoms) v_j? Also, each node v_j has some attributes? So the goal is to represent each conformation x_i -implicitly- as a set of distances between the nodes v_j? I think that Sec 2.1. can be more clear about the definition of the data, such that to be accessible from non experts in molecule modeling. (Related to comment #1.)\\\"\\n\\nWe made significant changes to Section 2.1, including clear statements about the dimensionality of the node and edge features. Further, we believe that the new Fig. 3 will help the reader understand how the graph data flows through the model. Finally, you are right in thinking that nodes (which represent atoms) have attributes $v_i$ to encode the atomic element. Further, it is correct that it is our goal to model the distribution over distances between the atoms in the molecule.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"We thank you for your review and comments. We have incorporated all your suggestions in the latest revision of the manuscript.\", \"general_response\": \"================\", \"we_made_the_following_major_changes_to_the_paper_based_on_the_comments_from_all_reviewers\": [\"1. We improved the overall readability of the paper, especially for readers who are not familiar with the field of molecular modeling.\", \"We reworked the introduction to clarify the concept of a conformation.\", \"We replaced Fig. 1 in the introduction with a new figure that clearly shows how a molecule's conformations are related to its graph representation.\", \"2. In Section 2.1, we expanded on the explanation of the molecular graph and the distances between atoms in a conformation.\", \"We now define all node and edge features formally.\", \"We modified Fig. 2 to illustrate how the set of distances d are extracted from the conformation x.\", \"3. In Section 2.2, we significantly improved on the clarity of the explanation of the generative model.\", \"With additional equations and implementation details, the reader is now guided through the model and the overall data flow step by step.\", \"We added a new figure (Fig. 3) that shows the different components of the variational autoencoder and how the graph data flows through the model.\", \"We added additional information on how message-passing neural networks take graphs of arbitrary size and shape as input.\", \"All hyperparameters are now reported in the Appendix.\", \"4. We simplified Section 2.4 on the proposed importance sampling scheme to improve readability and clarity.\", \"5. We now propose a new benchmark (Conf17) instead of a new dataset.\", \"6. We ran additional experiments:\", \"We studied the effect of the latent space size and report our results.\", \"We added ran simulations to add information on the computational effort of all methods.\"], \"individual_response\": \"==================\\n\\nI. \\\"What is the variable \\\"x\\\"? Is it possible to define a bit formally? It appears that x=(G,d) i.e. the graph with d.\\\"\\n\\nIn the introduction, we now clearly define what a conformation x is:\\n\\\"Over the last few years, many highly-effective deep learning methods generating small molecules with desired properties (e.g., novel drugs) have emerged. These methods operate using graph representations of molecules in which nodes and edges represent atoms and bonds, respectively. A representation that is closer to the physical system is one in which a molecule is described by its geometry or conformation. A conformation $\\\\mathbf{x}$ of a molecule is defined by a set of atoms $\\\\{ (\\\\epsilon_i, \\\\mathbf{r}_i)\\\\}_{i=1}^{N_v}$, where $N_v$ is the number of atoms in the molecule, $\\\\epsilon_i \\\\in \\\\{\\\\text{H}, \\\\text{C}, \\\\text{O},... \\\\}$ is the chemical element of the atom $i$, and $\\\\mathbf{r}_i \\\\in \\\\mathbb{R}^3$ is its position in Cartesian coordinates. Importantly, the relative positions of the atoms are restricted by the bonds in the molecule and the angles between them. Due to thermal fluctuations resulting in stretching of and rotations around bonds, there exist infinitely many conformations of a molecule. A molecule's graph representation and a set of its conformations are shown in Fig. 1.\\\"\\n\\nWe also replaced Fig. 1 to illustrate this concept. Further, it is indeed possible to convert a graph G with the distances d back to a conformation x, as you assumed in your comment. This is exactly what our generative model does.\\n\\n\\nII. \\\"It would be nice to develop the generative model of (G,d) i.e. both G and d. Simply generating d looks a bit restrictive.\\\"\\n\\nIt would be interesting indeed to construct a predictive model for both the graph G and the set of distances between atoms d. However, our model addresses a different question: Given a molecule G, what is the distribution over conformations p(x | G)? This is a pressing question in drug design, for example. Once a candidate drug is identified, properties need to be predicted for this molecule. This requires drawing samples from p(x | G) which is notoriously difficult. Our method attempts to address this very issue by providing a fast method that generates conformational samples in a one-shot fashion. We updated our introduction to clarify this.\\n\\n\\nIII. \\\"How do you model/learn \\\\mathcal{O}(x)?\\\"\\n\\nIn this study, we do not learn the function $\\\\mathcal{O}(x)$ but employ a computational method based on quantum mechanics (density functional theory). The same applies to the function E(x). \\n\\nIn the latest revision of the paper, we now state this explicitly in Sections 2.4 and 5.3:\\n\\\"... and computing $\\\\mathcal{O}(x_i) \\\\in R$ with a quantum-chemical method (e.g., density functional theory).\\\"\\n\\n\\nWe would appreciate it if you could let us know if you have any other concerns or topics you would like us to address/clarify further.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper proposes a generative model for generating molecule with three dimensional structure. Given a graph, it leverages a variational autoencoder to embed the distance between two atoms into latent vectors (encoder) and then generate distance between two atoms based on the latent vector. Even if the problem is well motivated, I have several concerns:\\nI. What is the variable \\\"x\\\"? Is it possible to define a bit formally? It appears that x=(G,d) i.e. the graph with d.\\nII. It would be nice to develop the generative model of (G,d) i.e. both G and d. Simply generating d looks a bit restrictive. \\nIII. How do you model/learn \\\\mathcal{O}(x)?\\n\\nI would be curious to know the answer of these questions.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"Summary:\\nThe authors propose a generative model designed for molecules, which is essentially a conditional variational auto-encoder. The model learns to generate, when conditioning on a molecule graph, the distribution of distances between each of the atoms and its second and third neighbour. Finally, using these information new molecules can be generated, which satisfy the generated distances. In the experiments the authors show the effectiveness of the method.\", \"general_comments\": \"The paper is ok written, but as a non-expert of the field, I find it quite hard to follow. More specifically, I think that the main content aims to readers which are very familiar with the problem of molecule modeling, which is of course understandable. However, I find the technical (machine learning) content a bit unclear, because I think that the proposed method is not presented properly in details (see comments):\\n\\n1. I think that authors should specify the dimensionalities of each variable and function in the paper. Unfortunately, none of them is clearly defined, which makes the technical part a bit hard to follow.\\n\\n2. As a non-expert of the filed, I would like to know what exactly is a conformation x? Does it refer to the Euclidean coordinates of each node (atom) v of the graph? (Related to comment #1.)\\n\\n3. For each G, there are multiple possible conformations x_i, where each x_i specifies the Euclidean coordinates of all the nodes (atoms) v_j? Also, each node v_j has some attributes? So the goal is to represent each conformation x_i -implicitly- as a set of distances between the nodes v_j? I think that Sec 2.1. can be more clear about the definition of the data, such that to be accessible from non experts in molecule modeling. (Related to comment #1.)\\n\\n4. I think that the definition of the model in Sec 2.2. could have been better. As before, I believe that it is hard to understand explicitly what the functions input and output is, since the dimensionalities are not defined. In my opinion, it would have been better if instead of this high level text you could provide explicitly all the steps of the model e.g. the concatenations, the appends, etc.\\n\\n5. I could not understand if and how the model is able to handle graphs with different number of nodes, and different number of distances among them.\\n\\n6. The experimental section seems solid enough, and shows that the proposed approach works better than the other methods. Also, I like the fact that some of the limitations of the method are stated. Moreover, the related work seems to be properly included. However, since I am not an expert of the field, I am not able to provide precise feedback for these two parts.\\n\\nIn general, I think that the proposed model solves good enough the problem that is designed for. Also, in my opinion, the manuscript aims for specialized readers, which is definitely understandable. However, my main concern regarding the current version, is that the technical (machine learning) content is a bit unclear and probably too high level. In case the technical content was presented properly, I think that it would have been a much better fit for this community.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I do not know much about this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #3\", \"review\": \"In this paper, a new method for estimating the equilibrium measure of molecular compounds is proposed and tested on a new (enriched) version of an existing data set (CONF17 is derived from the pre-existing ISO17, itself extracted from QM9). \\nThe method takes as input a collection of graph structures of a given stoichiometry mixture, and learns the corresponding conformations.\\nThis results in a generative model, taking as input a graph structure (or actually a single conformation instance, thus setting the stoichiometry AND the isomer/graph structure), to generate (in a one-shot fashion) other conformations of that isomer.\\nThe distribution of conformations sampled in this way compare well with those of an extensive training set (CONF17, set of isomers and conformations of a single compound), and in particular they perform better than 2 alternative methods published recently (2015 and 2019, the former performing globally much better than the latter!).\\n\\n\\nConcretely, the model/method/algorithm consists in a combination of existing network architectures. The main original contribution is to map the Cartesian coordinates of atoms into a pairwise distance space, which is naturally translationally and rotationally invariant. Some of the second and third neighbors distances in the graph structure are considered, such that the conformation is completely set (this would not be the case if considering only 1st nearest neighbors distances).\\nThe precise combination of architectures is explained in section 2.2. This explanation is a little fast for me (but I am new to this field). \\nIn that section, it could be nice to precise that the job of the CVAE is to learn (among others) the parameters mu,sigma of q_phi(z|d,G), using a batch of conformations of a given isomer (and then generalize to other isomers, each time guessing the correct mu,sigma parameters, even for a new isomer).\\nThis is not obvious for the inexperienced reader (I hope I \\\"guessed\\\" correctly).\\n\\nThe \\\"new dataset\\\" shared in this work is indeed useful for reproducing results and for later comparison with other works which aim at sampling conformations.\\nHowever, it derives directly from ISO17, which is itself an extraction from QM9. ISO17 lists the Cartesian coordinates of 127 isomers of C7,H10,O2 (each having ~3000 conformations). The \\\"new\\\" CONF17 is novel only in that it provides these conformations in distance-matrix space, i.e. it sets which pairs of atoms need to have their distances computed. This is very useful because this choice of pairs is not unique, and so setting it once is necessary for comparison. I would not stress too much that this is a new dataset: at first it actually sounds like authors chose a dataset such that their method outperforms others'! Which turns out not to be the case (they chose C7,H10,O2, but this doesn't look like fine-tuning). Indeed, their sharing of this augmented version of ISO17 is absolutely necessary: I would simply put it differently, stating from the beginning that they share an augmented version of a pre-existing data set.\\n\\nAbout the experimental protocol/methodology:\\nthe work carried out seems quite impressive. Indeed, the paper presents a rather extensive comparison of the new model with 2 other ones, on a large data set.\\nHowever sometimes the presentation rushes a bit towards results, in the sense that it is not easy to figure out over which ensemble the mean/median/max/min has been computed. This is related to the intrinsic complexity of estimating accuracy, given the encapsulation (compound-> isomer-> conformations-> each bond-> distribution over bond properties). I think however it could be made clearer, and this would benefit the paper greatly.\", \"a_deeper_comment\": \"in the field of conformation sampling, there are two issues. One is to correctly sample states in each local minimum of the free energy. This is addressed in the paper. The second issue is to hunt for all possible local minima of the isomer, i.e. sample even unlikely but crucial conformations. This is not addressed here. For computing average energy, dipole moment etc, the most likely conformations will dominate, most of the time. However given the topic I think it would be fair to mention this point, and explain why in this specific case, the second issue is not crucial.\\nBecause of this, also, I think \\n\\nOverall, the paper is well written and is very well situated within the literature (including the very recent literature).\\nIt states clearly its contributions and results, and provides detailed presentation of experimental results.\\nI think it contributes to the field and provides an interesting new way of dealing with the problem of molecular conformations sampling. The results obtained are much better than those of the compared ML method (RDkit).\\nHowever, some claims (in introduction or in section 3) are stated more strongly than they are supported by direct experimental evidence.\\n- ''... the first work on sampling molecular conformations for molecules of arbitrary size and shape based on deep learning.''\\nHowever, the experiments have been performed only on isomers of C7,H10,O2.\\nFurthermore, there is no discussion on the computational cost of the method at all, so its extendibility to larger shapes is not obvious.\\n- '' We create a new, challenging benchmark dataset CONF 17 for conformation generation, which is made publicly available.''\\nI made remarks on this claim above.'' \\n- '' We develop a rigorous experimental approach for evaluating and comparing the accuracy of conformation generation methods based on the mean maximum deviation distance metric.''\\nHowever the paper's methodology doesn't address the issue of ''hunting for new local minima''. \\nThe data set could be enriched also with a labeling of ''which conformation is in which local minimum of free energy''\\n- ''It is a fast alternative to resource-intensive approaches based on MCMC or MD.''\\nAs there was no mention of speed of computation in the paper (although we can guess it is faster), this is unsupported. However it is easy to remedy, simply give an order of magnitude of sampling time (for thousands of conformations of 1 isomer).\\n- ''Our principled representation based on pair-wise distances does not restrict our approach to any type of system (e.g., proteins) or any particular graph structure. In addition, it is extendable in a systematic fashion.''\\nThis suggests that the method would also work for dense packings. However this is not shown in the experiments.\\nProbably I misunderstood the claim, and it only refers to the ability to tackle general molecular compounds, not just linear ones? However proteins are not always linear, but can present cycles, and other branch-style patterns, so the claim is not very clear to me.\\n\\nBecause some of the important claims are only weakly supported, and because of the relatively small novelty of the method, I lean on rejecting the paper.\\n\\n\\n\\nAdditional comments (for improving the paper):\\n\\nThere is no discussion of hyper-parameter tuning, or any detail of the intermediate GNN that constitute the encoder and decoder in the CVAE. \\nSo it was either not done at all, and then it should be mentioned; or it was done and it should be discussed quickly.\\nIn particular, in section 2.2 there is no discussion on the impact of the latent space' dimension, here set to Nv, i.e. 7+2+10=19 (correct?). I guess there is a balance between accuracy and generalizability, a short discussion on this point would be interesting.\", \"the_paper_should_clarify_the_hierarchy_of_molecular_configurations\": \"for a given composition (%C, %O, %H, etc) (here, fixed throughout the paper at C7,H10,O2)\\nthere are several structures possible (graphs or isomers),\\nfor a given graph (isomer), there can be a couple of chiral variants (mirroring of (pieces of) the molecule,\\nfor each chiral variant, there are conformations, which correspond to rather weak variations of the Cartesian coordinates, due to thermal fluctuations. Some of these variations can be important, e.g. when a bond can freely rotate. Some correspond to bending or torsion, etc, of the molecule.\\nThe algorithm presented in the paper attempts to sample appropriately the equilibrium distribution of conformations for each given graph structure (at fixed chirality, if I understood correctly).\\n\\nAdd a comment on why (in your opinion) the distributions are not exactly correctly sampled, which after all is the goal (since the initial graph structure is given as input here).\\n\\n\\nAdditional question, for curiosity. In section 2.3: only a single molecule output is taken from each distance matrix.\\nCould you comment on the possibility to sample several Cartesian coordinates, for each distance matrix ? It seems like several solutions should (often) exist. \\nSampling continuously from that distribution would surely provide a more continuous (rich) output (am I correct) ?\"}", "{\"comment\": \"In the introduction, a sentence on existing methods based on statistical learning is not fully precise. The statement that they are not targeting a distribution over configurations but only the most stable folded configuration is not true for all of the methods. We will correct that at a later stage.\", \"title\": \"Sentence in introduction on existing machine learning methods needs to be more precise\"}" ] }

[ "Steven Hansen", "Will Dabney", "Andre Barreto", "David Warde-Farley", "Tom Van de Wiele", "Volodymyr Mnih" ]

[ "Reinforcement Learning", "Variational Intrinsic Control", "Successor Features" ]

{ "note_id": [ "NRPZobaUUg", "Skg5-FY2iH", "r1lY0dtnjH", "Hyl1FOK3sr", "B1lpxwFhoB", "SyekDcTg9B", "BJx2exIRFS", "rygRKCvOtr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730522, 1573849345862, 1573849296782, 1573849206598, 1573848821186, 1572031063419, 1571868659700, 1571483269741 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1711/Authors" ], [ "ICLR.cc/2020/Conference/Paper1711/Authors" ], [ "ICLR.cc/2020/Conference/Paper1711/Authors" ], [ "ICLR.cc/2020/Conference/Paper1711/Authors" ], [ "ICLR.cc/2020/Conference/Paper1711/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1711/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1711/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Talk)\", \"comment\": \"This work uses a variational autoencoder-based approach to combine the benefits of recent methods that learn policies with behavioral diversity with the advantages of successor representations, addressing the generalization and slow inference problems of competing methods such as DIAYN. After discussion of the author rebuttal, the reviewers all agreed on the significant contribution of the paper and that concerns about clarity were sufficiently addressed. Thus, I recommend this paper for acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Review Response\", \"comment\": \"Thank you for the lucid summary of our work and the well-reasoned review. We agree that Eq 10 is a good representation of our core contribution, and are glad that our empirical validation was satisfactory. You are correct that no fine-tuning was performed in these experiments, but we are confident that this could further boost our results, and is thus a promising avenue of future work.\"}", "{\"title\": \"Review Response\", \"comment\": \"Thank you very much for the kind review. We are glad you believe our work represents a valuable contribution to the field and appreciate your framing of VISR in terms of providing a grounded feature space.\"}", "{\"title\": \"Review Response\", \"comment\": \"Thank you for your considerate review and detailed comments/questions. We agree that clarity was a weak point for this work, so we have updated both the main text and the appendix to rectify this. Additionally, we\\u2019ve created a self-contained python notebook that implements the full VISR algorithm in a simplified setting. We are currently cleaning up that code with an eye on didactic utility, and promise to release it as soon as possible, and well before publication of the work.\", \"q\": \"include experiments in simple and easy to understand synthetic environments\", \"a\": \"We\\u2019ve included additional qualitative results in the appendix that train VISR on a simple grid-world environment. These include new figures that illustrate the features learned by VISR, as well as a representative sample of learned reward functions and the corresponding value functions. We will release a python notebook implementing the VISR algorithm and replicating this grid-world experiments upon publication.\"}", "{\"title\": \"Revision summary and general response to all reviewers\", \"comment\": \"Thank you all for your thoughtful reviews. We have done our best to address all your concerns in the revised version of the paper. One of the main concerns in the reviews was the clarity of the paper. To address this issue, we have carefully revised the paper and added more detailed explanations to several passages of the text, with particular emphasis on the mathematical derivations (see detailed comments to reviewer number 1, referred to as R1). In addition, as suggested by R1, we included in the appendix a simple experiment to provide intuition on the mechanics of the proposed method. Finally, we have tried to clarify diction and also made all the corrections suggested by the reviewers.\\n\\nAdditionally, (as mentioned in the previous version of our appendix) two of our random-feature baselines (RF-VISR and GPI-RF-VISR) had their performance metric calculated incorrectly, using on-policy data rather than the fixed data collection regime used on the other task-inference baselines. This has now been corrected, with all affected results updated. We emphasize that this has had no impact on any of the conclusions of the paper; we fixed it in the name of experimental rigor. \\n\\nLastly, we will be open-sourcing a self-contained implementation of the VISR algorithm in a python notebook in time for publication. This is being explicitly designed for didactic purposes, with a simplified version of our experimental setup that allows one to get a version of VISR trained up in under an hour, making it easy to test our task-inference procedure.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary:\\nThis paper proposes an algorithm to combine the ideas of unsupervised skill/option discovery (Eysenbach et. al., 2018, Gregor et. al., 2016, referred to as \\u201cBMI\\u201d in the paper) with successor features \\u201cSFs\\u201d (Barreto et. al., 2017, 2018). While unsupervised skill/option discovery algorithms employ mutual information maximization of visited states and the latent variables corresponding to options (typically discrete), this paper adds a restriction that this latent variable (now continuous) should be the task vector specified by some learnt successor features.\\n\\nWith such a restriction, the algorithm can now be used in an unsupervised pre-training stage to learn conditional policies corresponding to several different task vectors and can be used to directly infer (without training or fine-tuning) a good policy for a supervised phase where external reward is present (i.e. via GPI from Barreto et. al., 2018) by simply regressing to the best task vector.\\n\\nSuch unsupervised pre-training is shown to outperform DIAYN (Eysenbach et. al., 2018) in 3 different Atari suites (including the full 57 game suite) and also ablations to the proposed model where GPI and SFs are excluded individually.\", \"decision\": \"I vote for accept as this paper proposes a novel technique to combine mutual information based intrinsic control objectives with successor features, which allow for combining the benefits of both in a complementary way. An unsupervised phase can now discover good conditional policies with successor features which can be used to infer a good policy to solve an external reward task in a supervised phase, with such a policy capable of attaining human level performance in several Atari games and outperforming several baselines such as DQNs in limited data regimes.\", \"other_comments\": [\"The technique for enforcing the restriction in Eq. 10, as well as being able to use it with generalized policy improvement is a good novel contribution in the paper.\", \"The detailed comparison with baselines on the full Atari suite is sufficient to back the claims in the paper that the strengths of BMI and SFs do complement each other.\", \"The fact that fast task inference is sufficient to get good performance is impressive i.e. without the need to fine-tune the best inferred policy.\"], \"minor_typos\": [\"In section 5 para 5, \\u201cUFVA\\u201d -> \\u201cUVFA\\u201d, \\u201cUFSA\\u201d -> \\u201cUSFA\\u201d.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The authors address the problem of finding optimal policies in reinforcement learning problems after an initial unsupervised phase in which the agent can interact with the environment without receiving rewards. After this initial phase, the agent can again interact with the environment while having access to the reward function. To address this specific setting, the authors propose to use the successor feature representation of policies and combine it with methods that estimate policies in the unsupervised setting (without a reward function) by maximizing the mutual information of a policy-conditioning variable and the agent behaviour. The result is a method called Variational Intrinsic Successor Features (VISF) which obtains significant performance gains on the full Atari task suite in a setting of few-step RL with unsupervised pre-training. The main contribution seems to parameterize the successor features in terms of a variable specifying the policy. This variable will be the same as the linear weights in the linear model for the reward assume by the successor features representation. Finally, a discriminator aims to predict the linear weights of the policy from the observed state-feature representation.\", \"clarity\": \"I think this is one of the weaknesses of the paper. The writing and clarity could be significantly improved. How do you go from equation 8 to equation 9? How does q lower boun p? In the paragraph after equation 9 the authors mention the score function estimator. But very few details are given. After reading the paper, my impression is that the reproducibility of the results could be very hard, because of the lack of details of the specific implementation. The authors also do not mention that the code will be publicly available after acceptance.\\n\\nI feel that, to better understand the method, the authors should include experiments in simple and easy to understand synthetic environments which can be more illustrative than ATARI.\\n\\nWhat is the difference between the policy parameters theta and the conditioning variable z?\", \"novelty\": \"The proposed approach is novel up to my knowledge. I find the idea of parameterizing the successor features in terms of the policy parameters very innovative.\", \"quality\": \"The proposed approach seems well justified and the experiments performed indicate that the method can be useful in practice. However, I think it would be very useful to have results on other environments besides ATARI. For example, DIAYN contains experiments on a wide range of tasks, including gridworld style tabular experiments to illustrate what their method does. This work would benefit from similar simpler and easier to understand synthetic environments (unlike ATARI).\", \"significance\": \"The proposed contribution seems significant as illustrated by the experimental results and the novel methodological contributions. However, the lack of clarity and the difficulty in the reproduction of the results limit this.\", \"some_minor_comments\": \"I recommend the authors to remove references in the abstract.\\n\\nUpdate after the authors' rebuttal:\\n\\nAfter looking at the response from the authors, I believe that they have successfully addressed my concerns. \\nTherefore, I have decided to update my rating and vote for acceptance. I am looking forward to seeing the python notebook with the implementation of the VISR algorithm.\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have published in this field for several years.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper thoroughly.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper builds upon the idea of Successor Features; this is a principle used to facilitate generalization beyond the finite set of behaviors being explicitly learned by an MDP. The proposed paper ameliorates the need of defining the reward function as linear in some grounded feature space by resorting to variational autoencoder arguments. The derivations are correct, the motivation adequate, the experiments diverse and convincing. The literature review us up to date and the comparisons proper. This is a valuable contribution to the field.\"}" ] }

[ "Ping-yeh Chiang*", "Renkun Ni*", "Ahmed Abdelkader", "Chen Zhu", "Christoph Studor", "Tom Goldstein" ]

[ "certified defenses", "patch attack", "adversarial robustness", "sparse defense" ]

{ "note_id": [ "-E_zJWTDtS", "rkxL6t_njB", "B1gxVBPnir", "rJgZlGvhoH", "rkeTrpE2iH", "SJxfCOehsr", "B1g8i7IoiB", "H1eib7IisS", "Byg569Dcsr", "S1xAU9vqoB", "r1lTNkw6tS", "r1lKQwihYS", "rJefUZ0iFr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730493, 1573845437684, 1573840168003, 1573839337400, 1573829956549, 1573812426268, 1573770141664, 1573769986846, 1573710529750, 1573710421690, 1571807029453, 1571759904523, 1571705161630 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1710/Authors" ], [ "ICLR.cc/2020/Conference/Paper1710/Authors" ], [ "ICLR.cc/2020/Conference/Paper1710/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1710/Authors" ], [ "ICLR.cc/2020/Conference/Paper1710/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1710/Authors" ], [ "ICLR.cc/2020/Conference/Paper1710/Authors" ], [ "ICLR.cc/2020/Conference/Paper1710/Authors" ], [ "ICLR.cc/2020/Conference/Paper1710/Authors" ], [ "ICLR.cc/2020/Conference/Paper1710/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1710/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1710/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper presents a certified defense method for adversarial patch attacks. The proposed approach provides certifiable guarantees to the attacks, and the reviewers particularly find its experiments results interesting and promising. The added new experiments during the rebuttal phase strengthened the paper. There still is a remaining concern that is novelty is limited as this paper could be viewed as the application of the original IBP to patch attacks, but the reviewers believe in that its empirical results are important.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Update\", \"comment\": [\"We updated the manuscript based on valuable conversations with the official reviewers:\", \"We include additional result supporting our main claims using experiments with a larger model\", \"We compare against an interesting training approach, based on a bound pooling formulation proposed by R2, which did not improve upon our top performing model.\", \"We report on experiments with multiple 1x1 patches, also proposed by R2, that our approach can extended to multi-patch setting. In the specific case of 1 x 1 patch, the multi-patch setting is equivalent to defending against $l_0$ attacks.\", \"We update all results, with only marginal differences, after fixing some minor issues in our codes.\"]}", "{\"title\": \"Thank you!\", \"comment\": \"Thank you for supporting our paper!\\n\\n>>> We will make sure to formulate the bounds for this case more formally, and include that in the manuscript.\"}", "{\"title\": \"Thanks for the encouraging results\", \"comment\": \"Thanks for providing the encouraging initial results. Make sure to formally formulate the bound in this case and include these results in the final version of the paper.\\n\\nA caveat is that the input is bounded (e.g., for MNIST the image is in range 0 to 1). For example, if a pixel is already close to 1, it cannot be perturbed up even its corresponding element $w$ is large. A more careful approach will consider both the magnitude of elements in $w$, and the available perturbation budget. But it still only needs to check the elements in $w$ and the input bounds once, so does not increase time complexity.\\n\\nYes I also believe that this bound can be relatively loose for convolutional layers - especially when $k$ is large, e.g., $k$ equals to the number of non-zero elements in $w$, then the perturbation is like a L infinity perturbation with unbounded $\\\\epsilon$. For small $k$ the provided results look reasonable. Some discussions on this can be useful in the paper.\\n\\nI have changed my rating to weak accept, and tend to accept this paper based on the new results.\"}", "{\"title\": \"Thanks for point this out!\", \"comment\": \"Thanks for following up. We were able to get new results for multiple 1x1 patches as suggested, and will include them in the manuscript upon acceptance.\", \"r2\": \"\\\"However I believe multiple 1x1 boxes are doable.\\\"\\n\\nAuthor(s): Indeed, you are correct. Thanks for pointing this out. We have included results for both MNIST & CIFAR for the 1,4,10, 25 1x1 patches below. We will also update this in the main text upon acceptance.\\n\\n\\u2554\\u2550\\u2550\\u2550\\u2550\\u2550\\u2566\\u2550\\u2550\\u2566\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2566\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2566\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2557\\n\\u2551 \\u2551 k \\u2551 Model \\u2551 Clean Accuracy \\u2551 Certified Accuracy \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 MNIST \\u2551 1 \\u2551 MLP \\u2551 98.4% \\u2551 96.0% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 MNIST \\u2551 4 \\u2551 MLP \\u2551 97.8% \\u2551 90.8% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 MNIST \\u2551 10\\u2551 MLP \\u2551 95.2% \\u2551 86.8% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 MNSIT \\u2551 25\\u2551 MLP \\u2551 68.2% \\u2551 35.6% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 MNIST \\u2551 1 \\u2551 Conv3x3 \\u2551 97.0% \\u2551 88.3% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 MNIST \\u2551 4 \\u2551 Conv3x3 \\u2551 92.7% \\u2551 75.9% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 CIFAR \\u2551 1 \\u2551 MLP \\u2551 48.4% \\u2551 40.0% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 CIFAR \\u2551 4 \\u2551 MLP \\u2551 42.2% \\u2551 31.2% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 CIFAR \\u2551 10 \\u2551 MLP \\u2551 37.0% \\u2551 25.6% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 CIFAR \\u2551 1 \\u2551 Conv11x11 \\u2551 34.8% \\u2551 27.4% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 CIFAR \\u2551 4 \\u2551 Conv11x11 \\u2551 25.1% \\u2551 18.3% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 CIFAR \\u2551 10 \\u2551 Conv11x11 \\u2551 17.2% \\u2551 13.8% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 CIFAR \\u2551 1 \\u2551 Conv13x13 \\u2551 38.6% \\u2551 29.7% \\u2551\\n\\u2560\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u256c\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2563\\n\\u2551 CIFAR \\u2551 4 \\u2551 Conv13x13 \\u2551 28.1% \\u2551 19.6% \\u2551\\n\\u255a\\u2550\\u2550\\u2550\\u2550\\u2550\\u2569\\u2550\\u2550\\u2569\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2569\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2569\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u2550\\u255d\\n\\nIt is interesting to note that fully connected only architecture performs the best among all architectures that we experimented in the past few hours. The approach does not perform very well when applied to the convolutional networks because the matrix is very sparse leading to very loose interval bounds with the top-k approach. Moreover, the interval bounds would be vacuous when the number of pixels k is larger than the size of the convolutional kernel. It seems like increasing the size of the convolutional kernel (in the first layer only) improves the performance, but still does not match that of the fully connected. We think it is a very interesting direction to adapt the L0 interval bound to work better with the convolutional layer, and we will further investigate this in our future work.\\n\\n>>> We will also add the method and results in the manuscript upon acceptance.\"}", "{\"title\": \"Thanks for the additional results; I am considering increasing the rating; however I believe multiple 1x1 boxes are doable\", \"comment\": \"Thanks for the additional results on combining multiple boxes. Please make sure to mention it in the main text for the final revision, as I believe it is an important contribution.\\n\\nI am considering increasing the rating. However I believe multiple 1x1 boxes are doable.\\n\\nLet consider there are 32x32 positions (so dimension of input $x$ 1024), and we have a linear layer with weights $W$ in shape $N \\\\times 1024$, where $N$ is the number of neurons. For each neuron, we can see its output as a dot product between the 1024-dimensional input, and a row of the weight matrix (denoted as $w$). Now we consider the case, where we can only perturb 1 of the 32x32 positions. What is the output lower bound and upper bound for that neuron? Apparently, we can increase the output of that neuron by increasing a pixel value that is corresponding to a positive coefficient in $w$, and vise versa. And apparently to reached the upper and lower bound, we perturb the single pixel corresponding to largest element in $w$. Importantly, you just need to look at the largest element in $w$ to determine its output upper and lower bound. We don't need 32x32 boxes at all.\\n\\nNow we consider the general case, where we can perturb $k$ of the 32x32 positions. Then you need to look into the $k$ largest element in $w$ to determine the neurons output lower and upper bounds. There is no need to construct 32x32 boxes, again.\\n\\nI strongly encourage the authors to look into the multiple 1x1 patch case above, as it is a very important use case, and also should be fairly straightforward and efficient to implement.\"}", "{\"title\": \"Response to Reviewer 1 (1/2)\", \"comment\": \"Thanks for the careful remarks with a healthy dose of skepticism that helped improve the manuscript. We ran additional experiments using a larger model, as requested, and elaborated on the significance of the reported results w.r.t. state-of-the-art in certified accuracy.\\n\\n*** Main Comments ***\", \"r1\": \"\\u201csometimes the IBP approach is useful for defense, although often with a significant loss in accuracy on clean data\\u201d\\n\\nAuthor(s): The trade-off between robustness and generalization has been argued to exist both from a theoretical perspective and also empirically [3, 4]. Some loss in accuracy in certified robustness compared to clean accuracy is typical for all certified (and most empirical) defenses. Our proposed certified defense, which is the first certified defense against adversarial patch attacks, is no exception to this. To put our reported results in perspective, we consider state-of-the-art results for (a) certified robustness against L-inf perturbations, and (b) (non-certifying) empirical defenses against patch attacks; see below. \\n\\n(a) The State-of-the-art certified model for CIFAR with 16/255 L-inf perturbation using deterministic algorithms [5] achieves clean accuracy of 34.0%, whereas our clean accuracy for our most robust CIFAR 5 \\u2a09 5 model is 47.8% (for the updated large model). To further appreciate the performance of our model, note that our threat model allows the attacker to change each pixel within the patch to any desired value, whereas the L-inf perturbation model of [5] assumes the adversary can only change each pixel by at most some epsilon, e.g., 16/255. \\n\\n>>> This remark was added to the manuscript at the bottom of Section 5.1.\\n\\n(b) State-of-the-art empirical (non-certified) defenses against patch attacks achieve much lower robustness (i.e, adversarial accuracy) than our approach. In particular, local gradient smoothing (LGS) [2] achieves 24.2% empirical accuracy on CIFAR 2\\u2a092 while our method achieves 51.9% certified accuracy (for the largest model trained). Similarly, digital watermarking (DW) [1] achieves 15.6% empirical accuracy on CIFAR 5\\u2a095 while our method achieves 30.3% certified accuracy (for the largest model trained).\\n\\nFinally, we acknowledge that certified defenses may not be ready for real world applications. However, our contribution makes an important step towards securing ML systems against adversarial patch attacks by substantially improving both the clean and robust accuracy achievable by a certified model.\\n\\n\\nReferences\\n\\n[1] Jamie Hayes. On visible adversarial perturbations & digital watermarking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 1597\\u20131604, 2018.\\n\\n[2] Muzammal Naseer, Salman Khan, and Fatih Porikli. Local gradients smoothing: Defense against localized adversarial attacks. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1300\\u20131307. IEEE, 2019.\\n\\n[3] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy?\\u2013a comprehensive study on the robustness of 18 deep image classification models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631\\u2013648, 2018.\\n\\n[4] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P Xing, Laurent El Ghaoui, and Michael I Jordan. Theoretically principled trade-off between robustness and accuracy. ICML, 2019.\\n\\n[5] Zhang, Huan, Hongge Chen, Chaowei Xiao, Bo Li, Duane Boning, and Cho-Jui Hsieh. \\\"Towards Stable and Efficient Training of Verifiably Robust Neural Networks.\\\" arXiv preprint arXiv:1906.06316 (2019).\"}", "{\"title\": \"Response to Reviewer 1 (2/2)\", \"comment\": \"Continued ..\", \"r1\": [\"\\u201creal-life images with patches that are of real-life size\\u201d\", \"Author(s): As CIFAR images are only 32\\u2a0932, we are unable to experiment with patch size this big. If we want to do certification for 40x40, we would have to conduct our experiments on ImageNet. As it currently stands, IBP has difficulty extending to imagenet sized datasets, and we leave this as future work. Note that CIFAR-10 is currently a widely used benchmark for adversarial defense (both empirical and certified).\", \"*** Other Comments ***\", \"We have fixed the typos in equation (5) & (6)\", \"Yes, we forward pass all patches in order to get a certificate for all of our experiments. We have included an additional paragraph in section 4.2 to clarify.\", \"Our lower and upper bound is simply the image domain (0-1), and we don't really have an epsilon.\", \"We added description for the element-wise absolute value\", \"There was a minor bug in our code that didn\\u2019t have a significant impact on any of the results. All tables were updated with new runs.\"]}", "{\"title\": \"Response to Reviewer #3\", \"comment\": \"Thanks for the positive assessment and supportive remarks!\", \"r3\": \"\\u201chave to conduct many forward pass of the network with respect to all patches at all locations?\\u201d\\n\\nAuthor(s): Yes, we forward pass all patches in order to get a certificate. We have included an additional paragraph in section 4.2 to clarify.\\n\\n***Other Comments***\\n- There was a minor bug in our code that didn\\u2019t have a significant impact on any of the results. All tables were updated with new runs.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"Thanks for the support and valuable suggestions!\\n\\n*** Main Comments ***\", \"r2\": \"\\u201cConsidering multiple patches (at different locations) on a single input. [..] multiple 1x1 in my opinion is both technically feasible and practically important, and should definitely be included in the paper\\u201d\\n\\nAuthor(s): Unfortunately, this interesting setup is computationally infeasible. We actually considered this formulation earlier but could not pursue it due to the combinatorial explosion in the number of patches, see below for more details.\\n\\nIn the specific example of 1\\u2a091 patch, when we take the maximum of upper bounds generated by all patches, we are not getting the upper bounds for moving all pixels at the same time, rather we get the worst case bounds when we are allowed to move a single pixel, but not moving any of them together. To give an actual certificate, we have to consider (32\\u2a0932) position for the first patch and (32\\u2a0932-1) position for the second patch, and we would have to forward pass all of these possible combinations forward in order to get an actual certificate. Unfortunately, this is computationally infeasible, but is very interesting nonetheless.\\n\\n\\n*** Other Comments ***\\n- We have removed reference to Weng's\\n- We have fixed the typos in equation (5) & (6)\\n- We have included empirical adversarial accuracy to Table 3\\n- There was a minor bug in our code that didn\\u2019t have a significant impact on any of the results. All tables were updated with new runs.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper attempts to extend the Interval Bound Propagation algorithm from (Gowal et al. 2018) to defend against adversarial patch-based attacks. In order to defend against patches which could appear at any location, all the patches need to be considered. This is too computationally expensive, hence they proposed to use a random subset of patches, or a U-net to predict the locations of the patches and then use those patches to train. The algorithm is tested on the MNIST and CIFAR-10 datasets and it was shown that sometimes the IBP approach is useful for defense, although often with a significant loss on accuracy on clean data (e.g. on CIFAR the loss on clean accuracy is an astounding 300% -- from 66.5% - 35.7%).\\n\\nI think the technical contribution of this paper is a bit weak in that they mostly followed the original IBP and the only novelties are the random patch training and guided patch training. I partially like how the experiments are conducted, especially the one that generalizes to other shapes. On the other hand, the networks that are tested seem pretty poor by any standard. An experiment that is definitely missing is a CIFAR network that performs a little better than the current one. Clean accuracy of only 66.5% and 47.2% are very lousy for CIFAR.\\n\\nAnother missing experiment is one that would test on different epsilon values. I couldn't find what are the current epsilon values used?\\n\\nBesides, since this work is testing on adversarial patches, I would like to at least have it applied to some real-life images with patches that are of real-life size. I could care a bit less on how good it is, but one can still make an empirical test (e.g. certified defense accuracy on 5x5 patches, but empirical test using real-life sized patches 40x40 or 80x80) and see how the results would be. All the experiments mentioned above would significantly strengthen the experiments section of the paper.\\n\\nI don't think I read anywhere a confirmation that the testing is performed on all patches of the prescribed size. Could the authors please confirm whether this is true?\", \"minor\": \"There is a typo in Eq. (5) and Eq. (6), where the second term multiplied by |W^(k)| should be \\\\underline{z}^(k-1) - \\\\bar{z}^(k-1) instead of \\\\underline{z}^(k-1) + \\\\bar{z}^(k-1)\\n\\nYou should mention that |W^(k)| stand for element-wise absolute value when it first appears.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper proposes a certified defense for adversarial patch attacks.\\nTechnically, the authors use the well developed IBP based methods (Gowal et\\nal., 2018, Mirman et al., 2018, Zhang et al. 2019). The technique is simple\\nbut effective. Since the number of possible patches are quadratic w.r.t image\\ndimension, to reduce the number of bounds to propagate, the authors propose a\\nU-Net based NN to predict the worst case scenario, and only propagate \\\"worst\\ncase\\\" bounds predicted by the U-net.\\n\\nEmpirically, the proposed method gets good results, with certified accuracy\\nsometimes even higher than empirical accuracy by previous methods. The authors\\nalso provide results for transferring robustness properties to shapes that are\\nnot included during training.\\n\\nOverall, the contribution of this paper is novel, and results are promising,\\nbut it still has some missing components, especially the idea of combining\\nmultiple IBP bounds into one, which can be very effective for adversarial\\npatches, as I will elaborate below.\", \"suggestions_and_questions\": \"The core idea behind IBP is that for whatever input perturbation is given (any\\nLp norm or semi-norm, or non-norm based perturbations like patches at arbitrary\\nlocations), it converts them to per-neuron lower and upper bounds after the\\nfirst linear/conv layer. For example, if the input perturbation is *two* patches\\nB_1 and B_2, after propagating them through the first layer of the network, we\\ngot two lower bounds l_{i,1}, l_{i,2} and two upper bounds u_{i,1}, u_{i,2} for\\nthe i-th neuron. We then take the worst case bound, l_i = min(l_{i,1},\\nl_{i,2}), u_i = max(u_{i,1}, u_{i,2}) and propagate only one set of bounds l_i\\nand u_i to the next layer. The authors should explore on this direction, as\", \"detailed_below\": \"1. For the exhaustive patch enumeration in (11), we can actually greatly reduce\\nthe computation cost by combining the bounds of different patches after the\\nfirst layer of the network, as I mentioned above. At the input layer, the\\nnumber of bounds (each for one possible location of patch) are large; but after\\nthe first linear/conv layer, we can compress them to one or a small group of\\nbounds by taking the worst cast of them, like l_i = min(l_{i,1}, ...,\\nl_{i,|L|}), u_i = max(u_{i,1}, ..., u_{i,2}). The patches close with each other\\nshould also produces similar lower and upper bounds, so taking min or max over\\nthem will not make the combined bounds much worse. This is better than U-net\\nprediction since we are guaranteed to include the worst case scenario.\\n\\n2. Considering multiple patches (at different locations) on a single input. In\\nthe simplest case, consider multiple 1x1 patches (in fact, this is equivalent\\nto bounded L0 norm threat model); since each patch is 1x1 (only changing one\\npixel), the bounds should be relatively tight after the first layer, and after\\nthe first linear/conv layer we got 28*28 or 32*32 bounds which will be combined\\ninto one or very few sets of lower and upper bounds that will be propagate into\\nlater layers. Multiple larger patches (2x2, 5x5) can be difficult since bounds\\nare looser; multiple 1x1 in my opinion is both technically feasible and\\npractically important, and should definitely be included in this paper.\", \"minor_issues\": \"1. Eq. (5) and (6) are incorrect; the second term should be \\\\overline{z} -\\n\\\\underline{z}. Also it is missing the bias term.\\n\\n2. In related works (section 4.1, page 3), (Weng et al., 2018) is not a defense\\nmethod (it is a certification method, and no training is involved).\\n\\n3. In Table 3, it is better if the authors can provide empirical adversarial\\naccuracy to IBP defended networks as well.\\n\\nOverall, I think it is a good paper but the authors should explore more to\\nstrengthen their contributions. I gave a weak reject but I will not hesitate\\nto recommend an accept as long as the authors can provide additional results\\nmentioned above.\"}", "{\"rating\": \"6: Weak Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposed a certified defense method for adversarial patches. The paper is motivated by the finding that several existing works on adversarial patch defenses are easily \\\"breakable\\\" by in the white-box setting. The idea of the proposed method is derived Interval Bound Propagation (IBP), which is originally proposed for certified defense against adversarial noise. To simplify the certificate training in the patch defense setting (which original scales quadratically with respect to the image size), two randomized training methods are proposed. Lastly, experimental results indeed verify the effectiveness of the proposed method.\\n\\nThe paper is well-written and very well-organized. It is interesting to see supposedly strong adversarial patch defense methods \\\"break down\\\" in a very simple setting. And the contribution of the paper is significant to the field.\", \"i_do_have_a_question_with_regard_to_the_randomized_training_method\": \"Although random patches (or selected worse-case patches at random location) can be used for certificate training, in order to create a certificate during testing, does it mean that you do have to conduct many forward pass of the network with respect to all patches at all locations?\"}" ] }

[ "Yonglong Tian", "Dilip Krishnan", "Phillip Isola" ]

[ "Knowledge Distillation", "Representation Learning", "Contrastive Learning", "Mutual Information" ]

{ "note_id": [ "uWfGnHDk_u4", "w_JckE1MG", "H1lvNCwcjS", "S1l0WCw9or", "SklFaaP5jr", "HkevRLo4cB", "Hyl91UHRFB", "SkeiTRlTYS" ], "note_type": [ "comment", "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1633874005503, 1576798730464, 1573711407408, 1573711366216, 1573711296583, 1572284110856, 1571866081958, 1571782339276 ], "note_signatures": [ [ "~Xiangyu_He1" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1708/Authors" ], [ "ICLR.cc/2020/Conference/Paper1708/Authors" ], [ "ICLR.cc/2020/Conference/Paper1708/Authors" ], [ "ICLR.cc/2020/Conference/Paper1708/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1708/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1708/AnonReviewer3" ] ], "structured_content_str": [ "{\"title\": \"A misleading typo in Equation(10)?\", \"comment\": \"It seems that $N\\\\mathbb{E}_{q(T,S|C=0)}[1-\\\\log(h(T,S))]$ is actually $N\\\\mathbb{E}_{q(T,S|C=0)}[\\\\log(1-h(T,S))]$.\"}", "{\"decision\": \"Accept (Poster)\", \"comment\": \"This paper presents a new distillation method with theoretical and empirical supports.\\n\\nGiven reviewers' comments and AC's reading, the novelty/significance and application-scope shown in the paper can be arguably limited. However, the authors extensively verified and compared the proposed methods and existing ones by showing significant improvements under comprehensive experiments. As the distillation method can enjoy a broader usage, I think the propose method in this paper can be influential in the future works.\\n\\nHence, I think this is a borderlines paper toward acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer 3\", \"comment\": \"Dear Reviewer 3,\\n\\nThank you for your feedback.\\n\\n\\u201cthe problem then becomes nothing but trying to minimize another distance metric between teacher and student networks on an intermediate layer\\u201d\\nDifferent distance metrics can make a big difference. Theoretically, we show our objective is maximizing a lower bound on mutual information between the latent representations of the teacher and student network. Empirically, we show our method consistently outperforms all 12 recent knowledge distillation methods published in NeurIPS, CVPR, ICCV and other top conferences (we note that many of these previous works were also just different distance metrics between teacher and student networks).\\n\\n\\u201cComparing with original distillation method (KD) I'm not sure how significant the improvement is\\u201d\\nWe clarify that it\\u2019s important to see the relative improvement, rather than the absolute improvement. Suppose the gap of accuracy between well-trained teacher and student networks is only 3 percentage points, how can we ask the knowledge from the teacher model improves the student by 10 absolute percentage points, or even 5 percentage points. Indeed, our CRD outperforms the second best method (KD, see Table 1&2) with a 57% average relative improvement.\\n\\n\\u201chowever it is in the spirit of what the distillation work is all about and I do not see this as being fundamentally different, or at least not different enough for an ICLR paper\\u201d\\nThese 12 other methods we compare here are just as similar to KD but are published in top venues. However, our method not only outperforms recent SOTA methods, but also forges a connection between two literatures that have evolved mostly independently: knowledge distillation and representation learning.\\n\\n\\u201cIt would be helpful to also include the variance of each experiment\\u201d\\nWe have added the standard deviation in Sec 6.9 and Tables 9&10.\\n\\n\\u201cWhy this particular distance metric between the representations? Why not just L2?\\u201d\\nWe have both theoretical analysis and empirical validation for the improvement provided by this distance metric. An L2 metric is provided by FitNet, which we compare to in Tables 1, 2, 4, 5 and Figure 3 (CRD outperforms FitNet).\"}", "{\"title\": \"Response to Reviewer 1\", \"comment\": \"Dear Reviewer 1,\\n\\nThank you very much for the constructive comments.\\n\\n1. For our CRD results in Table 1&2, we only use CRD loss. We have added CRD+KD results in the revised version to avoid such confusion, and CRD+KD shows further small improvement.\\n\\n2. We have added a discussion of the \\u201cDeep Mutual Learning\\u201d paper in Sec 6.8. We have evaluated several different distillation methods under the mutual learning setting, see Table 8.\\n\\nPlease don\\u2019t hesitate to let us know for any further feedback. Thanks!\"}", "{\"title\": \"Response to Reviewer 2\", \"comment\": \"Dear Reviewer 2,\\n\\nThank you for the very constructive comments.\\n\\n1. We have revised the contribution section to make it more accurate. \\n\\nHyperparameters (architecture types, transfer connections between layers). For architecture types or student-teach pairs, we just randomly sampled over a candidate pool which contains widely-used networks. As for the transfer connections for feature based distillation methods, we tried two methods: (1) connect the layers with 0.25, 0.5 and 0.75 depth of the whole network; (2) connect the last layer of each specific spatial size, e.g., the last layers, which have spatial size of 32, 16, 8 and 4, are connected correspondingly. We found (2) generally works better and stick to it.\\n\\nWe agree that we might achieve higher accuracies if we separately tune hyperparameters for each specific student-teacher pairs. But on the other hand, we consider it to be more interesting to see how each method can generalize across different architecture types without much tuning. This is also our motivation of randomly sampling different student-teacher pairs to form the benchmark.\\n\\n2. Compatibility. We have included another table in the revised paper (see Sec. 6.6 and Table 6) discussing the combination of different distillation objectives. We found: (1) other objectives combined with KD still underperform CRD; (2) combining CRD with other objectives, such as KD and PKT, can further improve the performance.\\n\\n3. The main reason why we don\\u2019t extensively compare with all other baselines in other Tables is that we have limited resources. Given such a situation, we are inclined to distribute our resources towards building complete results (Tables 1&2) on the standard CIFAR-100 benchmark, rather than the proof of concept on specific settings, such as Table 5 and Figure 3. \\n\\nWe appreciate the idea of comparing transferability with other methods. We have included a new table (see Sec 6.7 and Table 7) showing the transferability of all methods with different architectures.\\n\\n4. VID also claims to capture the MI between the representations of student and teacher, but their instantiation is very different from ours. In order to achieve a tractable computation of MI, VID used a Gaussian distribution as a variational approximation of the true conditional distribution. Our conjecture is that the true conditional distribution is perhaps very different from the Gaussian assumption, and therefore it leads to suboptimal results.\", \"other_issues\": \"We have added a note to indicate that Online-kd does not use pre-trained ResNet-34. We can also remove this item if needed.\\nWe tried CRD on middle layers but only see very marginal improvement, so we opt to keep our methods simpler. Indeed, only using penultimate layers makes CRD more robust to different architecture types.\\nFSP only works for teachers and students with the size of features being equal, so it\\u2019s not available there. We will mark it as N/A.\\nThe typos have been fixed.\\n\\nPlease don\\u2019t hesitate to let us know for any additional comments. Thank you!\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": [\"Summary & Pros\", \"This paper proposes a well-principled distillation method based on contrastive loss maximizing the mutual information between teacher and student models.\", \"This paper provides extensive experiments that demonstrate the effectiveness of the proposed method. The performance gap compared to the existing distillation approaches seem to be significant.\"], \"major_concerns\": [\"The authors claim that \\\"none of the other methods consistently outperform KD on their own\\\". I feel that this claim is somewhat aggressive because some of them outperform KD without combining with KD, e.g., Table 1 in FT (Kim et al., 2018) and Table 2 in SP (Tung & Mori, 2019). Since the distillation (or transfer) methods are typically sensitive to hyperparameters (e.g., architecture types, transfer connections between layers), I also wonder how to set the hyperparameters for baselines, especially in Table 2, because choosing the transfer connections between different architectures is very important when using feature-based methods such as FitNet and AT.\", \"Moreover, the baselines are developed for improving distillation performance, not replacing KD. Especially, feature-based methods (e.g., FitNet, NST) are easily combined with logit-based ones (e.g., KD). So I think the compatibility between the proposed and exisiting methods should be checked. However, in this paper, only Table 4 shows the compatibility with KD (CRD+KD).\", \"The authors compare the proposed method with only KD, AT, FitNet except Table 1-3. For example, when evaluating the transferability (Table 4), why other baselines are not compared?\", \"VID also maximizes MI between penultimate layers. What is the key difference and why CRD perform better? I think detailed verfication should be provided in the paper.\", \"Minor Comments\", \"Comparison with Online-KD is unfair because it does not use pre-trained ResNet32.\", \"Why only use penultimate layers? CRD between intermediate layers is also available like VID.\", \"A result in Table 1 is missing (FSP WRN40-2 -> WRN40-1).\", \"In Section 4, both CKD (contrastive knowledge distillation) and CRD (contrastive representation distillation) are used, so one of them should be removed.\", \"In Section 4.1 Transferability paragraph, \\\"Table 3\\\" should be changed to \\\"Table 4\\\".\", \"On page 4, \\\"space\\\" after \\\"before the inner product.\\\" should be removed.\", \"I think the proposed method is well-principled and provides meaningful improvements on various distillation settings, thus this paper seems to be above the borderline. It would be better if additional supports for the above concerns is provided in a rebuttal.\"]}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper combines a contrastive objective measuring the mutual information between the representations learned by a teacher and a student networks for model distillation. The objective enforces correlations between the learned representations. When combined with the popular KL divergence between the predictions of the two networks, the proposed model shows consistently improvement over existing alternatives on three distillation tasks.\\n\\nThis is a solid work \\u2013 it is based on sound principles and provides both rigorous theoretical analysis and extensive empirical evidence. I only have two minor suggestions.\\n\\n1, From Section 3.2 to Section 3.4, it is not clear to me that on the model compression task, are both the proposed contrastive loss and the loss in Eq. (10) used?\\n\\n2, The \\u201cDeep mutual learning\\u201d, Zhang et al, CVPR\\u201918 paper needs to be discussed. I\\u2019d also like to see some experiments on the effects of training the teacher and student networks jointly from scratch using the proposed loss.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"I do not necessarily see something wrong with the paper, but I'm not convinced of the significance (or sufficient novelty) of the approach.\\n\\nThe way I understand it, an independent assumption on internal representation is relaxed by capturing correlations between them and higher order dependencies between them using a different objective function(distance measure), the problem then becomes nothing but trying to minimize another distance metric between teacher and student networks on an intermediate layer. \\n\\nComparing with original distillation method (KD) I'm not sure how significant the improvement is. And technically this is just a distance metric between the internal representation or output of the student and teacher. Sure it is a more involved distance metric, however it is in the spirit of what the distillation work is all about and I do not see this as being fundamentally different, or at least not different enough for an ICLR paper.\\n\\nThe experimental results also suggest only a marginal improvement compared to other methods. It would be helpful to also include the variance of each experiment i.e., it was mentioned that the results were averaged by repeating 5 experiments to make sure the proposed method consistently better than others. Rightnow, It is hard to compare with other approaches. The paper gives off a feeling that this method is not novel. Why this particular distance metric between the representations? Why not just L2?\"}" ] }

[ "Moritz Hoffmann", "Frank Noe" ]

[ "euclidean distance matrices", "wgan", "point clouds", "molecular structures" ]

{ "note_id": [ "s9nlt8thbT", "BkeaLRpdoB", "SylhY66_iB", "HyxknhaOiB", "BkeMvX5TFH", "rkeVnp86YH", "Skx9wBpHFB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730434, 1573604948781, 1573604740031, 1573604519437, 1571820377677, 1571806636080, 1571308898438 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1707/Authors" ], [ "ICLR.cc/2020/Conference/Paper1707/Authors" ], [ "ICLR.cc/2020/Conference/Paper1707/Authors" ], [ "ICLR.cc/2020/Conference/Paper1707/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1707/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1707/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes a parametrisation of Euclidean distance matrices amenable to be used within a differentiable generative model. The resulting model is used in a WGAN architecture and demonstrated empirically in the generation of molecular structures.\\n\\nReviewers were positive about the motivation from a specific application area (generation of molecular structures). However, they raised some concerns about the actual significance of the approach. The AC shares these concerns; the methodology essentially amounts to constraining the output of a neural network to be symmetric and positive semidefinite, which is in turn equivalent to producing a non-negative diagonal matrix (corresponding to the eigenvalues). As a result, the AC recommends rejection, and encourages the authors to include simple baselines in the next iteration.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to AnonReviewer2\", \"comment\": [\"Thank you for your kind review and positive feedback!\", \"We added to the explanation of the chosen loss function, especially with respect to the atom types via a cross entropy term and added a short explanation of the drift term. Furthermore an intuitive description of attaining high and/or low values in the critic network $C$ was added.\", \"Expanding the application scope is indeed something that we plan to do along with bias/penalty terms towards energetically favorable states and non violated valencies.\"]}", "{\"title\": \"Response to AnonReviewer1\", \"comment\": [\"We thank you for your review, questions, and input. In general however we do not think that the rating of \\\"3 - weak reject\\\" is properly motivated. Addressing the points:\", \"We agree that, since the kernel distance is defined as $d(X, Y) = K(X, X) + K(Y, Y) -2K(X, Y)$ for a kernel $K$, producing a valid EDM $D_{ij}=M_{ii}+M_{jj}-2M_{ij}$ (Eq. (3) in the manuscript) is equivalent to learning the symmetric positive semi-definite (kernel) matrix evaluations $M_{ij}$. The rank of $M$ corresponds to the embedding dimension.\", \"Equation (5) holds whenever the function $g$ has a preimage space that contains the eigenvalues. There are different ways to define a matrix function, one of them is via the Jacobi canonical form, which reduces in our case to an eigenvalue decomposition and application of the function on the individual eigenvalues (see, e.g., Higham, N. J. (2008). Functions of matrices: theory and computation (Vol. 104). Siam.).\", \"Algorithm 1, step 14: It is now noted that G is a neural network. All the steps are differentiable so depending on the optimization objective (for instance in a GAN architecture) backpropagation can yield a target distribution in EDM space (indicated by \\\\nabla L). SchNet in this case is used as a Critic (or in classical GANs as Discriminator) network but is not strictly required to sample in EDM space.\", \"Comparison to simple and straightforward baselines is going to be included in future research. While the GAN implementation of our approach generates substantial diversity in configuration space, the diversity in chemical space is limited, but can be improved by choosing different generative structures, e.g., autoregressive (arXiv 1906.00957).\", \"The standard way to compare two structures strongly depends on the considered type of graph. In proteins for example there is a polymeric backbone structure that can be leveraged for identity assignment via the help of sequence alignment. Since these are small molecules without polymeric structure and finding the right assignment is in principle NP hard, the Hungarian algorithm is often used to tractably find an optimal solution, where optimal refers to the choice of cost matrix.\"]}", "{\"title\": \"Response to AnonReviewer3\", \"comment\": [\"Thank you for the positive, very detailed, and constructive review. We have adressed your points in the following manner:\", \"Main request 1a: The unclear sentence about the type vector and cross entropy is expanded with further details in the manuscript. There is another linear transformation between the output of the neural network parameterizing the symmetric matrix $\\\\tilde{L}$ and a vector $\\\\mathbf{t}\\\\in\\\\mathbb{R}^{m\\\\times n\\\\times n_\\\\mathrm{types}}$ which is due to the use of a softmax non-linearity a probability distribution over the types per atom. This probability distribution is compared against a one-hot encoded type vector $\\\\mathbf{t}_\\\\mathrm{ref}$ representing the type composition in the considered isomer with a cross entropy term $H(\\\\mathbf{t}, \\\\mathbf{t}_\\\\mathrm{ref})$. As in this example the chemical composition never changes, the type vector $\\\\mathbf{t}_\\\\mathrm{ref}$ is always constant.\", \"Main request 1b: Bonds and bond types are not generated by the network but inferred by OpenBabel which uses a method based on spatial proximity and atom type for assignment. We discard samples which are disconnected and/or violate valencies based on the inferred bonds. This is something that can potentially be included at least approximately during training by using penalty terms. We consider this a good addition for future work, thank you for the suggestion! A sentence reflecting this (as well as including an energy term) has been added to the conclusion.\", \"Main request 2: Indeed adding penalty terms which encourage a valid bonding structure as well as including a force field (in this case for small molecules) should help results and we now mention this in the conclusion.\", \"Algorithm 1 now mentions that G is a neural network, z a sample from a simple prior, and that sampling in EDM space is in general not uniform. We added an explanation to the application of the softplus activation in the main text, in particular that one can also apply it for the first $d$ eigenvalues and set it to $0$ for the rest, in which case the rank of $M$ does not need to be optimized for. Which of the two methods (setting to $0$ versus penalizing) is better demands further research. In Line 8 it is not the eigenvalues of $D$ but the eigenvalues of $-0.5JDJ$, which is now mentioned in the text for clarification. It is an if and only if condition on EDMness.\", \"In the WGAN section we mention that the prior distribution can be chosen as normal distribution to make the connection to the algorithm and application better/easier.\", \"An intuitive explanation for the values of the critic for an individual sample has been added: Low values indicate the belief a sample stems from the generated distribution, high values indicate the belief that a sample stems from the data distribution.\", \"The condition on the Lipschitz constant comes from the Kantorovich-Rubinstein duality theorem for the Wasserstein-1 metric. Without it, we are not actually minimizing a metric between distributions. The constant can actually be any positive bound, in which case it is not the Wasserstein-1 metric but a fixed multiple of it, i.e., $L\\\\cdot W^1(p, q)$. Intuitively it is a restriction on the maximal gradient that the critic function can attain, which produces a one-dimensional embedding in which a decision can be made which sample belongs to which distribution.\", \"The drift term (Eq. (9)) helps to fixate the critic values for real samples around $0$, otherwise the optimization objective only considers the relative difference; ambiguity is reduced. A short explanation has been added to the main text.\", \"We emphasize more on the energies result by an additional sentence in the main text. Also we added that while our GAN implementation has substantial diversity in configuration space and only limited diversity in chemical space, this can be improved by using different generative structures, e.g., (Gebauer et al 2018, 2019).\", \"We agree that enriching Figure 3 with new structures by differentiating the very similar ones would improve it and plan to work on that in the future.\", \"We fixed the duplicate entry in the bibliography, thanks for spotting that!\"]}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"8: Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper addresses the important problem of molecular structures generation, and more generally of efficient point-cloud distributions learning in d-dimensional space.\\nThe paper is written clearly, the pseudo-code is presented in a clear way, striking a good balance between explicit-ness and concision.\\n\\nAfter my first reading I was disappointed in the experiments, because I read the paper quite quickly at first. After a second more careful read I understood most of it and was much more enthusiastic. Disclaimer: I am inexperienced in this particular field (GANs and GANs for molecular generation) so I lack literature knowledge, and I may be over-evaluating the quality of performances (compared to recent results).\\nBut currently I believe that the authors are a bit too humble about their results (a rather uncommon phenomenon). There is a part of the method that I would like to have clarified (about bond/atom types), but other than this clarification, I strongly recommend the paper to be accepted.\\n\\nThe paper deals with a couple of distinct, related problems. One is that of sampling valid (Euclidean) distances matrices (EDM). This is done with algorithm 1.\\nThe paper uses these EDM to train a generator G directly in EDM-space, against a Critic network C (the architecture of which is taken from existing literature).\\nSome of the output configurations have to be discarded due to incorrect bond types assignment (this is the part that is still obscure to me).\\nThe remaining outputs are chemically sensible in terms of bond types/valence/local chemistry/etc, but also, and quite impressively, have reasonable ground state energies ! In this sense, the paper produces samples that may deserve to be added into the QM9 data set (After some data augmentation using DFT or other physics-validated methods of course).\", \"i_have_mainly_two_requests\": \"1. The bond type or atom assignment is not clearly explained. I ask further detailed questions below, but overall I think you should attempt to make a pedagogical explanation of how the atom types are assigned/learned, and how some of them are later discarded.\\n\\n'The generator produces an additional type vector in a multi-task fashion which is checked against a constant type reference with a cross-entropy loss.'\\nCould you explain this better ? At least in the appendix. Otherwise this very dense sentence remains mysterious. Eq.11 is currently the single occurrence of the H, t, and tref terms.\\n\\nAbout the validity test using Open Babel. Again, what happens with bond types is not very clear. Are they set in stone at generation time, and then most of these affected types are 'wrong' and are discarded by Open Babel ?\\nWhy don't you include this validation at training time (if it's very complicated to do, explain why) ?\\nIn any case, please clarify this paragraph, as for now it is cryptic.\\n\\n2. Part of the paper goal is to achieve learning of point clouds distributions and not about chemistry. However, discovery of new structures (and their conformations) is in itself a big topic. I think it would be good to mention a couple of follow ups to your work in the conclusion. For instance, it would seem rather natural to me to include some equivalent of OpenBAbel and Energy-estimates within the learning loop, so that the generator directly generates valid (open-babel wise) and reasonable (energy level-wise) structures.\\nBesides the conclusion, you should stress out better the significance of your results with a couple of comments here and there. (I have more precise suggestions below).\\n\\n\\nSome other comments to improve the paper (in clarity or other):\\n\\nIn algorithm 1, you should precise that G() is your NN-based generator network. At this point of the paper G and z and N_z have never been mentioned. This is a problem when reading the paper for the first time.\\nLine 5, you could explicitly precise, 'with (5)'. Also I am puzzled by the use of the softplus here, but later in page 5 you say g() is softplus for the first three eigenvalues, and set to 0 for all others. Why not already anticipate and say this in section 2 ?\\nLine 8, I would have written 'eigenvalues .... of D' (what is the role of Eq. 1 here?)\\n\\nAt some point (i.e. around end of section 2), a comment would be welcomed, about whether you may sample the space of EDMs uniformly at random using algorithm 1 (I think you do not, and it is ok, but the question naturally arises and it's not addressed).\\n\\n\\u00b4\\u00b4which transforms a prior distribution into a target distribution\\u00b4\\u00b4\\nyou could specify that this prior is the Gaussian N(0,1)^Nz in this case, to better connect with algorithm 1.\", \"about_equation_6\": \"it would be nice to have some intuitive explanation of what the output values of C(x) mean. They are scalars that represent the opinion of C on the molecule x, so they are the probability that the observed molecule is 'a true molecule x' ? You should recall that for the inexperienced reader.\\n\\nStill about Eq.6 : could you quickly provide the motivation for demanding C to be L-Lipschit with $L \\\\leq 1$ ?\\n\\nThe drift term Eq.9 seems to be some sort of regularization, maybe it could be mentioned once in the text for completeness?\", \"using_mopac_stewart_and_figure_4\": \"you could insist more on this result. Up to that point I was very dubious about the usefulness of the whole work, because I would have expected many generated samples to be highly unrealistic, i.e. having huge energies, and so being extremely unstable. Instead here you show it is not the case, and even all your energies lie within the range of observed energies ! This is a very strong result, that is not obvious to expect, and is highly valuable (even after keeping only 7.5% of structures, this is still a strong result.)\\nI think this test and the corresponding result should be emphasized more.\\n\\nYou observe new topological types and complain there are not many. I think ~20 new topologies is not a small number (for so few atoms), and you may be rather proud of it. What is too bad is that you do not have a tool for differentiating between the very similar conformations (that correspond to thermal fluctuations around a given structure) and the conformations that encode a new structure (which does not necessarily means new topology). \\nIf you had this tool, you could enrich figure 3 with a curve showing the new structures (not counting conformational variants).\\n\\nrefs 2017a and 2017b are the same.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"Authors of this paper present architecture to produce valid Euclidean distance matrices. A Wasserstein GAN is then constructed by utilizing a permutation invariant critic network based on the architecture. Generating molecular structures in a one-shot fashion is conducted using the produced distance matrices in 3-d embedding.\\n\\nIn Section 2, the constraint on L makes M symmetric and positive semi-definite. This seems to be equivalent to treating M as a kernel matrix, and D is the pairwise distance between the kernel function induced by M. So, learning a valid Euclidian distance matrix is same as learning a kernel function.\\n\\nAuthors need to provide the evidence that Equation (5) holds for all function g, especially for the used softplus activation function. As I know, it holds if function g is a polynomial matrix function.\\n\\nIn Algorithm 1, what is the meaning of step 14? There is no definition or discussion about G and the \\\\nabla L. The construction function is the key contribution of this paper, which is incorporated the existing SchNet, so it is better to show the advantage of the proposed construction method comparing with SchNet.\\n\\nAlthough this paper is an application-oriented paper, the comparisons with baseline methods are preferred, such as some simple and straightforward baselines. Due to the lack of comparisons, it is hard to understand the statement given by authors that \\u201c the current performance\\u2026 is not optimal and can likely be improved by a better hyperparameter selection\\u201d.\\n\\nIn Section 4, authors calculate the similarity of generated molecules by the closest respective matches, which is determined by the maximal atomic distance after assignment of atom identities and superposition. Is this the standard way to compute the similarity between two graph structures?\"}", "{\"rating\": \"8: Accept\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper is extremely interesting, solid and very well written. The idea is simple but nonetheless developed in a smart and effective fashion. The underlying theory is solid, even if some choices should have been discussed more deeply (e.g. the chosen loss function). Introduction and references are adequate, and the paper is readable by a quite broad audience, despite the detailed technical sections. The main issue related to the manuscript is very narrow target of the experimental part, limited to the isomers of a given compound - it would have been interesting to check its potentialities in generating more different structures and distance matrices, and thus to compare its effectiveness versus alternative generative approaches.\"}" ] }

[ "He Zhao", "Trung Le", "Paul Montague", "Olivier De Vel", "Tamas Abraham", "Dinh Phung" ]

[ "adversarial examples", "perturbations", "spatial distortions", "susceptible", "images", "small perturbations", "classifiers", "spatial distortions perturbations", "adversarial evasion attacks" ]

{ "note_id": [ "pxupoAD8W", "HklsdMjcor", "rJxLyGs9sH", "SylTVgs5sr", "BkgXRYNaYB", "r1e3Y-QTFH", "r1xnXDVnFH", "SklVActsdS", "B1gf_h28Or", "H1er8xkSOS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "comment", "official_comment", "comment" ], "note_created": [ 1576798730405, 1573724787236, 1573724638373, 1573724213400, 1571797450669, 1571791235975, 1571731236362, 1570638539841, 1570323562310, 1570201677262 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1706/Authors" ], [ "ICLR.cc/2020/Conference/Paper1706/Authors" ], [ "ICLR.cc/2020/Conference/Paper1706/Authors" ], [ "ICLR.cc/2020/Conference/Paper1706/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1706/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1706/AnonReviewer3" ], [ "~Vu_Nguyen1" ], [ "ICLR.cc/2020/Conference/Paper1706/Authors" ], [ "~Dimitris_Tsipras1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The method proposed and explored here is to introduce small spatial distortions, with the goal of making them undetectable by humans but affecting the classification of the images. As reviewers point out, very similar methods have been tested before. The methods are also only tested on a few low-resolution datasets.\\n\\nThe reviewers are unanimous in their judgement that the method is not novel enough, and the authors' rebuttals have not convinced the reviewers or me about the opposite.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Many thanks for your precious time and valuable comments.\", \"comment\": \"Our response is as follows:\\n\\n1. Please allow us to re-emphasize the main novelty of our method: We focus on generating adversarial examples that look realistic to humans but also attack the classifier well; We achieve this goal by proposing a generator that conducts both spatial distortions and perturbations. Importantly, the proposed generator is fully differentiable so that we can train it to generate spatial distortions and perturbations jointly. In the joint process, spatial distortions and perturbations are \\u201caware of\\u201d each other and \\u201cwork collaboratively\\u201d, so that we are able to use small spatial distortions plus small perturbations to achieve better attack performance.\\n\\n2. We are conducting experiments on CIFAR and CelebA. We will try our best to report the results in the rebuttal. If the experiments cannot be concluded by the rebuttal deadline, we will report them in the revised paper.\\n\\n3. We have conducted the experiments of adversarial training + PGD, i.e., Adv-Train-PGD. The performance results are shown in the following table. It can be observed that Adv-Train-PGD defends well against perturbation-based methods but is less effective than our proposed approaches. \\n\\n\\n+---------------+--------------------------------+--------------------------------+--------------------------------+--------------------------------+\\n| | MNIST, Model A | MNIST, Model B | Fashion MNIST, Model A | Fashion MNIST, Model B |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\\n| | Adv-Train-FGSM | Adv-Train-PGD | Adv-Train-FGSM | Adv-Train-PGD | Adv-Train-FGSM | Adv-Train-PGD | Adv-Train-FGSM | Adv-Train-PGD |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\\n| No attack | 0.9916 | 0.9915 | 0.9757 | 0.9830 | 0.9057 | 0.9060 | 0.8869 | 0.8854 |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\\n| STM | 0.9481 | 0.4241 | 0.1200 | 0.0413 | 0.1296 | 0.1323 | 0.1112 | 0.1132 |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\\n| SdAdv (ours) | 0.074 | 0.0631 | 0.0744 | 0.08 | 0.1502 | 0.1049 | 0.1420 | 0.1850 |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\\n| FGSM | 0.9481 | 0.9710 | 0.8753 | 0.8189 | 0.8838 | 0.7499 | 0.8558 | 0.6076 |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\\n| PGD | 0.0926 | 0.9427 | 0.0147 | 0.7419 | 0.0764 | 0.6619 | 0.0431 | 0.5140 |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\\n| MIM | 0.1584 | 0.9358 | 0.0373 | 0.7201 | 0.1054 | 0.5987 | 0.0515 | 0.4401 |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\\n| AdvGAN | 0.9278 | 0.9906 | 0.2868 | 0.8680 | 0.1854 | 0.3762 | 0.1015 | 0.1387 |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\\n| SdpAdv (ours) | 0.033 | 0.0752 | 0.0741 | 0.092 | 0.0444 | 0.1113 | 0.0392 | 0.1081 |\\n+---------------+----------------+---------------+----------------+---------------+----------------+---------------+----------------+---------------+\"}", "{\"title\": \"Many thanks for your precious time and valuable comments.\", \"comment\": \"Our responses are as follows:\\n\\n1. Novelty:\\n\\n\\ta. Please note that we did not claim the idea of \\u201cintroducing spatial distortions into adversarial attacks\\u201d as our innovation. Instead, we have proposed a new approach with spatial distortions, which is better than existing spatial distortion based attacks.\\n\\n\\tb. We are aware of the existing study in spatial distortion based attacks; the most related ones to ours are discussed in Sections 2.2 and 3.5 in our paper; more importantly, we have compared our approach to this in our experiments.\\n\\n\\tc. The novelties of our paper compared with others are discussed in Sections 3.5 and please allow us to re-emphasize the main novelty of our method: We focus on generating adversarial examples that look realistic to humans but also attack the classifier well; We achieve this goal by proposing a generator that conducts both spatial distortions and perturbations; Importantly, the proposed generator is fully differentiable so that we can train it to generate spatial distortions and perturbations jointly; In the joint process, spatial distortions and perturbations are \\u201caware of\\u201d each other and \\u201cwork collaboratively\\u201d, so that we are able to use small spatial distortions plus small perturbations to achieve better attack performance.\\n\\n\\td. We undertook a comprehensive search of the related literature of spatial distortion based attacks. However, we may have missed some papers. It would be great if the reviewer could point out any additional related papers to help us improve our paper. We are more than happy to discuss and compare these papers with our paper.\\n\\n2. \\u201cMisleading\\u201d: \\n\\nIt is unfortunate that the reviewer misunderstood our main claims. We did not make any claims that \\u201cwe are the first to use spatial distortions in adversarial attacks.\\u201d On the contrary, we have compared the recent advances in spatial distortions in detail by means of both discussions (see Sections 2.2 and 3.5) and experiments. As the writing style and presentation order can be very subjective, we respectfully disagree on judging our paper by using them as major points.\\n\\n3. Theory:\\n\\nWe will remove the theory as suggested.\\n\\n4. Experiments:\\n\\n\\ta. \\\\gamma controls the magnitude of the spatial distortion. We have conducted experiments using varying \\\\gamma values for the proposed methods, with performance results shown in the following table. Please also note that we have released the code to the community so as to reproduce our results. It can be seen that SdAdv is relatively sensitive to \\\\gamma as it only uses spatial distortions. However, SdpAdv is less sensitive to \\\\gamma as perturbations \\u201care aware of\\u201d spatial distortions and will help attack.\\n\\n+--------+--------+--------------------+--------+--------+\\n| \\\\gamma | 0.1 | 0.3 (in the paper) | 0.5 | 1.0 |\\n+--------+--------+--------------------+--------+--------+\\n| MNIST, Model A, Non-robustified |\\n+--------+--------+--------------------+--------+--------+\\n| SdAdv | 0.7446 | 0.0517 | 0.0352 | 0.0417 |\\n+--------+--------+--------------------+--------+--------+\\n| SdpAdv | 0.034 | 0.0204 | 0.0131 | 0.0130 |\\n+--------+--------+--------------------+--------+--------+\\n| MNIST, Model B, Non-robustified |\\n+--------+--------+--------------------+--------+--------+\\n| SdAdv | 0.3892 | 0.0502 | 0.0501 | 0.0477 |\\n+--------+--------+--------------------+--------+--------+\\n| SdpAdv | 0.0117 | 0.0233 | 0.0223 | 0.0222 |\\n+--------+--------+--------------------+--------+--------+\\n| Fashion MNIST, Model A, Non-robustified |\\n+--------+--------+--------------------+--------+--------+\\n| SdAdv | 0.4213 | 0.1762 | 0.1369 | 0.1285 |\\n+--------+--------+--------------------+--------+--------+\\n| SdpAdv | 0.0221 | 0.0121 | 0.0225 | 0.0224 |\\n+--------+--------+--------------------+--------+--------+\\n| Fashion MNIST, Model B, Non-robustified |\\n+--------+--------+--------------------+--------+--------+\\n| SdAdv | 0.2976 | 0.1079 | 0.1299 | 0.1210 |\\n+--------+--------+--------------------+--------+--------+\\n| SdpAdv | 0.0158 | 0.0075 | 0.0120 | 0.0125 |\\n+--------+--------+--------------------+--------+--------+\\n\\n\\n\\tb. We have conducted experiments with adversarial training + PGD i.e., Adv-Train-PGD. Due to the character limits of our response, please find the results in our response to Reviewer 3.\"}", "{\"title\": \"Many thanks for your precious time and valuable comments.\", \"comment\": \"Our responses are as follows:\\n\\n1. Novelty and related work:\\n\\n\\ta. Our proposed attack falls into the category of \\u201cspatial distortion based attacks\\u201d, which also belongs to a more general research line of \\u201cadversarial attacks for images with visually meaningful transformations\\u201d (e.g. \\u201ccolour-shifting\\u201d in \\u201csemantic adversarial examples\\u201d). Please note that we focus on the category of \\u201cspatial distortion based attacks.'' In this category, we believe that our approach has significant novelties and advantages over others in the same category in both effectiveness and efficiency. This is discussed in Sections 2.2 and 3.5 of our paper.\\n\\n\\tb. Please allow us to re-emphasize the main novelty of our method: We focus on generating adversarial examples that look realistic to humans but also attack the classifier well; We achieve this goal by proposing a generator that conducts both spatial distortions and perturbations; Importantly, the proposed generator is fully differentiable so that we can train it to generate spatial distortions and perturbations jointly; In the joint process, spatial distortions and perturbations are \\u201caware of\\u201d each other and \\u201cwork collaboratively\\u201d, so that we are able to use small spatial distortions plus small perturbations to achieve better attack performance.\\n\\n\\tc. Thanks for pointing out the interesting papers, which we will add as the references for our paper. For those papers, we have the following discussions:\\n\\n\\t\\ti. \\u201cExploring the landscape of spatial robustness\\u201d, ICML19: This paper falls into the same category (spatial distortion based attacks) as ours. Actually, in our discussions and experiments, we compared a spatial distortion method called the \\u201cSpatial Transformation Method (STM)\\u201d, which is implemented in Cleverhans and is exactly the method proposed in \\u201cExploring the landscape of spatial robustness\\u201d. We were unaware of this paper because Cleverhans did not give a reference to this method. We re-summarise the differences between ours and STM: STM only allows translations and rotations while ours allows all kinds of affine transformations; STM uses grid searches to find the optimal translation and rotation while ours is a differentiable method can be jointly trained with perturbation-based methods; STM takes grid searches for every test sample which can be inefficient, while ours only takes a pass of neural networks to conduct attacks (Shown in Table 3 of the paper, where our technique is 3 times faster than STM, even when including perturbations). \\n\\n\\t\\tii. \\u201cCatastrophic Child's Play\\\", CVPR19: This paper only considers random affine transformations with rotations less than 15 degrees while ours optimises the affine transformations according to the loss of the classifier.\\n\\n \\t\\tiii. \\\"Semantic adversarial examples\\\", CVPR18 and \\\"Semantic adversarial attacks\\\", ICCV19: Spatial distortions are not used in these papers, so they might be relatively less related to ours. But we agree that they fall into the general area of \\\"adversarial attacks for images with visually meaningful transformations.\\\"\\n\\n2. Experimental evaluation:\\n\\nWe are conducting experiments on CIFAR and CelebA. We will try our best to report the results in the rebuttal. If the experiments cannot be concluded by the rebuttal deadline, we will report them in the revised paper.\\n\\n3. Weakness of theoretical part:\\n\\nWe will remove the theory as suggested.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper introduces a new approach to generate adversarial examples for deep classifiers. As opposed to the majority of work on adversarial attack models, which generally limit the attacker on pixel-space distortions measured with respect to an Lp norm, the authors here consider a slightly more general attack model that is a combination of an affine transformation and additive L2 perturbation of the input example.\\n\\nFinding optimal attacks for this model can be non-trivial (standard due to the highly nonlinear coupling between the affine parameters and the additive perturbation), so the authors instead propose training a surrogate neural network that generates the attack affine-transformation and distortion- parameters sequentially. This can, in principle, be done in a traditional supervised training setup; however, to force the adversarial images to look perceptually close to natural looking images, the authors throw a discriminator loss on top, and train the attack generator network adversarially.\\n\\nThe paper is well-written in general, the idea is intuitive, and the experiments are well-described. However, I have a few concerns that lead to me to give a low score (at least in the first round of reviews).\\n\\n- Novelty. \\nLeveraging spatial distortions (or other visually meaningful transformations) to generate adversarial attacks is not a new idea, but the authors seem to have been unaware of this very large body of work. See, for example:\\n** Engstrom et al, \\\"Exploring the landscape of spatial robustness\\\", ICML 2019\\n** Poovendran et al, \\\"Semantic adversarial examples\\\", CVPR 2018\\n** Ho et al, \\\"Catastrophic Child's Play\\\", CVPR 2019\\n** Joshi et al, \\\"Semantic adversarial attacks\\\", ICCV 2019\\namong many others.\\n\\nUsing GAN-like transformation models to generate attacks is also not a new idea. A few of the above papers use this approach, and the authors refer to a few other such papers as well.\\n\\nSo as such, the conceptual novelty of the contribution seems to be low (beyond the specific choice of combining affine and L2 perturbations).\\n\\n- Experimental evaluation.\\nThe authors do a commendable job thoroughly laying out the experimental setup. However, a couple of red flags emerge in the experiments. First, why not look at L-infty perturbations (as opposed to L2)? Second, why not test on more challenging datasets (CIFAR, CelebA, etc) as opposed to simple black/white datasets such as MNIST/Fashion-MNIST? One would imagine that the smaller, simpler datasets are easier to optimize for, and therefore the \\\"amortized\\\" attack generator networks are not necessary here.\\n\\n- Weakness of theoretical part.\\nI am not sure the theorem is saying anything strong or useful (since the underlying transformer neural network is assumed to possess infinite capacity). I would suggest just removing it.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper builds upon the work of AdvGAN and proposes to add spatial transformations on top of it. The resulting attacking framework is demonstrated to outperform AdvGAN on attacking several defense approaches, such as Defense-GAN, AdvCritic and adversarial training. Compared to previous approaches on generating spatially transformed adversarial examples, this approaches amortizes the attacking procedure and can produce spatially transformed adversarial examples much faster. This approach also simultaneously combine spatial transformations and perturbations to make the attack stronger.\", \"i_cannot_recommend_acceptance_of_this_paper_because_of_several_reasons\": [\"The idea is not novel enough. It is simply an A + B paper where A = AdvGAN and B = spatial transformer networks. The idea of adversarial attacks with spatial distortion is not the innovation of this paper and has been proposed and extensively studied by many previous papers. This paper does not have additional innovation and does not lead to additional insight that can warrant an acceptance at ICLR.\", \"The general narrative of this paper is misleading. The title seems to indicate this paper is the first to discover the importance of considering spatial perturbations, which is misleading. There is no mention of previous work on spatial transformation attacks in either the abstract nor the introduction (except at the very last). The introduction simply analyzes some well-known phenomenon in the literature, does not place this work well in the literature (even true in the related work section as well), and can mislead readers in believing that this work was the first to realize the importance of spatial transformation attacks.\", \"Theorem 1 is a vacuous statement. It is automatically true based on the universal approximation assumption of neural networks. Including the statement of Theorem 1 is decorative and a waste of space.\", \"The experiments are not convincing. For example, there is no \\\\gamma value reported in the tables. Since \\\\gamma is as important as \\\\epsilon in the proposed attacking method, the missing of this important variate is suspicious. Also the adversarial training only uses FGSM not PGD. The Defense-GAN is already shown not robust by Athalye et al. and cannot be considered as one of the state-of-the-art defenses.\"]}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper proposes a new adversarial attack method by combining spatial transformations with perturbation-based noises. The proposed method uses two networks to generate the parameters of spatial transformation and the perturbation noise. The whole architecture is trained by a variant of GAN-loss to make the adversarial examples realistic to humans. Experiments on MNIST prove that the proposed attack method can improve the success rate of white-box attacks against several models.\\n\\nOverall, this paper considers an important problem of adversarial robustness of classifiers, and present a new approach to craft adversarial examples. The writing is clear. However, I have some concerns about this paper.\\n\\n1. This paper seems to integrate multiple ideas studied before into a single attack method. Perturbation-based adversarial examples, spatial transformation-based adversarial examples, generating adversarial examples based on the GAN loss are all studied before. And the proposed method integrates them together to form a new attack.\\n\\n2. The experiments are only conducted on MNIST and Fashion MNIST. More experiments on CIFAR-10 and ImageNet can further prove the effectiveness of the proposed method.\\n\\n3. More robust defense models should be incorporated in experiments, at least the PGD-based adversarial training model (Madry et al., 2018).\"}", "{\"comment\": \"Great work addressing an important problem in deep learning !\\n\\nI would like to mention a closely related work [1] that uses Bayesian optimisation (BO) to sequentially suggest an attack to make the deep models failed. The attacks used in [1] has considered different ways of affine transformation including translation, rotation, shearing. Using BO in [1] will be sample-efficient in making the attacks. That is, the targetted deep model will be failed under less number of attacks. \\n\\nThe problem definition is quite similar. That is, the objective function defined in Eq1, used to select the attack $x_A$, in the current paper is similar to objective function defined in Eq1 in [1]. Having said that, the search spaces considered in two papers are different. Particularly, the Eq1 in [1] considered the parameter space of the attacks (such as the shearing parameters) while Eq1 in this paper considered the L2-ball in the raw feature space.\\n\\n[1] Gopakumar, Shivapratap, et al. \\\"Algorithmic assurance: an active approach to algorithmic testing using Bayesian optimisation.\\\" Advances in Neural Information Processing Systems. 2018.\", \"nb\": \"The source code is also available.\", \"title\": \"Related work in generating distortions under affine transformation\"}", "{\"comment\": \"Dear Dimitris,\\n\\nThanks a lot for pointing out this interesting and related paper, which we weren't aware of at the time of working on our submission.\\n\\nAfter a quick look at this ICML19 paper, we find that in terms of attacks, there are several differences between ours and the ICML19 one. For example, the ICML19 one conducts rotation and translation to generate an adversarial example by doing optimisation given a test sample, while ours uses an amortized way, which learns a neural network to conduct rotation, translation, scaling, and shear. Moreover, ours is a joint amortized process (with two neural networks), that combines spatial distortions and perturbations, aiming to generate realistic adversarial examples with fewer perturbations. While the ICML19 one seems to use two separate steps for PGD and grid search for rotations and translations, respectively. We will discuss more on differences as well as connections between the two attacks in the updated version of our paper.\\n\\nWe also appreciate that the code of the ICML19 paper is released. Therefore, we are currently doing a comparison with the attack and defence methods introduced in the ICML19 paper and will provide a detailed discussion in the updated version of our paper.\\n\\nBesides, thanks for pointing out the BMVC paper as well, which will be added to our reference.\\n\\nThanks again,\\nPaper 1706 Authors\", \"title\": \"Thanks for the reference\"}", "{\"comment\": \"I wanted to bring to your attention our work studying spatial distortions---rotations and translations---and their combination with standard pixel-based perturbations: \\\"Exploring the landscape of spatial robustness\\\" (ICML'19, https://arxiv.org/abs/1712.02779). Specifically, we consider an adversary that tries all possible spatial transformations (through exhaustive grid search) and then applies a standard PGD attack on top.\\n\\nAlso, I believe it is worth mentioning \\\"Manitest: are classifiers really invariant?\\\" (BMVC'15, https://arxiv.org/abs/1507.06535) which---to the best of my knowledge---is the first work studying the robustness of deep networks to spatial transformations.\", \"title\": \"Prior work on combining spatial distortions with pixel-wise perturbations\"}" ] }

[ "Mohak Bhardwaj", "Ankur Handa", "Dieter Fox", "Byron Boots" ]

[ "entropy regularized reinforcement learning", "information theoretic MPC", "robotics" ]

{ "note_id": [ "8P7NpCs8z5", "HyxcBAg2sH", "r1g_AgcooH", "H1xwr1V9iB", "HJeCDWEtsS", "rygnPvmFoS", "ryeVLDmYiS", "r1gSAHXFjB", "HyeB2rXYoB", "S1gNUGQKjr", "Hye5ppju5H", "HJgk1myAFr", "B1gjSMattS" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730376, 1573813826388, 1573785808098, 1573695294803, 1573630309888, 1573627748443, 1573627723986, 1573627341508, 1573627309263, 1573626444326, 1572548033977, 1571840726761, 1571570242821 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1705/Authors" ], [ "ICLR.cc/2020/Conference/Paper1705/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1705/Authors" ], [ "ICLR.cc/2020/Conference/Paper1705/Authors" ], [ "ICLR.cc/2020/Conference/Paper1705/Authors" ], [ "ICLR.cc/2020/Conference/Paper1705/Authors" ], [ "ICLR.cc/2020/Conference/Paper1705/Authors" ], [ "ICLR.cc/2020/Conference/Paper1705/Authors" ], [ "ICLR.cc/2020/Conference/Paper1705/Authors" ], [ "ICLR.cc/2020/Conference/Paper1705/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1705/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1705/AnonReviewer4" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The authors develop a novel connection between information theoretic MPC and entropy regularized RL. Using this connection, they develop Q learning algorithm that can work with biased models. They evaluate their proposed algorithm on several control tasks and demonstrate performance over the baseline methods.\\n\\nUnfortunately, reviewers were not convinced that the technical contribution of this work was sufficient. They felt that this was a fairly straightforward extension of MPPI. Furthermore, I would have expected a comparison to POLO. As the authors note, their approach is more theoretically principled, so it would be nice to see them outperforming POLO as a validation of their framework.\\n\\nGiven the large number of high-quality submissions this year, I recommend rejection at this time.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Author response\", \"comment\": \"1. The wall clock times depend on the particular implementation of the MPC algorithm but in the case of MPPI they are generally a function of the number of particles*planning horizon which can prohibit use of larger horizon especially in real-time implementation. Additional reasons for preferring shorter horizon are the difficulty of optimizing for longer action sequences and compounding model errors.\\n\\n2. Fig 1b-d show number of successes which is a binary quantity and hence does not error bars. We will include multiple seed results in final version.\"}", "{\"title\": \"Reviewer response\", \"comment\": \"Thanks for the detailed response! This clears up many of my concerns with the paper. I have two minor experimental questions about your response.\\n\\n> \\\" This demonstrates that MPQ allows us to significantly reduce the MPC horizon which is the major computational bottleneck in MPC and needs to be highly tuned in practice\\\"\\n1. What are the relative wall clock times of the various methods? If a longer horizon causes substantially larger wall clock time, that would help convince me that avoiding long-horizon MPC is an important aim.\\n\\n2. Can you add error bars to Fig 1b-d (or plot multiple random seeds)?\"}", "{\"title\": \"Update regarding experimental results\", \"comment\": \"We would like to bring to the attention of all reviewers and meta-reviewers an update to the experimental results in the current version. We have updated results for the following two tasks:\\n\\n1. FetchPushBlock\\nWe fixed a bug in our previous implementation and re-trained the MPQ and SoftQ for this task. We added comparison against MPPI with H=64 and observe that MPQ with H=10 is able to outperform it within a few episodes of training. (Refer Figure 1 in revision).\\n\\n2. FrankaCabinetEnv\\nWe added comparison against MPPI with H=64 and observe that the performance of MPQ with H=10 is still considerably better.\\n\\nBoth these results further support the method presented in the paper. We have updated our responses below to match these new numbers and would be glad to answer any questions that the reviewers may have.\"}", "{\"title\": \"Additional experiment and summary of responses\", \"comment\": \"We would like to thank all reviewers for providing their valuable feedback on our work. We first provide details of an additional benchmark task that we have added to the revision\\n\\nFrankaDrawerOpen\\nThe agent velocity controls a 7DOF Franka Panda arm to open a drawer on a cabinet. A simple cost function based on Euclidean distance and relative orientation of the end effector with respect to the handle and the displacement of the slide joint on the drawer is used. The agent is provided a biased distribution over damping and frictionloss of robot and drawer joints. Every episode is 4 seconds long after which the agent's starting configuration is randomized. Success corresponds to opening the drawer within 1cm of target displacement.\\n\\nThis task is based on a real-world manipulation problem from [1] and is a high-dimensional continuous control task based on MuJoCo. In order to solve it successfully, the agent requires long horizon planning in 7DOF to calculate a trajectory for the arm to reach the handle and open the cabinet and is hard to solve for model-free RL or MPPI with a short horizon. We use this task to provide further evidence for MPQ\\u2019s ability to overcome the shortcomings of both MPPI and soft Q-learning.\", \"results\": \"As shown in Figure 1, MPQ with a horizon H=10 outperforms MPPI with the same horizon and achieves five times the success rate by the end of training. It also performs considerably better than MPPI with H=64. This demonstrates that MPQ allows us to truncate the MPC horizon even in high-dimensional problems. We also observe that soft Q-learning is unable to solve any problem. This can be attributed to the general sample inefficiency of model-free RL which is further aggravated in high-dimensional control tasks. \\n\\nWe now provide a summary of responses to the concerns raised by reviewers\\n******\\nNovelty\\n******\\nThe paper provides a principled way to solve the infinite horizon entropy-regularized RL problem and is the first to our knowledge to derive the important connection between information theoretic MPC and entropy-regularized RL. The resulting algorithm is a natural combination of MPPI and soft Q -earning that can overcome the shortcomings of both. While previous approaches such as [2,3] have attempted to combine MPC with learned value functions ([3] even uses MPPI in their implementation), they do not explore this connection and hence are not solving the correct problem in principle. In addition, our experiments provide empirical evidence that Q-functions learned from interactions with the true system can help mitigate the effects of model-bias and add global information allowing us to use a shorter horizon in practice. Using MPC for calculating Q-value targets helps in faster convergence by providing more stable targets early on. This reflects in our experimental results as MPQ learns a competitive policy in few training iterations and considerably outperforms soft Q-learning. We further test the applicability of MPQ in sim-to-real transfer by comparing it to dynamics randomization.\\n\\n*********************\\nExperiments and Baselines\\n*********************\\nWe considered continuous control tasks that require long horizon planning to solve efficiently. The BallInCupSparse, FetchPushBlock and FrankaDrawerOpen are hard tasks and better represent real-world robotics problems than standard MuJoCo benchmarks like Ant or HalfCheetah.\", \"we_compare_against_three_baselines\": \"MPPI using same horizon as MPQ (eg. MPQH10 and MPPIH10 in FetchPushBlock), MPPI using a longer horizon and soft Q-learning. These are natural baselines and in all tasks, including 4DOF FetchPushBlock and 7DOF FranksDrawerOpen and the sparse reward BallInCupSparse, MPQ consistently outperforms them. We consider this as strong evidence to prove its robustness, scalability and sample efficiency.\\n\\nReviewer 2 and 4 wished to see comparison against model-based RL methods. However, as elaborated in the individual replies, we believe that learning accurate system models is an interesting but entirely different problem and is not within the scope of the current work. In fact, we would like to note that MPQ is a complementary approach to model learning and one can benefit from the other. \\n\\nWe hope that we have satisfactorily addressed all the reviewers' concerns and they would consider updating their ratings. We are looking forward to more fruitful discussions.\", \"references\": \"[1] Y. Chebotar, A. Handa, V. Makoviychuk, M. Macklin, J. Issac, N. Ratliff, and D. Fox. Closing the sim-to-real loop: Adapting simulation randomization with real world experience. In ICRA, 2019 \\n[2] Kendall Lowrey, Aravind Rajeswaran, Sham Kakade, Emanuel Todorov, and Igor Mordatch. Plan online, learn offline: Efficient learning and exploration via model-based control.arXiv:1811.01848, 2018\\n[3] M. Zhong, M. abd Johnson, Y. Tassa, T. Erez, and E Todorov. Value function approximation and model predictive control. IEEE ADPRL, 2013.\"}", "{\"title\": \"Response to review (2/2)\", \"comment\": \"************************************\\nResponses to Minor Comments\\n************************************\\n\\u201cThe derivation of the update rule in Section 4.2 was unclear to me. In Equation 17, why can the RHS not be computed directly? In Equation 21, where did this iterative procedure come from?\\u201d\\n\\nEq. 17: In order to directly compute the optimal actions it is required to estimate the expectation with respect to the optimal control distribution defined in Eq. 8. Since this is intractable to sample from directly, we employ importance sampling to convert the optimal action sequence into an expectation over a simpler sampling distribution such as a Gaussian.\\n\\nEq. 21: The iterative procedure is a result of the fact that we are using finite number of samples to estimate the importance sampling weights. If we had infinite compute, we would not require any iterative procedure and could estimate the optimal controls in one shot.\\n\\nWe will make these assertions clear in the final revision.\\n\\n* Eq 1 -- The \\\"\\\\forall s_0\\\" suggests that this equation depends on s_0. Can you clarify this dependency?\\n\\nThis implies that the policy should minimize the cost-to-go for every state in the state space i.e it must be globally optimal.\\n\\nWe would like to thank the reviewer for pointing out all the typos, places where the language could be improved and relevant citations. We will include these in the revision.\\n\\nReferences\\n[1] Stephane Ross and J Andrew Bagnell. Agnostic system identification for model-based reinforcement learning. arXiv preprint arXiv:1203.1007, 2012.\\n[2] Kendall Lowrey, Aravind Rajeswaran, Sham Kakade, Emanuel Todorov, and Igor Mordatch. Plan online, learn offline: Efficient learning and exploration via model-based control. arXiv preprint arXiv:1811.01848, 2018\\\\\"}", "{\"title\": \"Response to review (1/2)\", \"comment\": \"We would like to thank the reviewer for the valuable feedback. We agree with the comment that this is an interesting area of research and believe it requires further attention from the RL and controls community. We address your specific concerns below\\n\\n*************************\\nExperiments\\n*************************\\n\\n\\u201cIn Figure 1, it seems like the proposed method (MPQ) only beats that MPPI baseline in BallInCupSparse.\\u201d\\n\\nWe would like to clear up the confusion regarding the performance of MPQ with respect to baselines. Figure 1 shows the comparison among MPQ, MPPI and Soft Q-Learning during training. In all the tasks, MPQ outperforms MPPI when both use the same horizon after only a few episodes (eg. In FetchPushBlock MPPIH10 versus MPQH10). The results also show that MPQ ultimately performs better than MPPI, even when MPPI uses a much longer horizon. In the Pendulum task, MPQ with horizon (H) = 8 achieves an average reward higher than MPPI with H=32 at the end of the training. In FetchPushBlock, MPQ with H=10 obtains a success rate higher than MPPI with H=64. This demonstrates that MPQ allows us to significantly reduce the MPC horizon which is the major computational bottleneck in MPC and needs to be highly tuned in practice. While a longer horizon generally helps in MPC, optimizing for longer action sequences can be harder and model errors can compound over timesteps. MPQ also outperforms the soft Q learning baseline in all tasks.\\n\\nWe observe similar results in the new FrankaDrawerOpen task, where MPQ with H=10 outperforms MPPI with H=64 after only a few episodes of training demonstrating that the approach is scalable to high-dimensional tasks as well. Please refer to overall comments and paper revision for details of the task.\\n\\nWe believe that the consistent results across all different tasks sufficiently demonstrate the efficacy of the approach.\\n\\n\\u201cMoreover, the tasks considered are all quite simple. \\u201c\\n\\nThe three benchmarks, in addition to the Pendulum task, that we used to evaluate our method are standard tasks for real-world manipulation problems that require planning to solve successfully. We believe that these tasks are harder and more realistic proxies for real-world robotics problems than MuJoCO benchmarks such as Ant or HalfCheetah. Please refer to responses to Reviewer 1 and Reviewer 2 for more details.\", \"comparison_against_model_based_rl\": \"Learning globally consistent neural network models using approaches such as [Chua 2018, Kurutach 2018, Janner 2019] mentioned by the reviewer is indeed an interesting research problem. The problem is inherently hard to solve primarily because of covariate shift between the training and test distributions i.e the distribution of states visited by the controller at test time can be significantly different from the training distribution [1]. Therefore, in this work we focus on learning a policy that can rapidly adapt to the true system dynamics. Additionally, as shown in [2] even when the true model of the system is available, learning a terminal value function for MPC can significantly improve performance and reduce planning horizon. This implies that our approach is in-fact complementary to Model-based RL and can benefit from more accurate models. However, model learning is beyond the scope of the current work and thus we chose not to compare against the above mentioned algorithms.\"}", "{\"title\": \"Response to review (2/2)\", \"comment\": \"**************\\nMinor Comments\\n**************\\n\\u201c\\\"Solving the above optimization can be prohibitively expensive and hard to accomplish online.\\\"\\n\\n- This is not true. There are already plenty of work solving the entropy-regularized MDP online [2, 3, 4] and achieving good empirical performance. \\u201c\\n\\nThank you for pointing out the relevant work on entropy-regularized RL. We would first like to clarify what we mean by optimizing the policy online. MPC optimizes a finite-horizon sequence of controls calculate the optimal action at every timestep on the system. This results in a state-feedback policy. Efficient MPC approaches are able to perform the optimization in real-time by using simple policy classes, predictive models, truncated horizon optimization and leveraging parallel hardware such as GPUs [2, 4, 9]. Most model-free RL approaches, such as the ones pointed out by the reviewer, use policy gradient or value based learning algorithms to optimize complex policy classes such as deep neural networks. Optimizing them online would be akin to finding the best neural network weights at the current state. Doing so can be prohibitively expensive given the large number of parameters of the network. Hence, these approaches mostly use an iterative process of collecting data from the system and performing SGD on a loss function offline i.e between iterations of interaction. In general, if we assume access to an accurate dynamics model, we could solve for the optimal policy using dynamic programming techniques. However, such techniques suffer from the curse of dimensionality and often need additional assumptions on the form of the dynamics and cost function. Also, exact models are usually not available for real-world robotics problems like manipulation unlike games such as Atari or Go. Efficiently combining model-free and model based approaches can help mitigate these shortcomings which forms the motivation for the work. \\n\\n\\u201c\\\"The re-optimization and entropy regularization helps in mitigating model-bias and\\ninaccuracies with optimization by avoiding overcommitment to the current estimate of the cost.\\\"\\n- There is no evidence that the entropy-regularization will reduce the mode-bias. Actually, as discussed in [4, 5] the entropy regularization will also incur some extra bias.\\u201d\\n \\nWe did not intend to make the claim that entropy-regularization will directly reduce model-bias. We will correct this in the final revision. We intended to say that entropy regularization helps in maintaining exploratory behavior and prevents over-optimizing the current estimate of the cost which is biased due to model-errors, finite horizon and limited number of samples. \\n\\nFinally, we would like to thank the reviewer for pointing out typographical errors. We will correct them in the revision. We hope we have sufficiently addressed the reviewers concerns and would be glad to provide any additional clarifications.\", \"references\": \"[1] Roy Fox, Ari Pakman, and Naftali Tishby. Taming the noise in reinforcement learning via soft updates. arXiv preprint arXiv:1512.08562, 2015\\n[2] Nolan Wagener, Ching-An Cheng, Jacob Sacks, Byron Boots. An Online Learning Approach to Model-Predictive Control. Robotics:Science and Systems, 2019\\n[3] Pieter Abbeel, Adam Coates, and Andrew Y. Ng. Autonomous helicopter aerobatics through apprenticeship learning. The International Journal of Robotics Research, 2010.\\n[4] Grady Williams, Nolan Wagener, Brian Goldfain, Paul Drews, James M Rehg, Byron Boots, and Evangelos A Theodorou. Information theoretic mpc for model-based reinforcement learning. In 2017 IEEE International Conference on Robotics and Automation (ICRA)\\n[5] Wen Sun, J Andrew Bagnell, and Byron Boots. Truncated horizon policy search: Combining reinforcement learning & imitation learning. arXiv preprint arXiv:1805.11240, 2018.\\n[6] Kendall Lowrey, Aravind Rajeswaran, Sham Kakade, Emanuel Todorov, and Igor Mordatch. Plan online, learn offline: Efficient learning and exploration via model-based control. arXiv preprint arXiv:1811.01848, 2018\\n[7] M. Zhong, M. abd Johnson, Y. Tassa, T. Erez, and E Todorov. Value function approximation and model predictive control. IEEE ADPRL, 2013.\\n[8] U. Rosolia, F. Borrelli, \\\"Learning model predictive control for iterative tasks. a data-driven control framework\\\", IEEE Transactions on Automatic Control, 2017.\\n[9] Grady Williams, Paul Drews, Brian Goldfain, James M Rehg, and Evangelos A Theodorou. Aggressive driving with model predictive path integral control. In 2016 IEEE International Conference on Robotics and Automation (ICRA).\\n[10] Y. Chebotar, A. Handa, V. Makoviychuk, M. Macklin, J. Issac, N. Ratliff, and D. Fox. Closing the sim-to-real loop: Adapting simulation randomization with real world experience. In 2019 IEEE International Conference on Robotics and Automation (ICRA).\"}", "{\"title\": \"Response to review(1/2)\", \"comment\": \"Thank you for investing the time to review our paper and providing valuable feedback. We address your questions and concerns below\\n\\n******\\nNovelty\\n******\\nWe are interested in solving the infinite horizon entropy-regularized RL problem. This objective helps in incorporating prior knowledge, maintaining exploratory behavior and mitigating the biased update rule in vanilla Q-learning [1]. Model-free RL methods are difficult to implement on real-world systems because of sample inefficiency and safety concerns. In contrast, model-based approaches like MPC bypass these concerns and consistently work well in the real world [2, 3, 4]. The paper focuses on leveraging MPC to solve the infinite horizon entropy-regularized objective. While several previous works have demonstrated the utility of using a Q function or value function as a terminal cost in the MPC objective [6, 7, 8], our paper is the first to show that the principled way to do this is to use a soft Q function. In particular, previous work such as [6], which combines MPPI and, e.g., a vanilla Q function, is not consistent with either the standard infinite horizon RL problem or the entropy regularized RL problem. The key insight in our work is the important connection between entropy regularized RL and information theoretic MPC and the accompanying derivation that shows how methods like MPPI and soft Q learning should be combined. \\n\\nA few key differences as compared to MPPI [2] and Q learning:\\nMPPI\\nIn Section 4.1 we provide a more straightforward derivation for the optimal sampling distribution as compared to Williams et. al. 2017 who simply assume the expression for the free energy without any justification or motivation and use Jensen\\u2019s inequality to upper bound it by the entropy regularized objective. In contrast, our derivation starts from first principles and uses the original infinite horizon entropy-regularized objective to derive the free-energy and optimal policy making the connection between MPPI and soft RL extremely clear. Since the optimal policy is intractable to sample from, we follow the approach of [2] to find the optimal policy from a class of simpler Gaussian policies as presented in Section 4.2. In the original draft we have cited Williams et. al 2016 instead of Williams et. al 2017 [9] accidentally and we will correct this in the revision.\\n\\nQ-Learning\\nApproaches such as DQN are entirely model-free. However, in several real-world robotics problems efficient simulators are available [eg. MuJoCO] that can be leveraged to obtain better Q-value estimates. Previous work in RL such as [5, 6] has demonstrated that given an accurate model of the system, H-step optimization can improve an inaccurate value function. Similar to [6], MPQ leverages MPC not only as the policy executed on the system, but also as a critic, thus making it \\u201cModel-Predictive Q Learning\\u201d. The following are key distinguishing features from [4]: (1) Our mathematical connection between information theoretic control and soft RL theoretically justifies using MPPI as the control algorithm and free-energy targets for the Q-function. [4] doesn\\u2019t explore this connection and uses hard value target despite using MPPI in the implementation. (2) [4] assumes access to an accurate dynamics model whereas in our experiments we explicitly consider model-bias and demonstrate the robustness of MPQ. \\n\\n*****************\\nEmpirical Comparison\\n*****************\\n\\nWe compared our approach to several different baselines on three different environments. The purpose of the experiments was to investigate the extent to which learning a soft Q function improves over MPC with potentially biased models on realistic robotics manipulation problems. We believe that three benchmarks are sufficient to support and provide insight into the methods and theory detailed in the paper. However, we have now included an additional benchmark, which further supports our conclusions. Please refer to overall comments and revision for details of the task.\\n\\nThe three benchmarks, in addition to the Pendulum task, that we used to evaluate our method are standard tasks for real-world manipulation problems that require planning to solve successfully. We believe that these tasks are harder and more realistic proxies for real-world robotics problems than MuJoCO benchmarks such as Ant or HalfCheetah. Please see response to reviewer 2 for more details.\\n\\nThe BallInCupSparse and FetchPushBlock environments are in fact continuous control tasks that use MuJoCo. As mentioned in Section 5, the environment for the BallInCupSparse task is provided in the DeepMind Control Suite (github.com/deepmind/dm_control) and the one for FetchPushBlock task is available as part of OpenAI Gym\\u2019s robotics environments (github.com/openai/gym/tree/master/gym/envs/robotics). The PendulumSwingup environment however uses equations of motion and is not based on MuJoCo. The FrankaDrawerOpen environment is also based on a real-world manipulation task from [10].\"}", "{\"title\": \"Response to review\", \"comment\": \"Thank you for taking the time to review our paper and providing valuable feedback. We greatly appreciate the fact that you found the paper well-written and the derivation easy to follow. We address your questions and concerns below\\n**********\\nExperiments\\n**********\\n\\u201cFor experiments, I'd like to see some results on more complex environments (e.g., continuous control tasks in OpenAI Gym) and more comparison with recent model-based RL work. \\u201c\\n\\nFirst, we clarify that all the tasks considered in the paper are continuous control tasks. The FetchPushBlock environment is a part of OpenAI Gym\\u2019s MuJoCo robotics environments (github.com/openai/gym/tree/master/gym/envs/robotics). BallInCupSparse is based off of the environment provided by DeepMind Control Suite (github.com/deepmind/dm_control) that also uses MuJoCo. \\n\\nWhile designing the experiments our aim was to test the efficacy of MPQ in overcoming the shortcomings of both MPC and model-free Q learning. We considered tasks with sparse rewards and that require long horizon planning. The complexity is further aggravated by biased distribution over dynamics parameters making it hard for MPC to solve. By learning a terminal value function from real data we posit that MPQ should adapt to the true system dynamics as well as truncate the horizon of MPC which is the main computational bottleneck. By using a predictive model for Q learning, we also expect to be able to learn an effective value function with significantly less data as compared to entirely model-free soft Q-learning. Hence, we compare against natural baselines of MPPI and soft Q-learning. \\n\\nResults on all tasks provide evidence in favor of these hypotheses (Figure 1). In all the tasks, MPQ outperforms MPPI that uses the same horizon after only a few training iterations and ultimately performs better than MPPI using a much longer horizon. The BallInCupSparse experiment demonstrates that MPQ is able to discover the sparse rewards by planning into the future, thus mitigating the need to design a dense reward function. FetchPushBlock requires long planning horizon to figure out a trajectory for the arm that can get the block into the goal location. Using a very simple reward function of distance between block and goal, we observed that MPPI with a short horizon (H=10) is unable to even reach close to the block. MPQ with H=10 is able to achieve a success rate higher than MPPI with a significantly longer horizon (H=64) in only a few minutes of experience gathered on the true system. \\n\\nWe have additionally added another benchmark experiment called FrankaDrawerOpen which is a high-dimensional continuous control problem based on a real-world manipulation task. We observe similar results in this task which further strengthens the arguments in the paper. Please refer to overall comments and paper revision. \\n\\nWe do not compare against model based RL methods that learn neural network models for control as learning globally consistent neural network models adds an additional layer of complexity and is beyond the scope of the current work. We would like to note that MPQ is a complementary approach to model learning and one can benefit from the other. \\n\\n********\\nQuestions\\n********\\nQ. The experiment setup is that a uniform distribution of dynamics parameters (with biased mean and added noise) are used. Why not using a neural network?\\n\\nPhysics based simulators such as MuJoCo provide globally consistent dynamics models and are widely used in controls [1] and in sim-to-real approaches [2]. However quantities such as friction coefficients, tensile strength, mass and inertia are hard to estimate accurately. Thus, we chose to randomize them as a reasonable approximation of model bias as often done in dynamics randomization [2,3]. MPQ can be used with neural network dynamics models and learning such models has been the part of significant prior work, but is orthogonal to the problem we are attempting to solve. \\n\\nQ2. Eq (22): The estimation for eta is unbiased, However, 1/eta might be biased, so is the estimator for w(E) also biased? \\n\\nYes, since we use importance sampling, the estimate of the weights will be biased for a finite number of samples, but is asymptotically unbiased. \\n\\nWe would like to thank the reviewer for pointing out the typographical errors. We will correct these in the final revision. We hope that we have sufficiently addressed the reviewer\\u2019s concerns and would be happy to answer any further questions.\\n\\nReferences\\n[1] Tassa, Yuval, Erez, Tom, and Todorov, Emanuel. Synthesis and stabilization of complex behaviors through online trajectory optimization. In Intelligent Robots and Systems (IROS), 2012\\n[2] X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic control with dynamics randomization. CoRR, abs/1710.06537, 2017\\n[3] I. Mordatch, K. Lowrey, and E.Todorov. Ensemble-CIO: Full-body dynamic motion planning that transfers to physical humanoids. In IROS, 2015.\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"In this paper, the authors proposed the algorithm to introduce model-free reinforcement learning (RL) to model predictive control~(MPC), which is a representative algorithm in model-based RL, to overcome the finite-horizon issue in the existing MPC. The authors evaluated the algorithm on three environments and demonstrated the outperformance comparing to the model predictive path integral (MPPI) and soft Q-learning.\\n\\n1, The major issue of this paper is its novelty. The proposed algorithm is a straightforward extension of MPPI [1], which adds a Q-function to the finite accumulated reward to predict the future rewards to infinite horizon. The eventual algorithm ends up like a straightforward combination of MPPI and DQN. The algorithm derivation in Sec. 4 follows almost exact as [1] without appropriate reference. \\n\\n2, The empirical comparison is quite weak. The algorithm is only tested on three environments and only one (Pendulum) from the MuJoCo benchmark. Without more comprehensive comparison on other MuJoCo environments, the empirical experiment is not convincing.\", \"minors\": \"1, The derivation from 15-16 is wrong. The first term should be negative.\\n\\nThere are many claims without justification. For example:\\n\\\"Solving the above optimization can be prohibitively expensive and hard to accomplish online.\\\"\\n\\n- This is not true. There are already plenty of work solving the entropy-regularized MDP online [2, 3, 4] and achieving good empirical performance.\\n\\n\\\"The re-optimization and entropy regularization helps in mitigating model-bias and\\ninaccuracies with optimization by avoiding overcommitment to the current estimate of the cost.\\\"\\n- There is no evidence that the entropy-regularization will reduce the mode-bias. Actually, as discussed in [4, 5] the entropy regularization will also incur some extra bias. \\n\\n[1] Grady Williams, Nolan Wagener, Brian Goldfain, Paul Drews, James M Rehg, Byron Boots, and Evangelos A Theodorou. Information theoretic mpc for model-based reinforcement learning. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pp. 1714\\u20131721. IEEE, 2017.\\n[2] Haarnoja, Tuomas, Aurick Zhou, Pieter Abbeel, and Sergey Levine. \\\"Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor.\\\" arXiv preprint arXiv:1801.01290 (2018).\\n[3] Nachum, O., Norouzi, M., Xu, K. and Schuurmans, D., 2017. Bridging the gap between value and policy based reinforcement learning. In Advances in Neural Information Processing Systems (pp. 2775-2785).\\n[4] Dai, Bo, Albert Shaw, Lihong Li, Lin Xiao, Niao He, Zhen Liu, Jianshu Chen, and Le Song. \\\"SBEED: Convergent reinforcement learning with nonlinear function approximation.\\\" arXiv preprint arXiv:1712.10285 (2017).\\n[5] Geist, Matthieu, Bruno Scherrer, and Olivier Pietquin. \\\"A Theory of Regularized Markov Decision Processes.\\\" arXiv preprint arXiv:1901.11275 (2019).\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper builds a connection between information theoretical MPC and entropy-regularized RL and also develops a novel Q learning algorithm (Model Predictive Q-Learning). Experiments show that the proposed MBQ algorithm outperforms MPPI and soft Q learning in practice.\\n\\nThe paper is well-written. The derivation is clean and easy to understand. Adding a terminal Q function, which makes horizon infinite, is a very natural idea. Most of the derivation is unrelated to the additional Q^pi^* (Eq 11-17) and Eq (18-20) are simply plugging in the formula of pi^*, which is derived at Eq (9), which is also quite natural if one knows the result for finite horizon horizon MPPI (i.e., without the terminal Q). For experiments, I'd like to see some results on more complex environments (e.g., continuous control tasks in OpenAI Gym) and more comparison with recent model-based RL work.\", \"questions\": \"1. The experiment setup is that a uniform distribution of dynamics parameters (with biased mean and added noise) are used. Why not using a neural network? \\n2. Eq (22): The estimation for eta is unbiased, However, 1/eta might be biased, so is the estimator for w(E) also biased?\", \"minor_comments\": \"1. Eq (16): should the first term be negated? \\n2. Page 6, last paragraph: performved -> performed.\\n3. Page 7, second paragraph: produces -> produce.\\n4. Algorithm 1, L6: N % N_update -> i % N_update. \\n5. Appendix A.1, second paragraph: geneerating -> generating.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper studies how a known, imperfect dynamics can be used to accelerate policy search. The main contribution of this paper is a model-based control algorithm that uses n-step lookahead planning and estimates the value of the last state with the prediction from a learned, soft Q value. The n-step planning uses the imperfect dynamics model, which can be cheaply queried offline. A second contribution of the paper is an efficient, iterative procedure for optimizing the n-step actions w.r.t. the imperfect model. The proposed algorithm is evaluated on three continuous control tasks. The proposed algorithm outperforms the two baselines on one of the three tasks. Perhaps the most impressive aspect of the paper is that the experiments only require a few minutes of \\\"real-world\\\" interaction to solve the tasks.\\n\\nI am leaning towards rejecting this paper. My main reservation is that the empirical results are not very compelling. In Figure 1, it seems like the proposed method (MPQ) only beats that MPPI baseline in BallInCupSparse. The paper seems remiss to not include comparisons to any recent MBRL algorithms (e.g., [Chua 2018, Kurutach 2018, Janner 2019]). Moreover, the tasks considered are all quite simple. Finally, for a paper claiming to \\\"pave the way for deploying reinforcement learning algorithms on real-robots,\\\" it seems important to show results on real robots, or at least on a very close approximation thereof. Instead, since the experiments are done in simulation and the true model is known exactly, the paper constructs an approximate model that is a noisy version of the true model. \\n\\nI do think that this paper is tackling an important problem, and am excited to see work in this area. I would consider increasing my review if the paper were revised to include a comparison to a state-of-the-art MBRL method, if it included experiments on more complex task, and if the proposed method were shown to consistently outperform most baselines on most tasks.\", \"minor_comments\": [\"The derivation of the update rule in Section 4.2 was unclear to me. In Equation 17, why can the RHS not be computed directly? In Equation 21, where did this iterative procedure come from?\", \"\\\"Reinforcement Learning\\\" -- No need to capitalize\", \"\\\"sim-to-real approaches \\u2026 suffer from optimization challenges such as local minima and hand designed distributions.\\\" -- Can you provide a citation for the \\\"local minima\\\" claim? The part about \\\"hand designed distributions\\\" seems to have a grammar error.\", \"\\\"Model Predictive Control\\\" -- No need to capitalize.\", \"\\\"Computer Go\\\" -- How is this different from the board game Go?\", \"\\\"Entropy-regularized\\\" -- Missing a period before this.\", \"Eq 1 -- The \\\"\\\\forall s_0\\\" suggests that this equation depends on s_0. Can you clarify this dependency?\", \"\\\"stochastic gradient descent\\\" -- I believe that [Williams 1992] is a relevant citation here.\", \"----------------------UPDATE AFTER AUTHOR RESPONSE------------------------\", \"Thanks for the detailed response! The new experiments seem strong, and help convince me that the method isn't restricted to simple tasks. I'm inclined to agree that it's not overly onerous to assume access to an imperfect model of the world. Because of this, I will increase my vote to \\\"weak accept.\\\"\", \"In the next version, I hope that the authors (1) include multiple random seeds for all experiments, and (2) student the degree to which model misspecification degrades performance (i.e., if you use the wrong model, how much worse do you do?).\"]}" ] }

[ "Zhe Dong", "Deniz Oktay", "Ben Poole", "Alexander A. Alemi" ]

[ "rnns", "predictive information", "future", "remarkable capability", "history", "sensory inputs", "informative", "work", "artificial neurons" ]

{ "note_id": [ "aN2tKg-2-_", "SklLK06ijH", "B1l--0aioB", "r1x0Ja6isB", "r1l3FOGR9B", "Hyxmmk9kcS", "rkxNUnoftB" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730344, 1573801597803, 1573801465215, 1573801189767, 1572903043887, 1571950362622, 1571105867716 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1704/Authors" ], [ "ICLR.cc/2020/Conference/Paper1704/Authors" ], [ "ICLR.cc/2020/Conference/Paper1704/Authors" ], [ "ICLR.cc/2020/Conference/Paper1704/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1704/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1704/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"Nice start but unfortunately not ripe. The issues remarked by the reviewers were only partly addressed, and an improved version of the paper should be submitted at a future venue.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Reply to Review #1\", \"comment\": \"Thank you for your review and thoughtful feedback!\\n\\n--It is not very clear to me how the authors trained stochastic RNNs deterministically during training.\\n\\n We compared two different setups. In the first setup, we trained deterministic RNNs, and then added noise post-hoc at test-time, i.e. the model used to train and the model used to test are different. In the second setup, we trained stochastic RNNs with noise at training time, and added that same amount of noise at test-time, thus the train and test models are the same. These are two different ways to create a stochastic RNN at test-time. As one would expect, models trained with noise perform better when evaluated with noise, while models trained deterministically perform worse when noise is added. We included the first case primarily as a baseline, and as a means to estimate information in deterministically trained RNNs where it would otherwise be intractable.\\n\\n--Regarding: comparison to other methods\\n\\n We trained RNNs and LSTMs with output dropout (in Appendix A.5) and with the CPC objective instead of MLE (in Appendix A.4), while we were not able to compare to Z-Forcing in the short rebuttal period. We find that RNNs trained with dropout extract less information than RNNs without dropout, but the frontier of models when we sweep dropout rate and additive noise does not change. In other words, we can find models that are equivalent to adding dropout just by increasing the amount of Gaussian noise on the output. We have added an additional figure and discussion to the appendix including models trained with dropout. \\n\\n Similarly, we find that models trained with CPC perform similarly to models trained with MLE, and are not more optimal on the information plane. However, this may be due to the toy BHO dataset having Markovian dynamics, meaning that optimizing for one step ahead prediction with MLE is sufficient to maximize mutual information with the future of the sequence. Additional details and a figure can be found in the Appendix A.4 and Figure 10 in the paper.\\n\\n We hope that these additional experiments and improvements to the clarity of the text have addressed all of your concerns.\"}", "{\"title\": \"Reply to Review #2\", \"comment\": \"Thank you for your careful reading and feedback! We agree that stochastic training is a useful component of limiting information in RNNs, and performed additional experiments to address the connection with dropout as shown in Appendix A.5 and Figure 11.\\n\\n--The way stochastic training is introduced in this paper make it seem a bit contradictory with the fact that it actually helps generalization\\n\\n While stochastic training may limit performance on the training set (i.e. reducing I(hidden state; labels)), this often acts to regularize the model in a way that is beneficial for generalization to the test set, as we show in our experiments on the sketch dataset where training data is limited. Previous work on VIB (Alemi et al., 2017) has also shown how stochastic bottlenecks can improve generalization in the classification setting. \\n\\n--Intuitions that arise from the manuscript may not be as useful as we would like.\\n\\n As we emphasize throughout the manuscript, we find that RNNs trained without noise extract too much information about the past, and this can be harmful when training models on small datasets. Our simple and intuitive procedure of adding noise to hidden states to discard information presents one mechanism for limiting the capacity of hidden states, and our analysis in the information plane helps to reveal the tradeoffs between extracting information about the past and being able to predict the future of a sequence.\", \"reference\": \"Alexander A. Alemi, Ian Fischer, Joshua V. Dillon, Kevin Murphy. Deep Variational Information Bottleneck. ICLR 2017. https://arxiv.org/abs/1612.00410\"}", "{\"title\": \"Reply to Review #4\", \"comment\": \"Thank you for your careful reading and thoughtful feedback!\\n\\n--\\u201c the suggested method for improvement through adding noise to the latents (during training) is too much of handwaving for such a fundamental problem\\\"\\n\\n Adding noise to a bounded representation is a tractable and effective method for introducing a bottleneck into a hidden state with a bounded activation. Furthermore, for any level of additive noise, we can think of this process as introducing a hard constraint on the amount of information that can be stored in the hidden state, thus we\\u2019re solving a constrained optimization problem instead of the typical Lagrangian formulation in many applications of Information Bottleneck. While there are many other approaches for introducing bottlenecks (e.g. Dropout, computing a variational upper bound on the information and pushing that down), we found that adding noise was a simple strategy that was effective and interpretable.\\n\\n--Shouldn't the results on the BHC be quite surprising in terms of LSTMs performance? \\n\\n RNNs and LSTMs trained with deterministic hidden states perform well at next step prediction, but they do this by extracting far more information than is needed to solve the task. For large datasets this may not be an issue, but as we show in the QuickDraw experiments, when the data is limited, constraining the amount of information in the hidden state can be useful for improving generalization. \\n\\n--How are more future predictions generated.\\n\\n To evaluate mutual information between the hidden state and future states of the world, we do not need to generate future samples from our model, we only need hidden state samples paired with the true future observations. Furthermore, for a Markov process we only need to look a single step into the future to evaluate information with the infinite future.\\n\\n--Recently some researchers started to question whether 'reconstruction' is a good idea in order to learn generative-like models, for example you cite van den Oord 2018. How would such models perform in your metric.\\n\\n Thank you for this suggestion! We have performed additional experiments training RNNs and LSTMs with CPC. The model architecture is identical to our initial experiments, but instead of training with MLE we train with the CPC loss, which estimates the mutual information between the current timestep and K steps into the future. For our experiments on the BHO, we looked up to 30 steps into the future, and used a linear readout from the hidden state to a time-independent embedding of the inputs. We found that models trained with CPC had a similar frontier as those trained with MLE, and added a figure and additional details to Appendix A.4 and Figure 10 in the updated paper.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have read many papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper certainly poses an interesting question: How well do RNNs compress the past while retaining relevant information about the future. In order to quantitatively answer this question, the authors suggest to look at (the optimal solutions of) the Information Bottleneck Lagrangian (IBL). The investigated RNNs need to solve the task of next-step prediction, which can be used to evaluate the IBL. In the paper, the (deterministic) hidden state h is transformed through simple additive Gaussian noise into a stochastic representation which then is utilized to compute the IBL. In general the IBL is not tractable and hence the paper uses approximate computations.\\n\\nI definitely like the underlying question of the paper. Yet, to me it seems not ready for publication. For one, the presented experimental results look interesting but the suggested method for improvement through adding noise to the latents (during training) is too much of handwaving for such a fundamental problem. Second, shouldn't the results on the BHC be quite surprising in terms of LSTMs performance? Why is that, usually LSTM (or GRU) deliver excellent performance in typical (supervised) sequential tasks. Third, the task itself seems not well described, it seems to be next-step prediction, but how are more future predictions generated -- these seem to be not considered in the equation, but probably should when talking about 'retaining relevant information for the future'? Fourth, recently some researchers started to question whether 'reconstruction' is a good idea in order to learn generative-like models, for example you cite van den Oord 2018. How would such models perform in your metric.\\n\\nA final remark with respect to your citation for eq. 4, I think you meant Barber, Agakov, \\\"The IM algorithm...\\\", 2003?\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This manuscript shows that good ability to compress past information in RNNs help them to predict the future, and that improving upon this ability leads to more useful RNNs. The manuscript first adapts modern mutual-information estimators to mini-batch settings in order to measure the information that an RNN has on the past. It then considers stochastic training, adding Gaussian noise to the hidden states during the training of the RNNs to limit past information. A significant section is dedicated to an empirical study that shows that classically-train MLE RNNs lead to internal representations with a suboptimal mutual-information to the past and the future. For LSTM and GRU architecture, stochastic training actually significantly helps. Experiments on applications such as synthetizing hand-drawn sketches suggest that stochastic training leads to more useful RNNs.\\n\\nThis work has interesting observations and makes a credible case. The stochastic training does seem useful. However, I would like to understand better how it connects to the set of publications discussing dropout in RNNs. It is already known that stochastic perturbations during training help. In addition, the way stochastic training is introduced in this paper make it seem a bit contradictory with the fact that it actually helps generalization. I have the feeling that the benefit that is not understood and that intuitions that arise from the manuscript may not be as useful as we would like.\"}", "{\"rating\": \"3: Weak Reject\", \"experience_assessment\": \"I have published one or two papers in this area.\", \"review_assessment\": \"_thoroughness_in_paper_reading: I read the paper at least twice and used my best judgement in assessing the paper.\", \"title\": \"Official Blind Review #1\", \"review\": \"Summary:\\n\\nThe paper investigate how optimal recurrent neural networks (RNNs) are at storing past information such that it is useful for predicting the future. The authors estimated optimality in terms of mutual information between the past and the future. If the RNN was able to retain MI between the past and the future, it then has kept optimal information from the past for predicting the future. The experiments suggest that RNNs are not optimal in terms of prediction of the future. It also suggest that this is due to the maximum likelihood training objective.\", \"comments_for_the_paper\": \"1. Overall, the paper is a very interesting read and it explores and analyzes RNN under a different light. It answers a fundamental question about RNN training.\\n\\n2. There are a few things that would be nice to clarify. At the end of P3, the authors mentioned that the stochastic RNNs are trained either by a). deterministically during training and noise added during test or b) noise added during training and test. It is not very clear to me how the authors trained stochastic RNNs deterministically during training. It would be nice if this can be clarified. \\n\\n3. I am also curious how this compares to the training methods for example used in https://papers.nips.cc/paper/7248-z-forcing-training-stochastic-recurrent-networks.pdf. It seems that this would also help with retaining RNN optimality in terms of predicting the future. it would be interesting to include a comparison to this method for example.\\n\\nOverall an interesting paper. However, I think a few things could be improved and I would be willing to rise the score if the authors could addressed the above points.\"}" ] }

[ "Aviral Kumar", "Sergey Levine" ]

[ "data-driven optimization", "model-based optimization" ]

{ "note_id": [ "XCcOXrLGo_", "Syg3bE52jr", "SkgTIpFcsS", "HkeR_fSZoB", "Hkg9vv6koB", "SyeB72uRKS", "H1xX0KDRFB", "HyxTb_1TYr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730314, 1573852163870, 1573719380751, 1573110389905, 1573013345722, 1571879965193, 1571875274585, 1571776517409 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1703/Authors" ], [ "ICLR.cc/2020/Conference/Paper1703/Authors" ], [ "ICLR.cc/2020/Conference/Paper1703/Authors" ], [ "ICLR.cc/2020/Conference/Paper1703/Authors" ], [ "ICLR.cc/2020/Conference/Paper1703/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1703/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1703/AnonReviewer3" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This paper proposes Model Inversion Networks (MINs) to solve model optimization problems high-dimensional spaces. The paper received three reviews from experts working in this area. In a short review, R1 recommends Reject based on limited novelty compared to an ICDM 2019 paper. R2 recommends Weak Reject, identifying several strengths of the paper but also a number of concerns including unclear or missing technical explanations and need for some additional experiments (ablation studies). R3 recommends Weak Accept, giving the opinion that the idea the paper proposes is worthy of publication, but also identifying a number of weaknesses including a \\\"rushed\\\" experimental section that is missing details, need for additional quantitative experimental results, and some \\\"ad hoc\\\" parts of the formulation. The authors prepared responses that address many of these concerns, including a convincing argument that there is significant difference and novelty compared to the ICDM 2019. However, even if excluding R1's review, the reviews of R2 and R3 are borderline; the ACs read the paper and while they feel the work has significant merit, they agree with R2 and R3 that the paper needs additional work and another round of peer review to fully address R2 and R3's concerns.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Author Response: Summary of Revisions\", \"comment\": \"We thank the reviewers for their constructive feedback. We have revised the paper to improve clarity and included quantitative results in the main text. We summarize the updates below.\", \"section_2\": \"1. Added reference to Ngyuen et.al. in related work\\n\\nSection 3.2: \\n1. Added a probabilistic derivation of the inference scheme in Appendix A (newly added). \\n\\nSection 3.3:\\n1. Added interpretation of the function $g$ in reweighting \\n2. Improved Section 3.3 by omitting the rearrangement of the loss term\\n\\nSection 3.4:\\n1. Visualization of the randomized labeling scheme for a 1D function found here: https://ibb.co/album/i4d8qa (For reference, we used the same 1D function as https://tinyurl.com/min-1d-example )\\n\\nSection 3.5:\\n1. Described the procedure for creating the augmented dataset in practice \\n2. Improved description of how reweighting (Section 3.3) is implemented in practice\", \"section_4\": \"1. Added self-contained description of batch contextual bandits task \\n2. Added an ablation of MINs w/o reweighting in Table 1 \\n3. Added clear description to Figure 1 to describe how optimization is performed with MINs w/o inference (MIN - I); added a description of different rows in the caption of Figure 1. \\n4. Added quantitative score values to Figure 1\\n5. Added new tables (Table 2 and 3) with quantitative scores for face optimization task and MNIST inpainting\\n6. Added description of the architecture of $f^{-1}$ in Section 4.2 (protein fluorescence maximization) (previously in Appendix C.4)\\n7. Added self-contained description of the protein fluorescence maximization task, added a description of baselines\\n\\nIn all, we have quantitative results for five experiments, excluding benchmark functions, (Table 1, Figure 1, Table 2, Table 3, Table 5) and multiple ablations in Table 1 (w/o reweighting, w/o inference), Figure 1 (w/o reweighting, w/o inference), Figure 2 vs Figure 6, Appendix C.2 ( w/o reweighting), Table 5 (w/o reweighting, w/o inference), Table 4( w/o randomized labeling).\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"Thank you for the valuable feedback. We have included quantitative results for our experiments in Section 4 (Table 1, 2, 3, 5 and Figure 1). We have clarified details about tasks in the experimental section. We have also added a probabilistic interpretation for the method described in Section 3.2 in Appendix A. We answer specific questions below:\", \"q\": \"Quantitative results for optimization tasks\", \"a\": \"We have revised the paper to add quantitative results for all of the experiments, such that there are now five separate experiments with quantitative results. We hope that this addresses your concern about the quantitative evaluation, and we would welcome any further suggestions. These optimization scores were reported in the appendix in the submitted version. But now we have included these in the main paper (Table 2 and Table 3, Figure 1). A list of all quantitative results in the revised version of the paper is as follows --\\n\\n(1) Table 1 (Batch Contextual Bandits): We report the average test accuracy on unseen, novel contexts from the test split. The score is the average 0-1 reward across the test split, which is also the average test accuracy.\\n(2) Figure 1: We have added scores beneath each method. Higher scores are better, subject to the produced images being valid digit images. \\n(3) Table 2: We have added scores for the youngest face optimization task, where the score is given by the negative age of the face in the picture.\\n(4) Table 3: We have added scores for the contextual MNIST inpainting for two sub-cases.\\n(5) Table 5: Quantitative scores for protein design task.\\n\\nIn all, there are five experiments with quantitative results.\"}", "{\"title\": \"Response to Review #2\", \"comment\": \"Thank you for the constructive feedback. We have updated the paper (changes in red) to address the clarity concerns. Concisely, we have addressed clarity issues (in red) along these directions:\\n1. Added interpretation of the function $g$ in reweighting (Section 3.3)\\n2. Described the procedure for creating the augmented dataset in practice (Section 3.5)\\n3. Added clear description to Figure 1 to describe how optimization is performed with MINs w/o inference (MIN - I); added a description of different rows in the caption of Figure 1. \\n4. Improved Section 3.3 by omitting the rearrangement of the loss term \\n5. Added description of the architecture of $f^{-1}$ in Section 4.2 (protein fluorescence maximization). This has already been described in Appendix C.4.\\n\\nRegarding the concerns on ablations, we first summarize the existing ablations in our submission.\\n1. Table 1 (we compare MINs and MINs w/o inference)\\n2. Figure 1 (we compare MIN, MIN w/o inference, MIN w/o reweighting). Figure 8 (Appendix) shows quantitative scores for this task (measured as the number of active pixels (pixel value > 0))\\n3. Figure 2 and Figure 6, Appendix C.2 (Compares MIN and cGAN (no reweighting) on the youngest face optimization task; reweighting seems essential in this case)\\n4.Table 4 (final) (Ablation of randomized labeling; MIN + greedy = no randomized labeling)\\n5. Table 5 (final) (w/o reweighting) \\n\\nWe have updated the paper with an ablation study of MINs without reweighting on the contextual bandit task (Table 1). Also, please find gifs of runs (given different initial points) of the randomized labeling on a 1D function: https://ibb.co/album/i4d8qa (Earlier at: https://tinyurl.com/min-randomized1 https://tinyurl.com/min-randomized2 and https://tinyurl.com/min-randomized3) (For reference, we used the same 1D function as https://tinyurl.com/min-1d-example )\\n\\nPlease let us know if more ablations are necessary. Please let us know if any further clarity issues remain, we are happy to further update the paper. Further, all code used to run experiments will be released with the final version of the paper. \\n\\nPlease find specific replies to other questions below.\", \"q\": \"Are BO methods truly not applicable to the static setting?\\nTo the best of our knowledge, we haven\\u2019t seen a BO method applied in the static setting. Conditional Neural Processes (Garnelo et.al.) and Neural Processes (Garnelo et.al.) can perhaps be regarded as closest, but the main aim of these papers is to learn the forward function with correct uncertainty quantification. These approaches do not aim to perform optimization over the input $x$. If the reviewer has specific papers in mind, we would be happy to include those in related work.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"Thank you for this very relevant reference -- we were not aware of this prior work and will add and discuss it in our related work section. However, we believe that this work is broadly different both in terms of the problem it is addressing and the method that it is proposing.\", \"summary_of_our_response\": \"Nguyen et al. does not perform model-based optimization to produce an input $x$ that maximizes a scalar-valued $y$, but rather learns a generative inverse function of the form $f: y \\\\rightarrow x$, where $y$ is a vector-valued context, not the value of a function in an optimization problem. Further, there are many important differences in the technical approach which are elaborated upon below. Our paper already includes a comparison that is analogous to the most obvious application of this method for model-based optimization -- Figure 6 (Appendix C.2). MINs (Figure 2), which are designed for model-based optimization perform substantially better.\", \"detailed_response\": \"-----------------------------------------------------------------\", \"we_describe_the_differences_in_detail_below\": \"1. Differences in Problem Statement: Our aim is to solve optimization problems of the form $x^* = \\\\arg \\\\max_x f(x)$, where $f(x)$ (also referred to as y) is a scalar score value. Our goal is to find an x that *optimizes* the value of the score function. This is in contrast with the goal of Nguyen et al., which is to learn a generative model to generate values $x$ conditioned on a context $y$, where $y$ is a vector (for example, phase-transition graphs), and the goal is to produce any $x$ that is suitable for $y$. This problem statement is broadly distinct from model-based optimization.\\n \\n2. Comparison of the technical approaches: Nguyen et.al.'s approach is loosely analogous to conditional generative models (for example, cGANs [Mirza and Osindero, 2014]). Since MINs are designed to solve for the optimum $x$, there are many components in the MIN procedure that make it different from a cGAN (Algorithm 1):\\n(1) MINs learn an inverse map under reweighted data distribution (Section 3.3) \\n(2) MINs use specialized inference procedure to produce optimized $x$ values (Section 3.2).\\nWhile Nguyen et al.\\u2019s method is clearly relevant and will be discussed in our paper, its relationship with MINs is similar to that of other conditional generative models, such as cGANs, which are already discussed.\\n \\n3. cGAN baseline: In Appendix C.2, Figure 6, we present results from a cGAN baseline for the youngest face optimization task (compare to Figure 2). Note that a vanilla cGAN baseline tends to not perform optimization over the score value (i.e. doesn\\u2019t produce faces with small age), but rather produces images from the dataset distribution ignoring the score value. In general, static datasets could be highly skewed towards lower values of score that vanilla inverse-prediction might not be able to optimize score values.\\n \\nWe are happy to clarify any further questions that the reviewer has. We hope that our response clearly presents the differences between our approach and Ngyuen et al.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper prposes to learn an inverse network to predict x given a target y for optimisation, instead of the traditional way of optimisation (e.g. using Bayesian optimisation for the complex cases considered in the paper). However, unfortunately, this paper is too close in concept, and in my understanding lower in the solution quality to this recent paper:\\n\\nNguyen, Phuoc, Truyen Tran, Sunil Gupta, Santu Rana, Matthew Barnett, and Svetha Venkatesh. \\\"Incomplete conditional density estimation for fast materials discovery.\\\" In Proceedings of the 2019 SIAM International Conference on Data Mining, pp. 549-557. Society for Industrial and Applied Mathematics, 2019.\\n\\nPlease let me know if I missed anything. Otherwise it is a reject from me.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper tackles the problem of solving a black-box optimization problem where only some samples have been observed. This task requires a good model that can be both expressive and generalizable. Instead of learning only a single forward model of x -> y, this paper proposes to additionally use a mapping from y -> x. Optimizing in the space of z instead of x can be much simpler, and this should also act as a strong regularizer during training. Specifically, the paper uses a GAN that transforms [y,z] -> x, where z is stochastically sampled. This paper further proposes a reweighting scheme that interpolates between a uniform weighting and weighting the best sample so far, as well as a sampling procedure that iteratively samples points and refits a second model, which was inspired by Thompson sampling.\", \"pros\": [\"The proposed idea of using an inverse mapping is straightward but shown to be effective. The methods to make this work, namely reweighting and the randomized labeling procedure, seem to have some amount of theory behind them, though their presentation was confusing without multiple read-throughs.\", \"There are plenty of experiments across a wide array of domains, including images, 2D and 6D functions, and proteins which have a discrete representation.\", \"There are some comparisons to Spearmint and a scalable BO method (DNGO), though only on the 2D and 6D functions.\"], \"cons\": [\"The proposed pieces were often difficult to follow, and there doesn't seem to be sufficient information regarding the reweighting and randomized labeling for understanding and reproducing this work (see Questions).\", \"Only a few ablation experiments were carried out, and the effect of reweighting seems to only appear as a visual comparison in Figure 1. As the method seems quite different from related approaches, having more systematic comparisons would give us a better understanding of the core sources of improvement and better instigate future works.\"], \"questions\": [\"After Theorem 3.1, what is g? A linear function can only change a pdf by its bias (as any multiplication only affects the normalization constant), so not sure how to interpret g. I wasn't able to find this in the Appendix either.\", \"When creating the augmented dataset for defining & approximating some kind of posterior, can you be more concrete about how the y (and x) values are chosen? More specifically, is this procedure approximating some kind of meaningful posterior? If the y values are sampled randomly (from some fixed prior or perhaps from the p(y) described in the reweighting section?), what makes this procedure meaningful (does it rely on the learned GAN somehow)? Seeing a simple example in 1D would be a good visualization.\", \"What are the different rows of Figure 1 (different initializations)? Why are the different rows so similar for MIN but not for F? Also, what were the original examples (are they closer to the F results, or the MIN results)?\", \"For the protein task, how did you structure the output of f^-1(x) and how did you backpropagate through the discrete variables?\", \"In the experiments, it wasn't explicitly clear what the \\\"MIN without inference\\\" setting referred to; I assume this is where a single sample from the inverse mapping GAN was used?\"], \"additional_comments\": [\"When discussing the reweighted objective, the notation is a little confusing. The variables (j,k) is a different parameterization of the index i, but initially I thought k was indexing features of x. Perhaps a brief explanation of the rearrangement would make this clearer at first glance, or even remove this from the main text? (Without going through the proof, it isn't clear why this rearrangement is discussed.)\", \"After (1), it wasn't yet clear why this is called \\\"model-based\\\" optimization, as no model has been introduced yet.\", \"Are BO methods truly not applicable to the static setting? The static setting seems to be just a single prediction, which seems like a single step (ie. special case) of the non-static case?\", \"typo: \\\"method to solve perform optimization\\\"\", \"typo: Figure 1 \\\"Obsere\\\"\"]}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"In this paper, the authors present a method for model based blackbox optimization for cases where the design variables x are likely to lie on a manifold within a high dimensional space. The basic idea is to use a generative model to map targets y to potential design variables that would produce that target, and use this generative model in combination with a trained forward model to produce an optimal setting x* that both lies on the manifold and takes a high y* value, possibly higher than seen in the collected dataset (see the first two terms in equation 4).\\n\\nOverall, I feel positively about the paper, largely because of the idea it introduces. In particular, the task of manifold constrained optimization is not highly studied in BayesOpt, and it probably should be. Specifically, BayesOpt *is* occasionally studied on problems where the inputs are natural images, protein sequences, controller policies, and other settings where not all inputs are going to be valid, even if all inputs can technically give rise to a function value. The use of a GAN to constrain the possible inputs via a learned measure p(x|y) is a nice idea where applicable, and represents a concrete utility for GANs beyond generating images. All the individual components the authors' use to put this together are somewhat basic; however, the experimental results are vaguely convincing enough and I would choose to evaluate the merits of this paper on the core idea presented rather than whether or not the latest GAN architecture was used in the actual implementation.\\n\\nWith that said, I have a few comments. First, the constructions in 3.2 and 3.3 are perhaps a little more ad hoc than they need to be. If we are free to use probabilistic machinery for the forward model, then the fact that the inverse model effectively gives us a distribution fit to p(x|y) suggests that this can be naturally incorporated in to existing BayesOpt schemes. It might seem almost more satisfying to see a simple existing bayesopt pipeline extended with this idea, rather than wholly inference/recommendation and acquisition schemes introduced.\\n\\nAdditionally, two of the sections in the experimental results seemed somewhat rushed and need significant additional detail. The setup in the first portion of 4.1 is largely left to the reader to read Joachims et al., 2018 -- I'd like to see a more self contained description of the task. The more egregious example of this is the \\\"protein floresence [sic] maximization\\\" section. One of the baseline methods (GB) exists only as an acronym in Table 3 with no citation or discussion, and in general the section could be significantly expanded.\\n\\nThis last paragraph is something of a shame and one of the main weaknesses I see with the paper, as these two section are among the only quantitative results for the authors' optimization algorithm. Results are reported on Branin and Hartmann6, but frankly I don't see the value of these because they are far outside the intended application domain for the authors' work, as they are functions that are specifically intended to be optimized over compact domains. In general, it would be better to see more emphasis placed on quantitative experiments, particularly when the task is optimization. Perhaps adversarial image generation (where x arguably leaves the manifold of natural images, but not so far as to be random noise) or some other image task might substitute for the benchmark functions?\"}" ] }

[ "Defa Zhu", "Si Liu", "Wentao Jiang", "Guanbin Li", "Tianyi Wu", "Guodong Guo" ]

[ "Visual Relationship Detection", "Scene Graph Generation", "Knowledge", "Zero-shot Learning" ]

{ "note_id": [ "BuzOos2t2i", "ryeMPURijr", "BJg4MLRsjS", "HJgsbSRisH", "BJg-INCsjr", "H1lPSWAooH", "rkxZmVqCYS", "BJeeZWtaFr", "B1eKItmTFr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730283, 1573803610267, 1573803531633, 1573803266533, 1573803080613, 1573802302535, 1571886105410, 1571815672420, 1571793233052 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1702/Authors" ], [ "ICLR.cc/2020/Conference/Paper1702/Authors" ], [ "ICLR.cc/2020/Conference/Paper1702/Authors" ], [ "ICLR.cc/2020/Conference/Paper1702/Authors" ], [ "ICLR.cc/2020/Conference/Paper1702/Authors" ], [ "ICLR.cc/2020/Conference/Paper1702/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1702/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1702/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"The paper proposes a new problem setting of predicate zero-shot learning for visual relation recognition for the setting when some of the predicates are missing, and a model that is able to address it.\\n\\nAll reviewers agreed that the problem setting is interesting and important, but had reservations about the proposed model. In particular, the reviewers were concerned that it is too simple of a step from existing methods. One reviewer also pointed towards potential comparisons with other zero-shot methods.\\n\\nFollowing that discussion, I recommend rejection at this time but highly encourage the authors to take the feedback into account and resubmit to another venue.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Response to Reviewer #1\", \"comment\": \"We thank the reviewer for tending towards accepting our work and mentioning that our work has \\u201ca few strong contributions\\u201d. The \\u201cminor edits\\u201d are done in the revised version.\\n \\n*** Response to the things that could be strengthened or addressed further ***\\n \\nQ 1. There could be more meaningful comparison to other zero-shot learning algorithms. \\nA 1. Thanks for the great suggestion. Due to the time limit, we will add such results after the rebuttal period. \\n \\n \\nQ 2. Why was the unbalanced sampled-softmax was being used for only predicate prediction and not entity prediction? \\nA 2. (Paragraph 2, Sec 5, P6). The recognition of the entity is a full-supervised task. Since entity recognition is not the focus of this work, we have controlled the number of object categories by clustering with hypernym relationships. After that, the distribution of entities is relatively balanced. In this setting, the softmax has a good recognition effect.\\n \\n \\nQ 3. It wasn\\u2019t totally clear to the reviewer whether all of the verbs/entities were in WordNet and/or Glove.\\nA 3. \\na. All of the verbs/entities are in WordNet. \\nIn this problem setting, we let all the predicate categories to exist in WordNet.\", \"an_original_annotation_in_visual_genome_is_shown_in_the_follows\": \"++++++++++++++++++++++++++++++++++++++++++++++++++++++++++\\n\\\"predicate\\\": \\\"has\\\", \\n\\u2026\\n\\\"relationship_canon\\\": \\n[ { \\\"synset_name\\\": \\\"have.v.01\\\", \\\"synset_definition\\\": \\\"have or possess, either in a concrete or an abstract sense\\\" } ] \\n++++++++++++++++++++++++++++++++++++++++++++++++++++++++++\\nThe relationship instances are not only labeled with the \\u201cpredicate\\u201d (in natural language), but also with the synset (node of WordNet) corresponding to the \\u201cpredicate\\u201d. We filter out all relationship instances without labeling synset and used the synset to replace the original predicate, e.g., \\u201chave.v.01\\u201d replace \\u201chas\\u201d. Thus, all the predicate labels are in WordNet. \\n \\nb. Are all of the verbs/entities in Glove?\\nTake the above sample as an example, the definition of \\u201chave.v.01\\u201d is \\u201chave or possess, either in a concrete or an abstract sense\\u201d. We take all the words of the definition sentence as tokens to obtain the word embeddings and average all the word embeddings. We ignore the word not in GloVe. \\n \\nQ 4. Can human performance on this task be measured?\\nA 4. Thanks for the great suggestion. The experiment is ongoing.\", \"we_are_calling_on_undergraduates_to_perform_the_experiment_and_the_process_of_the_user_study_is_proposed_as\": \"1. A test image with two boxes is shown to the user. \\n2. The user types the keywords into the interface. \\n3. The system searches the predicates with the input keywords. \\n4. The user selects several output predicates as the answers. \\nIn the final version paper, we will report the human experimental results and details.\"}", "{\"title\": \"Response to Reviewer #2\", \"comment\": \"We thank the reviewer for the appreciation of this paper and have followed the valuable advice to revise the paper.\\n \\n*** Response to the main questions ***\\nQ 1. The special properties and challenges of predicate zero-shot learning. \\nA 1.\", \"properties_of_pzsl_are_summarized_as_follows\": \"a) PZSL is very necessary since the difficulty of labeling. To avoid exhaustively traversing all pairs ($O(N^2)$) of entities ($O(N)$), the labels of predicates (VG) are just extracted based on image descriptions. In this process, the predicates are seriously missing. It reflects the necessity of PZSL. \\n\\nb) PZSL is significant for downstream tasks. VRD is often used to generate a scene graph (SG, edge: predicate, node: entity) for visual reasoning (like VQA). As long as the SG is restricted to a closed vocabulary, the reasoning engine [1,2] is inevitably difficult to cope with the open questions with very diverse unseen predicates or entities. For visual reasoning, PZSL may make the state transition (the edge of SG) more generalized. (Paragraph 2, Sec 1)\", \"challenges_of_pzsl_are_summarized_as_follows\": \"c) Recognizing predicates is very difficult. 1). It is difficult to represent predicate. On the one hand, predicates are often abstract not as specific as objects. On the other hand, the representation of predicates also depends on that of subjects and objects. How to effectively represent the abstract predicates is still a hard problem. 2) Analogizing the seen abstract predicates to the unseen ones further brings in new difficulty. Furthermore, unlike many object ZSL methods, they adopt the pre-defined attributes of objects to recognize the unseen object. However, it is difficult to define the attributes of predicates. (Paragraph 3, Sec 1)\\n \\nd) The long-tail distribution of predicates is significantly more severe than that of objects. As suggested by the reviewer, the extreme long-tail distribution of predicates is presented in Fig. 4 of Appendix A. In the VG-Zero dataset, the instances' number of the top $5$ predicates count for about $80\\\\%$ of the total predicate instances. Under this distribution, the model tends to collapse to output few frequent predicates. As an initial solution to PZSL, we have to give priority to solving the extreme long-tail distribution (unbalanced sampled-softmax). (Paragraph 3, Sec 1)\\n\\n[1] Hudson D A, Manning C D. Learning by abstraction: The neural state machine.\\n[2] Mao J, Gan C, Kohli P, et al. The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision.\\n\\n \\nQ 2. The proposed method looks correct but is a rather direct application of existing methods. \\nA 2. \\na. We are the first to propose the unbalanced sampled-softmax to handle the extreme long-tail distribution. The most similar method Sampled Softmax, proposed in NLP for the large-scale vocabulary, fails under the extreme long-tail distribution. \\n \\nb. The graph sampling is borrowed from the recommender system domain after sufficient problem modeling and is very suitable to deal with large-scale KG for VRD problem. \\n\\n*** Response to the additional questions ***\\nQ 3. How well would a human do on this task? \\nA 3. Thanks for the great suggestion; the experiment is ongoing. \\nWe are calling on undergraduates to perform the experiment, where the process of the user study is proposed as: \\n1. A test image with two boxes is shown to the user. \\n2. The user types the keywords into the interface. \\n3. The system searches the predicates with the input keywords. \\n4. The user selects several output predicates as the answers. \\nIn the final version paper, we will report the human experimental results and details. \\n \\nQ 4. How much would Hits@k be if the test label is seen during training (not zero-shot)?\\nA 4. We test the model in Row 7, Table 2. The results of the generalized setting are as follows: \\n \\nTest set | Hit@5 Hit@10 Hit@20\\n====================================\\nUnseen predicates | 4.1 7.4 11.8\\nSeen predicates | 67.5 75.7 81.6\\nMixed predicates | 62.0 69.7 75.4\\n \\n\\u201cSeen predicates\\u201d means we only test on the region pairs that labeled with the seen predicates. \\n\\u201cMixed predicates\\u201d means we test on all the region pairs from the test images.\"}", "{\"title\": \"Response to Reviewer #3 (Main discussion, 1/2)\", \"comment\": \"We thank the reviewer for the valuable comments and constructive suggestions on our paper. The minor errors have been revised and the related references have also been added.\\n \\n*** Response to the main discussion ***\\nQ 1. Comparing \\u201cfast GCN\\u201d and \\u201cPinSage\\u201d. Why the authors have changed the name of PinSage and just mentioned that \\u201ctheir\\u201d \\u201cFast Graph Convolution Network\\u201d is \\u201cinspired\\u201d from PinSage? \\nA 1. Thanks for the valuable comments. \\u201cFast GCN\\u201d is a simplified version of \\u201cPinSage\\u201d. The name \\u201cFast GCN\\u201d tends to cause confusion so that we have changed the name to \\u201cSimplified PinSage\\u201d. \\n \\nThe components in PinSage that we don't use include: \\n1. Producer-consumer minibatch construction\\n2. Efficient MapReduce inference\\n3. Importance pooling\\n4. Curriculum training\", \"the_components_in_pinsage_that_we_use_include\": \"1. The architecture of GCN. \\n2. The node sampling strategy in minibatch, which we call \\u201con-demand sampling\\u201d. \\n3. The neighbor sampling strategy is a little different:\", \"pinsage\": \"adopt the \\u201cimportance pooling\\u201d to sample neighborhoods.\", \"ours\": \"We randomly sample neighborhoods.\\n \\n \\nQ 2. \\u201cgeneralized\\u201d and \\u201ctraditional\\u201d settings lack clarity.\\nA 2. The revised definition is shown as follows: \\nThe seen predicates are the predicates that appear in the training set, while the unseen predicates are those not in the training set. For the generalized setting, the methods search on the seen and unseen predicates vocabulary $\\\\mathcal{V}_{tr} \\\\cup \\\\mathcal{V}_{te}$ for inference. For the traditional setting, the methods only search on the unseen predicates vocabulary $\\\\mathcal{V}_{te}$. Note that the generalized setting is more challenging since the search space of the generalized setting is significantly larger than that of the traditional one. \\n \\n \\nQ 3. Please define the metrics clearly (Hits@k). \\nA 3. The definition of the metrics is revised as: \\nWe evaluate methods with Hit@k on generalized and traditional settings (Table 1 and 2): Given any test region pair, we check whether the GT label falls within the predicted categories with top k score. If yes, the sample is counted as a $Hit$. The Hit@k is $\\\\frac{\\\\#Hit}{\\\\#Test\\\\ region\\\\ pair}$.\\n \\n \\nQ 4. The descriptions for Table 1 and Table 2 fail to provide enough details to help understand the difference between the results in these two tables. \\nA 4. \\nIn Table 1 (Accuracy of unseen predicate recognition), the methods are only required to \\u201chit\\u201d the unseen predicates ignoring the subjects and objects. \\n \\nIn Table 2 (Accuracy of recognition of triplets with unseen predicates), the methods are required to \\u201chit\\u201d the full relationship triplets (subject, predicate, object). Only when all the elements of the triplet are \\u201chit\\u201d simultaneously can it be counted as a \\u201chit\\u201d.\"}", "{\"title\": \"Response to Reviewer #3 (Response to comments and feedback, 2/2)\", \"comment\": \"Q 5. Do you plan to provide the proposed dataset splits so others can work on this setting?\\nA 5. Yes, we will release the dataset as well as the splits setting. Source code will also be released to facilitate further researches. \\n \\n \\nQ 6 (Extra). \\u201cI have a question regarding the ablation studies with GloVe, Normal and InferSent initialization. The question is whether this initialization is necessary?\\u201d \\nA 6. The experimental results in Row 1 (only with GloVe)\\uff0c14 (only with WordNet) and 3 (with both) in Table 2 are analyzed as follows. \\n \\nThe generalized setting of PZSL is a very difficult test setting. Note that we only constrain the visual features matching the embeddings of the seen predicates during training. Thus, the visual features are more likely to match the embeddings of seen predicates. As the GT labels in the test set only containing the unseen predicates, all the seen predicates are wrong answers during testing, which makes the task difficult (both versions \\u201conly with GloVe\\u201d and \\u201conly with WordNet\\u201d get 0.0 accuracy). \\n \\nObserving on the traditional setting (only search on the unseen predicates for inference), the version \\u201conly with GloVe\\u201d (37.9 Hit@20) significantly outperforms the version \\u201conly with WordNet\\u201d (19.6 Hit@20). Furthermore, combining these two versions (\\u201cwith both\\u201d) will achieve significant improvement (48.7 Hit@20). \\n \\nThe reason why the version \\u201conly with WordNet\\u201d get a poor performance can be analyzed as follows: \\nThe GCN is just trained to align the visual features and semantic embeddings of the seen predicates. In this process, we have NO explicit constraints to build a good embedding space, such as putting together the semantically similar predicates. Due to the randomness of the initial embedding, the initial embeddings of two semantically similar predicates may be quite different. In this setting, the GCN is not explicitly supervised to pull closer the two predicates in the embedding space. Therefore, there is no guarantee that the randomly initialized embedding is able to achieve good results with WordNet. \\n \\nQ 7. \\u201cThe represented graphs show Object to Object and Predicate to Predicate connections which I\\u2019m not sure if it is correct?\\u201d \\nA 7. Object to-Object and Predicate-to-Predicate connections do exist in WordNet. \\nActually, the synsets (nodes in WordNet) are used to replace the original predicate categories. As shown in the following example, we use \\u201chave.v.01\\u201d to replace \\u201chas\\u201d as the predicate category. \\n \\nAn original annotation in Visual Genome is shown in the follows (Please focus on the \\u201csynset_name\\u201d): \\n++++++++++++++++++++++++++++++++++++++++++++++++++++++++++\\n\\\"predicate\\\": \\\"has\\\", \\n\\u2026\\n\\\"relationship_canon\\\": \\n[ { \\\"synset_name\\\": \\\"have.v.01\\\", \\\"synset_definition\\\": \\\"have or possess, either in a concrete or an abstract sense\\\" } ] \\n++++++++++++++++++++++++++++++++++++++++++++++++++++++++++\\n \\nWordNet contains the linguistic relationships of the synsets. For example, \\u201chave.v.01\\u201d is the hypernym of \\u201cstock.v.01\\u201d (the definition of \\u201cstock.v.01\\u201d is \\u201chave on hand\\u201d). Thus, there is an edge between \\u201chave.v.01\\u201d and \\u201cstock.v.01\\u201d. This explains why Predicate-to-Predicate connections exist. Similarly, Object-to-Object connections also exist. \\n \\nAs mentioned in Paragraph \\u201cKnowledge graph\\u201d, Sec 5, we use the following linguistic relationships (provided in WordNet) to build edges: \\nhypernym, \\nhyponym, \\npart meronym, \\npart holonym, \\nsubstance meronym, \\nsubstance holonym, \\nentailment, \\nsubstance holonym \\nand sharing lemmas. \\n \\nAs long as there is one linguistic relationship between two nodes, there will be one edge.\"}", "{\"title\": \"To all reviewers\", \"comment\": \"We thank reviewers for the very constructive comments and appreciation. We have revised the paper based on most of the comments. Due to the time limit, some experiments are ongoing and will be added in the final version. Please refer to the updated paper and answers to each reviewer.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I carefully checked the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Title: Good work, requires some edits.\\n\\n1. Summarize:\\n\\nThis paper proposes a new problem setting in visual relation detection which is called \\u201cPredicate Zero-shot Learning (PZSL)\\u201d. They provide a clear motivation and description of this setting. They propose a solution to this problem which leverages linguistic priors and knowledge bases. Furthermore they propose an unbalanced sampled-softmax to tackle the long tail distribution of predicates.\\n\\n2. Clearly state your decision. One or two key reasons for this choice.\\n\\nI will go for a weak accept for the paper at this stage. (+) I think the proposed problem setting is well-motivated and useful. Also, (+) the proposed initial solution to this problem is interesting. However, (-) they propose a \\u201cfast graph convolution network\\u201d which seems to be precisely equivalent to a PinSage. Also, (-) the paper requires to be polished as it lacks clarity.\\n\\n3. Main discussion\", \"my_first_argument_is\": \"I\\u2019m not sure why the authors have changed the name of PinSage and just mentioned that \\u201ctheir\\u201d \\u201cFast Graph Convolution Network\\u201d is \\u201cinspired\\u201d from PinSage. To me it looks exactly the same. If there are any differences, it should be stated clearly. In fact, I would not be against using PinSage as a part of their approach. However, trying to rename it without clear reasons is not a good idea.\\n\\nMy second argument is that the paper lacks clarity in writing (for detailed suggestions please refer to comments and feedbacks). Specially the evaluation section lacks details and clarity: a) In the beginning of this section (page 7), the authors talk about \\u201cgeneralized\\u201d and \\u201ctraditional\\u201d settings without properly defining them. b) The descriptions for Table 1 and Table 2 fail to provide enough details to help understand the difference between the results in these two tables (one of them states \\u201cAccuracy of unseen predicate recognition\\u201d and the other one \\u201cAccuracy of recognition of triplets with unseen predicates\\u201d). \\n\\n4. Comments and feedback.\", \"introduction\": \"\", \"paragraph_one_in_the\": \"1. The relationship recognition methods are mainly supervised \\u201cthat\\u201d \\u2192 \\u201cto\\u201d.\\n2. last line: \\u2026. and do not study \\u201con generalizing\\u201d \\u2192 \\u201cthe generalization of\\u201d.\", \"paragraph_two\": \"1. no manual annotations or \\u201creal samples\\u201d \\u2192 \\u201cimage samples\\u201d. (a real sample is ill-defined)\\n2. For example, no instance of chew \\u2192 For example \\u201cgiven\\u201d no instance of chew.\", \"paragraph_three\": \"1. \\u2026 is difficult since predicates are often abstract not as specific \\u2192 is difficult since predicates are often abstract \\u201cand\\u201d not as specific.\\n2. Furthermore, unlike many object ZSL methods \\u2026 \\u2192 This line to the end is very complicated and hard to understand.\", \"related_works\": \"1. Visual Relationships: I would cite \\u201cGraph R-CNN for scene graph generation\\u201d since it is the most relevant work regarding the similarity of pipeline (using GCNs).\\n2. External Knowledge bases (KB): I would cite \\u201cImproving Visual Relationship Detection using\\nSemantic Modeling of Scene Descriptions\\u201d since it is one of the most relevant works using knowledge graph modelings to improve visual relation detection.\", \"problem_setup\": \"Do you plan to provide the proposed dataset splits so others can work on this setting? I consider this very important given your paper\\u2019s contribution. Maybe it is better if it is also mentioned in the paper.\", \"pipeline\": \"1. Paragraph 2: \\u2026 the output of which is fused with \\u2026: Given Figure 2, it does not seem like $V_p$ is being created by fusing $V_s$ and $V_o$. It looks more like it is extracted directly from the image (union of bounding boxes).\\n\\n2. In Figure 2: In the representation of Pipeline (A), the graph is colored by dark blue for objects and light blue for predicates. The represented graphs show Object to Object and Predicate to Predicate connections which I\\u2019m not sure if it is correct. Shouldn\\u2019t we always have a light blue between every pair of dark blue connections?\", \"evaluation\": \"1. Please consider the mentioned points in the Main Discussions.\\n2. In Table 1, I suggest re-naming \\u201cembedding\\u201d to \\u201cinitial embedding\\u201d.\\n3. In Table 1, Hit@k should be Hits@k.\\n4. Please define the metrics clearly (Hits@k).\", \"extra\": \"I have a question regarding the ablation studies with GloVe, Normal and InferSent initialization. The question is whether this initialization is necessary? It seems like in the setting \\u201cW/O KG\\u201d, even though the embeddings are initialized with GloVe, there is no gain (all of the Hits@k values are 0.0). So GloVe embedding without KG bring no external semantic knowledge? Then why use them? Regarding that, I can see that a GCN, initialized with normally distributed embeddings (row 14 in Table 1) has given 0.0 accuracies, but I find this very counter intuitive, as graph convolution layers already have trainable weights capable of compensating for the lack of \\u2018proper\\u2019 initialized embedding and getting 0.0 does not make sense to me.\", \"conclusions_and_future_work\": \"1. two lines before the last: please use \\u201c\\\\citep\\u201d.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"The paper considers the task of predicting visual predicates (e.g., eat, bite, take) between pairs of entities. In particular, the paper focuses on the zero-shot setting where the test predicates are unseen during training. The model uses linguistic prior from a knowledge graph (WordNet):\\nwith graph embedding (fast GCN), unseen predicates are embedded based on the information propagated from seen predicates. The model is trained so that the visual feature vector and the correct predicate embedding are nearby in the joint embedding space. The method was evaluated on a zero-shot split of the Visual Genome dataset.\\n\\nOverall, as a task and dataset paper, the paper should have sold the task more by highlighting its special properties. While the task of predicting unseen predicates is interesting, the setting and the technique are similar to previous work on predicting other types of unseen labels (e.g., unseen objects, as referenced in the paper). The new task could still be interesting if it presents different challenges (e.g., maybe predicates are more ambiguous than objects, or visual predicates are harder to embed). But from the description in the paper, most of the challenges seem to also exist in other zero-shot settings (long-tail distribution; large space of labels). The paper could benefit from providing examples or statistics that demonstrate the challenges of the task.\\n\\nThe proposed method looks correct but is a rather direct application of existing methods. The experiment setup looks OK. Based on the error analysis, the labels look very noisy and subjective, but this seems to be a common problem in the visual predicate prediction task (hence the recall-based evaluation).\", \"additional_questions\": [\"The provided examples in the error analysis look pretty tricky; e.g., \\\"swing\\\" and \\\"slug\\\" are judged as different. How well would a human do on this task?\", \"How much would Hit@k be if the test label is seen during training (not zero-shot)?\"]}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"This paper creates a new task for zero-shot learning of predicates (specifically in cases where the individual predicate components have never been seen in the training, rather than the more traditional setting where the full s-v-o relationship is unseen but each component is). They create a new subset of the Visual Genome dataset specifically targeted towards predicting unseen predicates. They also provide results using a knowledge graph (in this case WordNet), to integrate linguistic and visual features for prediction. Interestingly, their pipeline introduces a new softmax variant, the unbalanced sampled softmax, which addresses the problem of over-predicting common predicates.\\n\\nI generally tend towards accepting this paper. The reasons being that this paper has a few strong contributions: (1) they design a new task set-up with data they selected and cleaned from VG, (2) new modelling pipeline with empirical analysis backing the modeling choices, and (3) new softmax variant.\", \"a_few_comments_about_things_that_could_be_strengthened_or_addressed_further\": [\"There could be more meaningful comparison to other zero-shot learning algorithms. Even if they are not fully comparable because they were originally meant for a slightly different zsl set-up, it would be nice to have more baselines from external work.\", \"Why was the unbalanced sampled softmax was being used for only predicate prediction and not entity prediction?\", \"It wasn\\u2019t totally clear to me whether all of the verbs/entities were in WordNet and/or Glove. If not, can the authors clarify what the overlap was and how this might be affecting performance?\", \"As noted by the authors, there are some cases in Figure 3 where humans would consider the answer to be confusing or might actually prefer the machine response. Can human performance on this task could be measured? Perhaps humans could possibly evaluate a subset of the machine vs. gold answers?\"], \"minor_edits\": [\"Related work: WordNet is misspelled in the last line\", \"Figure 3: please consider using colors other than red/green, this is not readable for color-blind readers\", \"Section 5.2: \\u201cThe case c is confusing that even\\u201d \\u2192 \\u201cThe case c is so confusing that even\\u201d\", \"Section 5.2: \\u201cand output a more appropriate\\u201d --> \\u201cand outputs a more appropriate\\u201d\", \"In the references: the Devise paper by [Frome et al 2013] is listed twice\", \"In the references: the first author of ConceptNet5 should be Robyn Speer\"]}" ] }

[ "Simon Du", "Mengdi Wang", "Ruosong Wang", "Lin F. Yang" ]

[ "continuous control", "learning", "context" ]

{ "note_id": [ "ZR3jbPaDAF", "BkeJLdEMnS", "HkeH_BO3sS", "B1luvxAijS", "rJg0mNiYsS", "r1edA96SjH", "H1xficTBsr", "HJelb3w15r", "B1lQPS9pYH" ], "note_type": [ "decision", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review" ], "note_created": [ 1576798730253, 1574221894879, 1573844332647, 1573802079734, 1573659686493, 1573407440254, 1573407385558, 1571941368504, 1571820891465 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1701/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1701/Authors" ], [ "ICLR.cc/2020/Conference/Paper1701/Authors" ], [ "ICLR.cc/2020/Conference/Paper1701/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1701/Authors" ], [ "ICLR.cc/2020/Conference/Paper1701/Authors" ], [ "ICLR.cc/2020/Conference/Paper1701/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1701/AnonReviewer2" ] ], "structured_content_str": [ "{\"decision\": \"Reject\", \"comment\": \"This work considers the popular LQR objective but with [A,B] unknown and dynamically changing. At each time a context [C,D] is observed and it is assumed there exist a linear map Theta from [C,D] to [A,B]. The particular problem statement is novel, but is heavily influenced by other MDP settings and the also follows very closely to previous works. The algorithm seems computationally intractable (a problem shared by previous work this work builds on) and so in experiments a gross approximation is used.\\n\\nReviewers found the work very stylized and did not adequately review related work. For example, little attention is paid to switching linear systems and the recent LQR advances are relegated to a list of references with no discussion. The reviewers also questioned how the theory relates to the traditional setting of LQR regret, say, if [C,D] were identity at all times so that Theta = [A,B]. \\n\\nThis paper received 3 reviews (a third was added late to the process) and my own opinion influenced the decision. While the problem statement is interesting, the work fails to put the paper in context with the existing work, and there are some questions of algorithm methods.\", \"title\": \"Paper Decision\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I did not assess the derivations or theory.\", \"title\": \"Official Blind Review #4\", \"review\": \"This paper considers the problem of changing environments for LQR. The authors model this through the use of a decoder that maps an incoming context (C,D) to the LQR matrices (A,B). They provide an algorithm for this setting based on a UCB strategy, prove sample complexity and regret bounds, and experimental results.\\n\\nOverall the paper was well written but I had several concerns.\\n\\n1. The results of this paper were not contrasted with other papers in this area. For example, if C,D are constant, and \\\\Theta_* is a Block diagonal matrix with A,B on the diagonal - then the contextual case reduces to the standard LQR problem. It's unclear how the results given compare to past results in this setting, for example those of Abbasi-Yadkori/Szepesvari 2011.\\n\\n2. I did not fully understand the UCB nature of the algorithm. In each round \\\\Theta^(k) (the least squares estimator) seems to be used to compute the optimal policy (line 10 of the algorithm) instead \\\\tilde{\\\\Theta}^(k) -the optimistic estimate. The optimistic estimate is only used in line 16 - a randomized procedure that is unmotivated.\\n\\n3. Building on (1), it is hard to understand the results as given since there are no lower bounds given nor is there a discussion of the problem dependent parameters that arise. For example, in Theorem 1, is dp^2 suspected to be tight? Since the number of parameters in \\\\Theta^{\\\\ast} is d(p+p'), perhaps this is off by a factor of p?\\n\\n3. I struggled to understand the setup of the experiments - as described the algorithm given was not used at all, rather \\\\Theta^(k) was approximated and beta^k was set to be a constant. This does not seem like a fair evaluation of the method. \\n\\nIn summary, I would reject this submission unless the authors couch it better in past work, explain their results better, and improve the experiment setup.\\n\\nFinally, a typo: I think the indexing variable in the equation on the top of page 4 is h' not h.\"}", "{\"title\": \"General Response and Revision Summary\", \"comment\": \"We thank both reviewers for the constructive comments! All major changes are marked in red. We made the following main changes in our revision.\\nWe changed the title according to suggestion from Review #1.\\nWe updated Figure 2 to make the presentation clearer according to the suggestion from Review #1.\\nWe added a clarification that Q and R are known to the agent according to the suggestion from Review #2.\\nWe added a paragraph to discuss a na\\u00efve approach and motivate our algorithm according to the suggestion from Review #2.\\nWe added a clarification that V_k and C_k are changing at every epoch according to the suggestion from Review #2.\"}", "{\"title\": \"Response\", \"comment\": \"For the first question, notice that there are two main components in our algorithm: UCB-type exploration (Line 6 in Algorithm 1) and linear regression to estimate $\\\\Theta$ (Line 15 in Algorithm 1). Here we would like to explain the necessity of both components. Notice that the problem studied in this paper includes linear bandit [1] as a special case, and most previous algorithms for solving linear bandit do have UCB-type exploration. Moreover, linear regression is also a special case of the problem studied here (if one wants to estimate $\\\\Theta$). Our algorithm is a careful combination of these two main components, with additional steps to deal with the underlying LQR problem. Thus, our algorithm is in fact very natural from this point of view, and any algorithm with provable guarantees do need the two components mentioned above (since it needs to be able to solve the two special cases).\\n\\nFor the second question, our theoretical bound for $\\\\beta^{(k)}$ does depend on $k$, but only depends on $k$ logarithmically. For the setting of the experiments considered in this paper ($k$ is at most 100), it is reasonable to ignore such a minor dependence, and thus we use a fixed value of $\\\\beta^{(k)}$ in our experiments. Moreover, in our experiments we do use different values of $V^{(k)}$ for the confidence set $\\\\mathcal{C}^{(k)}$ as described in Algorithm 1. Thus, our confidence set does change as $k$ changes. We will make this point more explicit in the description of the algorithm and in the experiment description. \\n\\nFor the third question, thanks for the suggestion. We will make this clear in the paper. \\n\\n[1] Stochastic linear optimization under bandit feedback. Varsha Dani, Thomas P. Hayes, and Sham M. Kakade. COLT 2008\"}", "{\"title\": \"Comments on authors' response\", \"comment\": \"Thanks for the clarifications. I still have a few concerns.\\n\\n| The algorithm you proposed will not work. The reason is simple, since your double least square algorithm requires [A,B] estimated to be accurate.\\nThat depends on how long each trajectory is. Since the system is linear, efficient system identification is feasible [1]. Furthermore, please don't take my suggested algorithm literally. What I mean is that it would be helpful to compare your proposed scheme against some baselines. And there are several naive baselines one could think about, so showing that your method outperforms them could only be advantageous.\\n\\n| the specific choice of hyper-parameters (beta) is tuned to achieve the best performance\\nWhat I was concerned about is that the dependence on k was dropped. In UCB, the bonus grows with time, e.g., as sqrt(log t). However, looking closer at Eq. (10), it seems that even for a fixed beta, the confidence set C_k will still change in size because V_k and theta_k are changing, and it is the confidence set and not beta which is crucial for balancing exploration/exploitation. Perhaps making it a bit more explicit in the text would help the readers.\\n\\n| we assume [Q,R] are known. This is standard in LQR literature.\\nIt is also standard in LQR literature that matrices A and B are known... It would still be helpful if you clarify whether [Q, R] are context-dependent.\\n\\n[1] R. Mehra. Optimal input signals for parameter estimation in dynamic systems\\u2013survey and new results.\\nIEEE Transactions on Automatic Control, 19(6):753\\u2013768, 1974.\"}", "{\"title\": \"Response\", \"comment\": \"Thanks for raising these questions. Please find our responses to your questions below.\\n1.\\tThe algorithm you proposed will not work. The reason is simple, since your double least square algorithm requires [A,B] estimated to be accurate. Furthermore, it is clear from bandit and RL literature that without UCB or other exploration techniques, one cannot achieve $\\\\sqrt{T}$ type of regret bound. \\n\\n2.\\t- For the random sampling step in experiments, it is straightforward to show that in the low-dimensional setting using a small number of samples, one can obtain a near optimal solution. We leave devising a provably efficient computational efficient approach for the high-dimensional setting as a future work.\\n- The crucial point to use UCB is that UCB can balance exploration and exploitation. In practice, the specific choice of hyper-parameters ($\\\\beta$) is tuned to achieve the best performance. Note this is standard in bandit and RL literature. In our experiments we set $\\\\beta=10^4$ to encourage exploration, as mentioned on Page 7. \\n\\n3.\\tThroughout the paper, we assume [Q,R] are known. This is standard in LQR literature. We will clarify this assumption in the next version.\"}", "{\"title\": \"Response\", \"comment\": \"Thanks for the positive review. We will improve and polish our experiment section in the final version. We will also update the title. Thanks for the suggestion.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"In order to generalize the RL agent to unseen environment, in this work the authors studied the theoretical learning problem of building a decoder on top of linear continuous control using linear quadratic regulator (LQR). They presented a simple, UCB-based algorithm that refines the estimates of the encoder while doing LQR and balances the exploration-exploitation trade-off. In the online setting, the proposed algorithm has a O(\\\\sqrt{T}) regret bound, where T is the number of environments the agent played. This also implies after certain exploration, the agent is able to transfer the learned knowledge to obtain a near-optimal policy to an unseen environment. To justify their theoretical bounds the authors also present experiments that demonstrate the effectiveness of the algorithm.\\n\\n\\nThe work of designing decoder on top of RL/control in order to generalize to new, unseen environments is very interesting, and is pretty novel to my knowledge. The problem formulation of LQR is standard until the part where the authors introduced the output matrices (C,D), which extends the fully-observable case of LQR (that is based on state feedback only) to partially observable. Leveraging the theoretical analysis of LQR, the authors extended the analysis to the setting of output feedback with particular structures of decoder matrices (C,D) sampled from decoder \\\\mu. The algorithm proposed is quite standard in the output-feedback LQR literature (in control or RL). But the work is still interesting because to my knowledge I am not aware of general theoretical analysis of this setting (while most analysis is based on the full state feedback).\\nI haven't checked the proofs very carefully in the appendix, but from the description in the main paper it seems the analysis of contextual transfer learning performance is sound, and under certain regularity assumptions the authors did provide a high-probability regret bound for this contextual transfer learning problem. It would be great if the experiments are more involved as they are a bit too simple at this point, (where the unseen environment is the change in the physical constants). I also have some difficulties understanding all the dots in figure 2. Perhaps the authors can simplify the number of trajectories plotted there to make the presentation clearer. Another comment is about the current title, currently by looking at it I have no idea that is about contextual transfer learning and LQR. It would be great if that can be more specific.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"# Summary\\n- The paper proposes a UCB-inspired algorithm for a contextual LQR problem. The problem itself is introduced in this paper and is similar in spirit to CMDPs, with the difference that instead of learning a mapping from context to transition matrix, a mapping from context to matrices [A, B] figuring in the system dynamics of LQR is learned.\\n- The proposed algorithm is an online-learning algorithm shown to have sublinear regret in the number of experienced environments. A toy experiment with a 2D moving mass is presented to illustrate the theory.\\n\\n# Decision\\nAlthough the problem setting is interesting and it is encouraging to have a guarantee, several important unclear points in the paper and a missing comparison to a straightforward baseline stop me from recommending it for publication in its present form. I detail my concerns below.\\n\\n# Concerns\\n1) First, a conceptual question. I can see a straightforward algorithm that can learn the linear mapping \\\\theta from context to [A, B] as follows.\\n - In episode k, obtain trajectory (x_{1:H}, u_{1:H-1})\\n - By least squares, find [A, B] from the obtained trajectory\\n - Since context [C, D] is observed, find \\\\theta : [C, D] -> [A, B] again by least squares\\nOne can do this using data from K episodes if needed, one can sequentially update the controller for collecting data, etc.\\n=> A comparison to such a basic approach should be definitely included in the paper, in my opinion.\\n\\n2) The authors might argue that the algorithm suggested above has no guarantee. I would be curious to hear in this regard a comment on the practical implementation suggested in the paper. Namely, after deriving the bounds etc., the authors make further approximations and modifications in the practical algorithm. From my point of view, these modifications defeat the purpose of the bounds, because then only empirical evaluation can confirm that these approximations have not destroyed the analysis. Alternatively, one needs to incorporate the introduced approximation errors in the analysis. In more detail,\\n - Eq. (9) is not solved exactly but by random sampling. In the 2D toy task, it may be OK, but in higher-dimensional spaces, a significant error can be introduced which is not accounted for.\\n - More importantly, the UCB bound \\\\beta in Eq. (11) is not used at all in the experiments.\\n => To my understanding, it is the crucial point of UCB to use the UCB-bound. If it is not used, how should one judge the resulting algorithm?\\n\\n3) This is a concern regarding clarity. I didn't get (i) if matrices [Q, R] are context-dependent or not and (ii) if the agent observes them or not. This is not clearly communicated in the text.\\n=> Clarify whether [Q, R] are context-dependent and observed.\\n\\n# AFTER REBUTTAL\\nAfter authors' clarifications and improvements on the paper, I update my score to weak reject. The reason I am still against acceptance is the lack of stronger empirical evaluations. As R4 pointed out, some clarifications on the side of the algorithm are also required.\"}" ] }

[ "Emilio Parisotto", "Francis Song", "Jack Rae", "Razvan Pascanu", "Caglar Gulcehre", "Siddhant Jayakumar", "Max Jaderberg", "Raphaël Lopez Kaufman", "Aidan Clark", "Seb Noury", "Matt Botvinick", "Nicolas Heess", "Raia Hadsell" ]

[ "Deep Reinforcement Learning", "Transformer", "Reinforcement Learning", "Self-Attention", "Memory", "Memory for Reinforcement Learning" ]

{ "note_id": [ "n1m6lmJM207", "DxVrvy_Za", "S1xS4AVsjB", "rkeZiLPDiS", "r1gutMDPjS", "HJxtWzwviH", "Hkxirbwvjr", "SJxOvgPPsH", "S1x2q1vDjS", "HkeMR-1QiB", "rklAUbhRFr", "BJerZG59YS", "H1l3qFXvKr" ], "note_type": [ "comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "comment" ], "note_created": [ 1650352825306, 1576798730223, 1573764652645, 1573512857129, 1573511807788, 1573511680571, 1573511491296, 1573511264137, 1573511060140, 1573216714127, 1571893590076, 1571623420856, 1571400084477 ], "note_signatures": [ [ "~Marco_Pleines1" ], [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1700/Authors" ], [ "ICLR.cc/2020/Conference/Paper1700/Authors" ], [ "ICLR.cc/2020/Conference/Paper1700/Authors" ], [ "ICLR.cc/2020/Conference/Paper1700/Authors" ], [ "ICLR.cc/2020/Conference/Paper1700/Authors" ], [ "ICLR.cc/2020/Conference/Paper1700/Authors" ], [ "ICLR.cc/2020/Conference/Paper1700/Authors" ], [ "ICLR.cc/2020/Conference/Paper1700/AnonReviewer4" ], [ "ICLR.cc/2020/Conference/Paper1700/AnonReviewer1" ], [ "ICLR.cc/2020/Conference/Paper1700/AnonReviewer3" ], [ "~julien_perez1" ] ], "structured_content_str": [ "{\"title\": \"Lack of reproducibility due to missing source code and technical details\", \"comment\": \"Adding a Transformer architecture to Deep Reinforcement Learning algorithms is a non-trivial process. It is great that a few people, like the authors and Lampinen et al. 2021, got it working while showing promising results towards multiple POMDPs. However, there is one major concern. There is no clear implementation available. Without access to the source code, many important implementations details are missing that ensure reproducibility. It is unclear what kind of data is stored in the episodic memory (raw observations? encoded features?) and how the interface between the memory and the training algorithm works. There are likely more questions that the authors should address to make their research accessible and hence reproducible, which should be the default for scientific work. This issue is also apparent in the published ICML paper.\"}", "{\"decision\": \"Reject\", \"comment\": \"This paper proposes architectural modifications to transformers, which are promising for sequential tasks requiring memory but can be unstable to optimize, and applies the resulting method to the RL setting, evaluated in the DMLab-30 benchmark.\\n\\nWhile I thought the approach was interesting and the results promising, the reviewers unanimously felt that the experimental evaluation could be more thorough, and were concerned with the motivation behind of some of the proposed changes.\", \"title\": \"Paper Decision\"}", "{\"title\": \"Paper Revision #2\", \"comment\": \"Dear Reviewers,\\n\\nWe have submitted another revision where in Section 4.3.3 we include the results of a substantially larger LSTM baseline that is parameter-matched to the transformers we trained. This model performs only marginally better than the original baseline LSTM and the performance gains are mainly in the reactive set of levels.\"}", "{\"title\": \"Paper Revision\", \"comment\": \"Dear Reviewers,\\n\\nWe have uploaded a paper revision that we hope addresses some of the feedback made in reviews. We have expanded the section describing DMLab-30 with references to previous work that used the 30 levels or subsets of them as benchmark environments for algorithms and architectures. Additionally, as requested by some reviewers we include in the appendix additional results on the multitask Atari-57 benchmark. We see no improvements over an LSTM baseline, which we believe is expected as Atari-57 is not meant to test an agent's memory capabilities.\"}", "{\"title\": \"Response to Review #1 cont\", \"comment\": \"> Lack of understanding of the layer normalization. The authors provide some explanations about why the reordering works, but they seem not intuitive. More analysis about why the reordering works would significantly enhance this paper.\\n\\nWe first want to reiterate that reordering the layer norm is not an original contribution of our work, as we cited in our paper it has been established in transformers previously [7][8] and earlier for convolutional networks [9]. A definitive reason for keeping the residual stream untransformed through depth still remains an open question. Our work empirically demonstrates that it provides substantial improvements in stability for RL domains which we believe, by itself, is an important contribution for others working in the field. See our answer to point (3-4) for a summary of a previously established hypothesis for why reordering normalization works.\\n\\n>Have you tried simply removing the layer normalization layer?\\nAn initial experiment removing layer norm performed substantially worse, even on a shallow 4-layer transformer. \\n\\n> TrXL-I moves two layer-normalization layers together. Have you tried only moving one of them? Which modification contributes more?\\nWhile this is definitely an interesting question, it is a convention to place the layer norm at the same place and due to the already substantial amount of computation time the ablations required, we were not able to test this. \\n\\n> Specific Points (3-4) \\nWe believe our reasoning for reordering the layer norm operator largely follows the same lines as previous work examining normalization placement (see section 2 of [9]). Having an identity mapping (the skip connection in the residual connection) enables gradients to flow from any layer l to any shallower layer k<l untransformed, allowing an easier optimization as e.g. gradients can flow from the policy and value head outputs directly to the image encoder without any intermediate transformation. If the layer norm was placed after residual recombination, then during backpropagation the gradients at a shallower layer k<l are non-linearly transformed by the backwards pass of each layer m\\u2019s layer norm, where k<m<l. \\n\\n[1] Recurrent Experience Replay in Distributed Reinforcement Learning. Steven Kapturowski, Georg Ostrovski, Will Dabney, John Quan, Remi Munos. International Conference on Learning Representations. 2019.\\n[2] Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures. Lasse Espeholt, Hubert Soyer, Remi Munos, Karen Simonyan, Volodymir Mnih, Tom Ward, Yotam Doron, Vlad Firoiu, Tim Harley, Iain Dunning, Shane Legg, Koray Kavukcuoglu. International Conference on Machine Learning. 2018.\\n[3] Multi-task Deep Reinforcement Learning with PopArt. Matteo Hessel, Hubert Soyer, Lasse Espeholt, Wojciech Czarnecki, Simon Schmitt, Hado van Hasselt. Association for the Advancement of Artificial Intelligence. 2018.\\n[4] Information Asymmetry in KL-regularized RL. Alexandre Galashov, Siddhant M. Jayakumar, Leonard Hasenclever, Dhruva Tirumala, Jonathan Schwarz, Guillaume Desjardins, Wojciech M. Czarnecki, Yee Whye Teh, Razvan Pascanu, Nicolas Heess. International Conference on Learning Representations. 2019.\\n[5] Mix & Match \\u2013 Agent Curricula for Reinforcement Learning. Wojciech Marian Czarnecki, Siddhant M. Jayakumar, Max Jaderberg, Leonard Hasenclever, Yee Whye Teh, Simon Osindero, Nicolas Heess, Razvan Pascanu. International Conference on Machine Learning. 2018.\\n[6] Exploration by Random Network Distillation. Yuri Burda, Harrison Edwards, Amos Storkey, Oleg Klimov. International Conference on Learning Representations. 2019.\\n[7] Adaptive Input Representations for Neural Language Modeling. Alexei Baevski, Michael Auli. International Conference on Learning Representations. 2019.\\n[8] Language Models are Unsupervised Multitask Learners. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever. 2019.\\n[9] Identity Mappings in Deep Residual Networks. Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. European conference on computer vision. 2016.\"}", "{\"title\": \"Response to Review #1\", \"comment\": \"Dear Reviewer 1,\\n\\nWe thank you for your comments and constructive criticism. We hope to address your concerns below:\\n\\n> Lack of experiments on benchmark and large environments... it is unclear whether the proposed transformer can scale to large environments\\n\\nWe want to highlight that DMLab-30 is a large-scale, multitask benchmark arguably more challenging than Atari-57, comprising 30 first-person, 3D environments with image observations. The tasks require a range of agent competencies such as language comprehension, navigation, handling of partial observability, memory, planning, and other forms of long horizon reasoning, with episodes lasting over 4000 steps. It has been used in a large number of previous works, see e.g. [1][2][3][4][5] for a subset, with the previous multitask state-of-the-art being [3]. With the substantial amount of previous work that went into obtaining high scores on the DMLab-30 suite as context, the GTrXL results we report in this paper are the highest so far obtained for this benchmark.\\n\\n> The authors did not evaluate their model on the widely used benchmark Atari-57.\\n\\nThe transformer is mainly a memory architecture and previous work has shown minimal improvement from using memory on Atari-57 (see e.g.in RND [6], on Montezuma\\u2019s Revenge a feedforward CNN achieves scores close to a recurrent model). While the Atari-57 tasks themselves might not always be fully observable, they are not meant to test the memory of the agent. On the other hand, DMLab-30 was explicitly designed to require memory in order to solve it.\\n\\nDespite this, we have run the GTrXL on the Atari-57 suite. As in previous work, memory does not seem to play a large part in performance and there is no substantial improvement of the transformer over a baseline LSTM architecture. However, we can at least note that the GTrXL suffers no performance degradation on this set of environments, and that it shows stability comparable to an LSTM.\", \"atari_57\": \"Model | Median Human Normalized Score \\nLSTM | 136.6 +- 3.4 \\nGTrXL | 137.1 +- 5.0\"}", "{\"title\": \"Response to Review #3 cont\", \"comment\": \"> It should be able to improve the performance or stability of the Transformer in general (e.g., in NLP tasks)\\n\\nThis work explores the use of the transformer as a memory architecture for reinforcement learning agents and we accordingly limited the experiments to partially-observable reinforcement learning domains. We believe that supervised and reinforcement learning domains are sufficiently distinct (e.g. for RL: learning rate schedules being more difficult to tune, no regularization, gradient clipping having a larger effect, higher variance of the objective function, etc) and that focusing on only one of the domains is valid. We leave applying the gated transformer to supervised learning settings to future work.\\n\\n> the experiments on the reinforcement learning is a little narrow, and many famous RL benchmarks and environments were not tested\\n\\nBeyond DMLab-30, there are a very limited number of partially-observable reinforcement learning benchmarks where memory plays a critical role in performance, and that none of these reach the difficulty of multitask DMLab-30. Standard benchmarks such as Atari-57 and the Mujoco OpenAI gym environments do not make extensive use of memory (or are even fully observable), and applying the transformer there would not be informative. For further proof, we have run the GTrXL on Atari-57 and, as expected from previous work, the transformer\\u2019s memory provides no substantial gain over a baseline LSTM agent.\", \"atari_57\": \"Model | Median Human Normalized Score \\nLSTM | 136.6 +- 3.4 \\nGTrXL | 137.1 +- 5.0 \\n\\n\\n[1] Recurrent Experience Replay in Distributed Reinforcement Learning. Steven Kapturowski, Georg Ostrovski, Will Dabney, John Quan, Remi Munos. International Conference on Learning Representations. 2019.\\n[2] Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures. Lasse Espeholt, Hubert Soyer, Remi Munos, Karen Simonyan, Volodymir Mnih, Tom Ward, Yotam Doron, Vlad Firoiu, Tim Harley, Iain Dunning, Shane Legg, Koray Kavukcuoglu. International Conference on Machine Learning. 2018.\\n[3] Multi-task Deep Reinforcement Learning with PopArt. Matteo Hessel, Hubert Soyer, Lasse Espeholt, Wojciech Czarnecki, Simon Schmitt, Hado van Hasselt. Association for the Advancement of Artificial Intelligence. 2018.\\n[4] Information Asymmetry in KL-regularized RL. Alexandre Galashov, Siddhant M. Jayakumar, Leonard Hasenclever, Dhruva Tirumala, Jonathan Schwarz, Guillaume Desjardins, Wojciech M. Czarnecki, Yee Whye Teh, Razvan Pascanu, Nicolas Heess. International Conference on Learning Representations. 2019.\\n[5] Mix & Match \\u2013 Agent Curricula for Reinforcement Learning. Wojciech Marian Czarnecki, Siddhant M. Jayakumar, Max Jaderberg, Leonard Hasenclever, Yee Whye Teh, Simon Osindero, Nicolas Heess, Razvan Pascanu. International Conference on Machine Learning. 2018.\\n[6] Highway Networks. Rupesh Kumar Srivastava, Klaus Greff, J\\u00fcrgen Schmidhuber. 2015.\\n[7] Language Modeling with Gated Convolutional Networks. Yann N. Dauphin, Angela Fan, Michael Auli, David Grangier. International Conference on Machine Learning. 2017.\\n[8] A Simple Neural Attentive Meta-Learner. Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, Pieter Abbeel. International Conference on Learning Representations. 2018.\\n[9] Squeeze-and-Excitation Networks. Jie Hu, Li Shen, Samuel Albanie, Gang Sun, Enhua Wu. IEEE conference on computer vision and pattern recognition. 2018.\\n[10] Conditional Image Generation with PixelCNN Decoders. Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, Koray Kavukcuoglu. 2016.\"}", "{\"title\": \"Response to Review #3\", \"comment\": \"Dear Reviewer 3,\", \"we_address_your_criticisms_below\": \"> the paper was based on the hypothesis of the authors that the Transformer is not stable, however, there is no comprehensive study on the unstability\\n\\nThe instability of the transformer architecture is demonstrated first-hand in our paper with the canonical variant failing to learn above a random policy\\u2019s performance in both DMLab-30 and the Memory Maze environment. These results were obtained over a large number of seeds and hyperparameter settings. Therefore the instability of the canonical transformer was not a hypothesis, it was repeatedly empirically demonstrated within our paper. We ask the reviewer to be more specific on what is insufficient about these results to make it not possible for us to claim that canonical transformers are unstable in RL domains. Beyond our own results, as referenced in the text, previous work [9] explicitly states a negative result when attempting to use a transformer for RL.\\n\\n> It would be much more convincing to give a form definition of unstability and to add experimental study and theoretical analysis to the motivation part, instead of just based on a hypothesis.\\n\\nOur paper is an empirically-driven study on using transformers in large-scale partially-observable RL environments, encompassing domains with high-dimensional observations and discrete, continuous and hybrid action spaces. Through a large number of experiments and ablations we propose a new gated variant of the transformer that is robust and easy to train (Sec. 4.3.2), and achieves the highest scores obtained so far on the DMLab-30 environment using less data than previous methods (Sec. 4.1) [1][2][3][4][5]. Our results are statistically significant, obtained over several seeds and hyperparameter settings. We believe it is unreasonable to discard these extensive empirical results only because there isn\\u2019t a strict theoretical understanding of the architecture.\\n\\n> It seems to me that these are mostly heuristics, but not verified or strongly motivated solutions.\\n\\nIt is not clear that an improved architecture design needs theoretical insight as a prerequisite to being useful to the field, as even the original transformer paper does not present substantial theoretical results. Previous work which applied gating / multiplicative interactions in feedforward architectures and demonstrated significant empirical improvements had widespread impact to their respective fields [6][7][9][10]. With respect to these architectural modifications being \\u201cnot verified\\u201d, we request the reviewer to be more specific about what is missing in our experimental results.\"}", "{\"title\": \"Response to Review #4\", \"comment\": \"Dear Reviewer 4,\\n\\nThank you for your comments and feedback. We hope to address your concerns below: \\n\\n>DMlab-30 is still a relatively new environment suite and it is difficult to place the result of this paper in the context of broader work.\\n\\nWe will make more effort to better contextualize the DMLab-30 results that we reported and intend to expand this exposition in the next version of the paper. To summarize: (1) there has been significant previous work on improving scores on DMLab-30 (and subsets of the 30 levels), see e.g. [1][2][3][4][5] for a few works, and (2) our proposed architecture achieves the best scores so far reported on this benchmark and with better data efficiency, even reporting scores better than previous results where different models were trained on each of the 30 levels [1][2].\\n\\n> In addition, the comparison is against LSTM on a new objective. While Transformer is able to beat LSTM on the same objective, it is unclear whether that is a success of the objective or the architecture.\\n\\nWe want to highlight that the LSTM trained with the V-MPO objective is state-of-the-art for LSTM architectures on DMLab-30 over all RL algorithms reported so far, and we believe this provides evidence that the objective does not somehow unfairly bias improved performance only in transformer architectures. \\n\\n> In Numpad, the transformer architecture shows an improvement over LSTM, but no comparisons are made to any other memory-based agents.\\n\\nThe toy numpad environment was mainly to highlight that our model works much better than an LSTM even within partially-observable continuous control applications. \\n\\n> adequate theoretical understanding of architecture\\n\\nOur goal in this paper was not a theoretical understanding of the optimization stability of transformers and we make no claim towards that goal. Our goal was instead to (1) demonstrate that there is significant empirical instability when the canonical transformer is used in RL applications, and (2) propose a set of modifications which seem to alleviate instability, validated in a large-scale rigorous empirical study. We believe these results will be valuable to others who want to exploit the powerful memory capability of transformers in RL. Previous work that applied gating to feedforward architectures had a substantial impact on their respective fields without a formal theoretical understanding of why gating helps [6][7][8][9].\\n\\n> \\u2026 the paper notes that gating mechanisms have been used to handle the vanishing gradients problem. The paper also notes that vanishing gradients is not an issue in transformers.\\n\\nWe believe the transformer\\u2019s recent success on long-range modeling of temporal dependencies provides evidence that vanishing gradients are not an issue for transformers along the time axis. The gating mechanism we applied was along the depth axis, not time, and therefore we believe our statements were not in direct contradiction. \\n\\n[1] Recurrent Experience Replay in Distributed Reinforcement Learning. Steven Kapturowski, Georg Ostrovski, Will Dabney, John Quan, Remi Munos. International Conference on Learning Representations/ 2019.\\n[2] Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures.\\nLasse Espeholt, Hubert Soyer, Remi Munos, Karen Simonyan, Volodymir Mnih, Tom Ward, Yotam Doron, Vlad Firoiu, Tim Harley, Iain Dunning, Shane Legg, Koray Kavukcuoglu. International Conference on Machine Learning. 2018.\\n[3] Multi-task Deep Reinforcement Learning with PopArt. Matteo Hessel, Hubert Soyer, Lasse Espeholt, Wojciech Czarnecki, Simon Schmitt, Hado van Hasselt. Association for the Advancement of Artificial Intelligence. 2018.\\n[4] Information Asymmetry in KL-regularized RL. Alexandre Galashov, Siddhant M. Jayakumar, Leonard Hasenclever, Dhruva Tirumala, Jonathan Schwarz, Guillaume Desjardins, Wojciech M. Czarnecki, Yee Whye Teh, Razvan Pascanu, Nicolas Heess. International Conference on Learning Representations. 2019.\\n[5] Mix & Match \\u2013 Agent Curricula for Reinforcement Learning. Wojciech Marian Czarnecki, Siddhant M. Jayakumar, Max Jaderberg, Leonard Hasenclever, Yee Whye Teh, Simon Osindero, Nicolas Heess, Razvan Pascanu. International Conference on Machine Learning. 2018.\\n[6] Language Modeling with Gated Convolutional Networks. Yann N. Dauphin, Angela Fan, Michael Auli, David Grangier. International Conference on Machine Learning. 2017.\\n[7] Squeeze-and-Excitation Networks. Jie Hu, Li Shen, Samuel Albanie, Gang Sun, Enhua Wu. IEEE conference on computer vision, pattern recognition. 2018.\\n[8] Conditional Image Generation with PixelCNN Decoders. Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, Koray Kavukcuoglu. 2016.\\n[9] Highway Networks. Rupesh Kumar Srivastava, Klaus Greff, J\\u00fcrgen Schmidhuber. 2015.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #4\", \"review\": \"* Summary\\nThis paper introduces architecture modifications for self-attention to stabilize transformers in reinforcement learning.\\nThe new architecture, Gated Transformer-XL, replaces the order of the layer norm blocks to preserve an identity mapping.\\nMultiple existing gating layers are proposed to replace the residual connections of transformer.\\nThe new architecture is compared against MERLIN and LSTM on DMlab-30, and further ablation studies are done on Numpad and Memory Maze.\\nIt is noted that they use the recent V-MPO objective to train LSTM and transformer.\\nTheir results show that transformers are able to learn in memory-intensive environments, with some gating combinations surpassing LSTM.\\n\\n* Decision\\nThis paper presents promising empirical results, however the experiments are limited, making it difficult to place in the broader work.\\nIn addition, the contribution is incremental and not well-motivated.\\nI would recommend a weak rejection.\\nStill, I think the paper is well written and could be improved upon.\\n\\n* Reasons\\nWhile the empirical results are impressive, they are not put into context.\\nDMlab-30 is still a relatively new environment suite and it is difficult to place the result of this paper in the context of broader work.\\nIn addition, the comparison is against LSTM on a new objective.\\nWhile Transformer is able to beat LSTM on the same objective, it is unclear whether that is a success of the objective or the architecture.\\nIn Numpad, the transformer architecture shows an improvement over LSTM, but no comparisons are made to any other memory-based agents.\\nThe hyperparameter studies on Memory Maze also show improvements in memory-related tasks, but do not help in understanding of the proposed work.\\n\\nThe choice of architecture modification is also not well motivated.\\nAs the paper mentions, initializing near identity has been shown to be important in the supervised learning literature.\\nFor the topic of this paper however, I do not think this adequetly explains the instability of transformers in reinforcement learning.\\nIn the related work section for example, the paper notes that gating mechanisms have been used to handle the vanishing gradients problem.\\nThe paper also notes that vanishing gradients is not an issue in transformers.\\nHence, it is unclear why gating would stabilize transformers for reinforcement learning.\\n\\nThe paper is overall well written and the ideas developed are clear.\\nUnfortunately, the impressive results on DMlab are not sufficient for both the lack of deeper empirical study and better theoretical motivation for the architecture modifications.\"}", "{\"experience_assessment\": \"I have published one or two papers in this area.\", \"rating\": \"3: Weak Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: N/A\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper explores a transformer for reinforcement learning. The authors demonstrate that Canonical Transformer is unstable. The authors introduce two modifications to the Canonical Transformer. The first is to move the layer normalization layer to the input stream. The second is to replace residual connections with gating layers. The experimental results show that (1) the first modification, i.e., moving the layer normalization layer to the input stream significantly stabilizes the training; (2) Gated Recurrent Unit (GRU) gating seems to be most effective gating mechanism.\\nMy decision is Weak Reject, considering the following aspects.\", \"positive_points\": \"(1) The experiments seem solid. The authors have evaluated the overall performance, as well as hyperparameters, seeds, and ablations. (2) Moving the layer normalization layer to the input stream seems to be surprisingly effective. This could be an interesting finding. (3) The paper is well organized.\", \"negative_points\": \"(1) Lack of experiments on benchmark and large environments. The authors did not evaluate their model on the widely used benchmark Atari-57. Also, it is unclear whether the proposed transformer can scale to large environments. (2) Lack of understanding of the layer normalization. The authors provide some explanations about why the reordering works, but they seem not intuitive. More analysis about why the reordering works would significantly enhance this paper.\", \"specific_questions\": \"(1) Have you tried simply removing the layer normalization layer? (2) TrXL-I moves two layer-normalization layers together. Have you tried only moving one of them? Which modification contributes more? (3) Could you provide more explanations about why the modification of the layer normalization layer works? (4) Have you experimentally validated the proposed hypothesis as to why the Identity Map Reordering, such as recording the evolution of the produced values in the submodules?\"}", "{\"experience_assessment\": \"I have published in this field for several years.\", \"rating\": \"1: Reject\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"This paper is motivated by the unstable performance of Transformer in reinforcement learning, and tried several variants of Transformer to see whether some of them can stabilize the Transformer. The experimental results look good, however, I have problems in understanding the motivation, the intuition of the proposed methods, the experimental design, and the general implication to the research community that is using the Transformer in their day-to-day research.\\nFirst, the paper was based on the hypothesis of the authors that the Transformer is not stable, however, there is no comprehensive study on the unstability, and deep understanding on the root cause of it. It would be much more convincing to give a form definition of unstability and to add experimental study and theoretical analysis to the motivation part, instead of just based on a hypothesis.\\n\\nSecond, the proposal of the new structures (e.g., reordering the layer normalization, adding the gating layer) are quite ad hoc. There is not very solid motivation and theoretical analysis on why they could solve the unstable problem of the Transformer. For example, by changing the order of layer normalization, there are direct identity mapping from the first layer to the last layer, making the information flow smoother. However, why this will make the Transformer more stable? The hypothesis and intuitive analysis are not very convincing. For another example, why replacing the residual connection with the gating layer can make the Transformer more stable? It seems to me that these are mostly heuristics, but not verified or strongly motivated solutions.\\n\\nThird, if the proposal is sound, it should not be effective only for reinforcement learning. It should be able to improve the performance or stability of the Transformer in general (e.g., in NLP tasks). However, there is no experiments and discussions regarding this.\\n\\nFourth, the experiments on the reinforcement learning is a little narrow, and many famous RL benchmarks and environments were not tested. This makes it unclear whether the proposed approach is generally effective.\\n\\n\\n**I read the author responses, however, they do not really change my assessment on the paper.\"}", "{\"comment\": \"Thanks for this paper, I'm sharing a possible related work in the context of language comprehension and dialog:\", \"https\": \"//www.aclweb.org/anthology/E17-1001/\\n\\nRegards,\", \"title\": \"possible related reference\"}" ] }

[ "Krishnamurthy (Dj) Dvijotham", "Jamie Hayes", "Borja Balle", "Zico Kolter", "Chongli Qin", "Andras Gyorgy", "Kai Xiao", "Sven Gowal", "Pushmeet Kohli" ]

[ "verification of machine learning", "certified robustness of neural networks" ]

{ "note_id": [ "2-61ZCc3Ue", "HyxxTiecsB", "rklUqQ9dsB", "rkxA4Q9uoS", "BJgnDzc_or", "BygdAZ9_jH", "B1gCq_6Msr", "rkged9hGsr", "rklQRwcfsr", "rygigJwkjH", "BkealMyptB", "r1eYLSiotr", "SylnaDVqYr" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1576798730191, 1573682104365, 1573589902480, 1573589813752, 1573589603844, 1573589456236, 1573210261909, 1573206631682, 1573197771372, 1572986611244, 1571774964535, 1571693905334, 1571600324201 ], "note_signatures": [ [ "ICLR.cc/2020/Conference/Program_Chairs" ], [ "ICLR.cc/2020/Conference/Paper1699/Authors" ], [ "ICLR.cc/2020/Conference/Paper1699/Authors" ], [ "ICLR.cc/2020/Conference/Paper1699/Authors" ], [ "ICLR.cc/2020/Conference/Paper1699/Authors" ], [ "ICLR.cc/2020/Conference/Paper1699/Authors" ], [ "ICLR.cc/2020/Conference/Paper1699/Authors" ], [ "ICLR.cc/2020/Conference/Paper1699/Authors" ], [ "~Guang-He_Lee1" ], [ "~Bai_Li1" ], [ "ICLR.cc/2020/Conference/Paper1699/AnonReviewer2" ], [ "ICLR.cc/2020/Conference/Paper1699/AnonReviewer3" ], [ "ICLR.cc/2020/Conference/Paper1699/AnonReviewer1" ] ], "structured_content_str": [ "{\"decision\": \"Accept (Poster)\", \"comment\": \"This submission proposes a black-box method for certifying the robustness of smoothed classifiers in the presence of adversarial perturbations. This work goes beyond previous works in certifying robustness for arbitrary smoothing measures.\", \"strengths\": \"-Sound formulation and theoretical justification to tackle an important problem.\\n\\nWeaknesses\\n-Experimental comparison was at times not fair.\\n-The presentation and writing could be improved.\\n\\nThese two weaknesses were sufficiently addressed during the discussion. All reviewers recommend acceptance.\", \"title\": \"Paper Decision\"}", "{\"title\": \"p.s. clarification and score revision\", \"comment\": \"We would like to follow up on the p.s request from the reviewer - we believe that this has been addressed in our revision (section 6.2, table 2) and are happy to provide any further clarifications the reviewer seeks.\"}", "{\"title\": \"Thank you for your comments\", \"comment\": \"We thank the reviewer for the careful review and encouraging feedback and hope that our updated draft and the following clarifications address the concerns raised by the reviewer.\", \"clarification_on_connection_to_cohen_et_al\": \"We have significantly revised Section 5 (Connections to prior work) and clarified the steps involved in deriving the Cohen et al result as a special case of our framework. We refer to the paragraph titled \\\"Cohen et al 2019\\\" below table 1 in the revised draft.\", \"clarification_on_practical_certification_algorithms\": \"We have revised section 3 of the paper\\nso that theorems 1 and 2 are immediately followed by descriptions of practical certification algorithms that derive from these theorems - we refer the reviewer to algorithm 2 (for the information-limited case) and algorithm 1 (for the full-information case). We have also added section 2.3 on obtaining bounds on f-divergences.\", \"clarification_on_experiments\": \"1. Full-information vs Information limited setting: We have revised definitions 2.1 and 2.2 to better reflect this distinction. We added section 4.2 where we prove that the full-information certificates are provably tighter than the information-limited ones when smoothing probabilistic classifiers. We have added section 6.1 in the experiments that clarifies the details of the experiment comparing the full-information setting with the limited-information setting for l2 perturbations.\\n2. L0 certification: To address the issue of fractional l0 values in the plot, we have removed the plots and produced tables that document the certified robustness as a function of the radius of the l0 perturbation (tables 2 and 3).\\n3. Librispeech: We have expanded the experiments section to add more details about the Librispeech dataset as well as motivation for performing these experiments in section 6.3. We are happy to answer any further questions on this.\"}", "{\"title\": \"Thank you for your comments\", \"comment\": \"We thank the reviewer for the careful review and encouraging feedback. We address the comments of the reviewer below:\\n1. We have re-run the experiments with ResNet-50 for ImageNet. We took the model trained by Lee et al and ran comparisons between our certification procedure and those from Lee et al, the results are shown in Table 2.\\n2. Yes, the Renyi divergence is not an f-divergence but is related via 1-1 mapping to an f-divergence. We have further reflected this in the paper see Section 2.2, paragraph titled \\u201cRelaxations using f-divergence\\u201d. Further, we have theoretical results (theorem 4) and experimental results (section 6.1) that use the hockey-stick divergence, and hence we think f-divergences is an appropriate unifying title.\\n3. We have reorganized sections 3 and 4 in the paper and now have a full statement of the tightness result in theorem 4 in the paper. \\n4. As shown in theorem 4, the tightness result in the information-limited holds specifically for the case of hockey-stick divergences. However, since it is not easy to compute hockey-stick divergences in general (Gaussians being a notable exception), it is also valuable to have certification procedures that use other f-divergences. We are happy to provide any further clarifications the reviewer seeks in this regard.\"}", "{\"title\": \"Thank you for your comments\", \"comment\": \"We thank the reviewer for the careful review and encouraging feedback. We respond to individual comments below:\\nTechnical\", \"on_the_distinction_between_information_limited_and_full_information_certificates\": \"We have updated the paper to clarify this distinction and refer the reviewer to section 2 (definitions 2.1 and 2.2) - we have updated the definitions. Further, section 4.1 shows theoretically how the full-information setting improves upon the information-limited setting and section 6.1 verifies this experimentally.\", \"experiments\": \"We have added more details to experiments (section 6 in the updated draft). The remaining comments are address below:\\n1. Solution of convex optimization problems: We use an off-the-shelf solver CVXPY to solve the convex optimization problems for the experiments in section 6.1. We have documented the certification running times in section 6.1 for this procedure as well (see caption of Figure 2). In the information-limited setting considered in sections 6.2 and 6.3, we can simply use the closed-form certificates computed in the table 4 (appendix section A.6).\\n2. CIFAR-10: We have added results for l1 certification on CIFAR-10 in appendix section A11. \\n3. Sec 5.1:We thank the reviewer for pointing this out, we have updated the paper to reflect this (it is now Fig. 2 in section 6.1). Please refer to the updated draft.\\n4. Sec 5.2: For L2 perturbations, the optimal certificate in the information limited setting can be obtained from Cohen et al, which is a special case of our framework that uses M=2 Hockey-Stick divergences (section 4, Corollary 5). We have removed this figure due to space constraints in the revised draft.\\n5. Comparison to [1]: The contribution of our paper is a general flexible framework for certifying robustness of smoothed classifiers while [1] focuses on using the certification method from (Cohen et al 2019) but improving the training so as to obtain classifiers with high certified accuracy. Thus, the contributions of our paper and [1] can be considered to be orthogonal, and indeed combining them is an interesting future research direction.\\n6. Sec 5.3: We have revised these results so that the model used for both certification procedures is the same and thus the clean accuracy is also the same. We have further clarified this in the paper (see Table 2, section 6.2).\\n7. Motivation behind librispeech experiments: We have added clarifying text in section 6.3 in the revised draft on the motivation.\", \"comments_on_writing\": \"We thank the reviewer for these suggestions and have incorporated them into the revised text. We have revised the abstract and contributions sections in the paper to make the specific contributions of the paper clear. The contributions section also now contains justification for the perturbations studied in the paper.\"}", "{\"title\": \"Summary of rebuttal\", \"comment\": \"We thank the blind reviewers as well as authors of the public comments for their careful reviews and feedback on our paper. We have made the following changes to our paper in response:\\n1. We have organized the technical material in the paper into three sections:\", \"section_3\": \"The main certification theorems, followed by descriptions of practical certification algorithms that follow from the theorems\", \"section_4\": \"Theoretical analysis of the certification algorithms\", \"section_5\": \"Comparisons to prior work, including a table comparing the contributions of our work relative to prior work (table 1) and detailed explanations relating our work to prior work\\n2. We have amended the abstract and contributions sections of our paper to better reflect our contributions relative to prior work and motivate the use of the perturbation measures (l0, l1 and l2) used.\\n3. We have rerun experiments on l0 certification to ensure an experimental protocol that is consistent with prior work ensuring fair comparisons can be made. We have expanded the descriptions in the experiments section to clarify precise details of our experimental protocol and highlight novel contributions from our work.\"}", "{\"title\": \"Thank you for the clarification\", \"comment\": \"Dear Guang-He,\\n\\nThank you for taking time to read our paper and pointing out these concerns.\\n\\nThe goal of our paper is to develop a general framework for certifying smoothed classifiers that allows arbitrary smoothing measures and perturbations. Indeed, in the case of l0 perturbations, the method from [1] is optimal. We will clarify this in our updated draft and change the experiments to better reflect this.\\n\\nMore concretely, we plan to make the following changes:\\n\\n1. Experiments: Regarding the experimental protocol, you are correct in pointing out the differences. We will rectify this by rerunning our experiments with a consistent experimental protocol, and if possible within the time constraints of the rebuttal period, with the same models (since our code is in tensorflow, it may take some time to convert the pytorch models and rerun the experiments).\\n\\n2. We agree that the O(d^3) computation is a one-time computation for a given smoothing measure and dataset. We will clarify this in the text. We would like to point that the quicker computation time in our approach allowed us to experiment with a broader range of smoothing measures which indeed can improve model performance significantly.\\n\\n3. We will change the sentence \\u201cFinally, these works are both restricted to the information-limited black-box verification setting where only \\u03b8a, \\u03b8b are known.\\u201d to clarify that [1] indeed looked at using more information to obtain better certificates. We would like to point out that our full-information certificates are still black-box (and hence apply to arbitrary classifiers including deep nets) and do not require access to the internals of the classifier, unlike the decision tree results obtained in [1].\\n\\n4. (\\u03b8a, \\u03b8b) are estimated using the same procedure as Cohen et al (the Clopper-Pearson CI test). We use 10M samples for MNIST, 1M samples for CIFAR-10 and 100K samples for ImageNet - these are sampled independently for every image.\"}", "{\"title\": \"Thanks for pointing this out\", \"comment\": \"Dear Bai,\\nThank you for bringing this work to our attention. \\n\\nWe agree that the Renyi divergence case is indeed closely related and we are preparing an updated draft clarifying the distinction and relationship with [1].\\n\\nIn particular, we believe that our work generalizes [1] in three ways:\\n1. Our framework handles arbitrary smoothing measures (not just Gaussian and Laplacian).\\n2. We allow arbitrary f-divergence measures. In particular, for certifying l2 perturbations using Gaussian smoothing, we show that the hockey-stick divergences are necessary to obtain optimal certificates (recovering the work of Cohen et al).\\n3. Even in the specific case of Renyi divergences, we can obtain better certificates by using convex combinations of Renyi divergences.\\n\\nWe will update the draft to reflect this and are happy to answer any other questions.\\n\\nRegards,\\nAuthors\"}", "{\"title\": \"Some comments\", \"comment\": \"Dear Authors,\\n\\nThank you for the effort in preparing this work, but we have some questions and concerns about the paper.\\n\\nTaking numbers directly from another paper [1] for comparison is appropriate only if the protocol was the same. This doesn\\u2019t seem to be the case. For example, [1] used 0.999 confidence level with 100K samples, while in this paper confidence level is 0.99 with 10M samples. So one could get improvements even with exactly the same certification algorithm. Also, since the difference is really the certification algorithms rather than models, are the models the same?\\n\\nBesides, we would like to clarify your descriptions of [1]. For instance, \\u201cthey require an O(d^3) computation (where d is the input dimension) to certify smoothness to L0 perturbations (in addition to the cost of estimating \\u03b8a, \\u03b8b by sampling).\\u201d The O(d^3) computation only requires to be done once for any given dataset. After the computation is done once, the certification takes constant time for any input data and model for the dataset: testing whether \\u03b8a > some pre-computed threshold.\\n\\nThe description \\u201cFinally, these works are both restricted to the information-limited black-box verification setting where only \\u03b8a, \\u03b8b are known.\\u201d is incorrect. To our best knowledge, [1] is the first one to extend robustness certificates for randomized smoothing beyond only knowing \\u03b8a, \\u03b8b. For example, given a decision tree predictor under the smoothed distribution, [1] showed that the exact certificate can be efficiently computed.\\n\\nFinally, could you please explain how (\\u03b8a, \\u03b8b) are estimated in the experiments? For each image, did you use the same 10M samples for estimating \\u03b8a and \\u03b8b?\\n\\n[1] \\u201cTight Certificates of Adversarial Robustness for Randomly Smoothed Classifiers.\\u201d (The old title is \\u201cA Stratified Approach to Robustness for Randomly Smoothed Classifiers\\u201d), NeurIPS 2019.\"}", "{\"title\": \"Closely Related Previous Work\", \"comment\": \"Dear Authors,\\n\\nThank you for the interesting work. I would like to point out that using Renyi divergence to calculate certified bound for randomized smoothing has been extensively studied in https://arxiv.org/abs/1809.03113 [1].\\n\\nAlthough this paper aims to generalize the framework to general f-divergence, Renyi divergence is particularly used in the experiments, which is the exact focus of [1]. The comparison of Renyi divergence based bound and the one obtained from PixelDP has also been done in [1].\\n\\nI suggest the authors discuss the connection and distinction between their work and [1] for the comprehensiveness.\\n\\n[1] Li, B., Chen, C., Wang, W., & Carin, L. (2018). Certified Adversarial Robustness with Additive Noise. arXiv preprint arXiv:1809.03113.\"}", "{\"experience_assessment\": \"I do not know much about this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #2\", \"review\": \"This paper extends existing work on certified robustness using smoothed classifier.\\nThe fundamental contribution is a framework that allows for arbitrary smoothing measure, in which certificates can be obtained by convex optimization. \\nA good number of technical contributions\\nFramework for certificate under arbitrary smoothing measure -> Theorem 1, and proof in A.4 (good use of duality)\\nFull calculation/result of certificates under different divergneces in Table 1.\\nReasonable set of empirical evidence. \\n\\nOverall a lot of good things to be said, below are some questions/comments that could improve the paper:\\n*Technical*\\nPersonally, I cannot get a lot of value out of the distinction between full-information and information-limited certification. It\\u2019d be great if I can get some clarification on this. \\n*Experiments*\\nGenerally, more details of the experiments should be included. \\nHow are the convex optimization problems actually solved (e.g., what methods/tools)? \\nHow much more computational overhead is there?\\nOddly, seems like we\\u2019re missing CIFAR10 results completely. \\n\\nSec 5.1., \\nWhat does each dot in Figure 3a represent?\\nSec 5.2, \\nSomewhat strangely, Figure 3b is results on l0 perturbation, but not l2. What happens for l2 when we use M>1? How does this compare to other extensions (likely the SOTA) like [1]?\\nSec 5.3,\\nIt is unclear from the writing whether the comparisons to previous works were done on the same ResNet architecutre with the same clean accuracy. Please clarify. \\nI\\u2019m not sure why we need the Librispeech results. It\\u2019s not motivated clearly. Also, from writing it seems the adversary zeros out a segment. It\\u2019s unclear if this is a reasonable kind of attack to expect on speech. If I block out a segment of the speaker, we probably don\\u2019t expect any system to do well on speaker recognition. I suggest removing this result, or somehow make it a lot more better motivated/conducted. Clarify if I missed reasons why simply showing your method works on speech is impressive.\", \"here_are_suggestions_on_writing\": \"Contribution --- in both the abstract and introduction, the experimental results should be stated clearer. Be more specific, e.g., \\u201cShow SOTA certified l2 robustness on X,Y,Z, establish first certified robustness on Librispeech, and first results on certified l0, l1 robustness on A,B,C.\\u201d\\nMore broadly in the introduction, please motivate why \\u201cadversarial attacks as measured by other smoothing measure is important\\u201d. Past studies focus on l2-norm not just because they are do-able, but also white noise (which is naturally measured by l2-norm) is something to expect in practice. Justify why l0, l1 would also be important.\\n\\nI recommend accepting this paper, but would do so with more passion if some of the comments/questions can be addressed.\\n\\nBest,\", \"reference\": \"[1] https://arxiv.org/abs/1906.04584\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #3\", \"review\": \"Summary: This submission proposes a unified framework for black-box adversarial certification. Based on cohen's and lee's result, it extends the \\\\lambda_TV distance to more general f-divergence. Besides, some other techniques, e.g. reference distribution, are also included in the framework. For L_0 and L_1 perturbation, the proposed framework comes to better results than all the related methods.\", \"strengths\": \"[+] Detailed analysis and theorem.\\n[+] Describing the background and conclusions clearly.\", \"weaknesses\": \"[-] In the experiment section, for ImageNet, baselines use ResNet-50 but the authors use ResNet-152. I wonder whether this is fair enough.\\n[-] Renyi divergence is not a proper f-divergence. Therefore, maybe the title should be changed.\\n[-] The theorem part contains too many details and some important parts (e.g. tightness) are in appendix. It should be re-organized.\", \"questions\": \"[.] In Appendix A.7, Lagrangian strong-duality has been used to show the tightness, but I wonder whether the tightness holds for any kind f-divergence? If the authors can write down some theorem about the tightness, it will be better.\\n\\np.s. Could you give me some quick responses to my questions (e.g. ResNet-152 v.s. ResNet-50) ? Once you convince me, I would revise the score to 6. If not, I will turn down the score to 3.\"}", "{\"experience_assessment\": \"I have read many papers in this area.\", \"rating\": \"6: Weak Accept\", \"review_assessment\": \"_checking_correctness_of_derivations_and_theory: I assessed the sensibility of the derivations and theory.\", \"title\": \"Official Blind Review #1\", \"review\": \"The paper introduces a generalization of the randomized smoothing approach for certifying robustness of black-box classifiers, allowing the smoothing measure to be an arbitrary distribution (whereas previous work almost exclusively focused on Gaussian noise), and facilitating the certification with respect to different metrics under the same framework.\\n\\nGiven the wide interest in certified robustness based on randomized smoothing, the generalizations considered in this paper could have a high potential. I found the motivation, the definition of the framework and the statement of the main theoretical results, up to Section 3.2, very clear. The following sections were not as well organized, in my opinion, and need improvement. In particular, I found it not very clear how the framework can applied to recover previous results, in particular the results from Cohen et al. (2019). All the ingredients seem to be there, among Theorem 3, the result from Balle & Wang (2018), and Corollary 5, but the arguments could be presented in a better organized way.\\n\\nWhat I also found is missing is a clear descriptions of practical algorithms for applying the framework. Again, most of the ingredients seem to be there (e.g. Table 1 in the main body, Lemma 6 as well as the discussion about sampling mechanism in the Appendix), but they are not well organized. Presumably, the same statistical methodology as in Cohen et al. is used to obtain estimates of theta_a and theta_b, but this doesn't seem to be clearly stated anywhere. How is certification practically performed for intersections of contraint sets as introduced at the end of section 2? How the full-information certification can be applied based on empirical estimates is unclear to me, too.\\n\\nFinally, the description of the experiments need improvements: For instance, in Section 5.1, which smoothing measure was used? Was the ResNet classifier trained using samples from this measure? Which sets of f-divergences were used? What does Figure 3 (a) exactly show? Do the blue points correspond to the 50 data samples? The number of random samples for computing empirical estimates and the confidence bounds are missing.\\nIn Figure 4 (a), as the l_0 norm counts the number of altered pixels I don't understand why the certified accuracy varies e.g. for epsilon *between* 0 and 1.\\nThe experiment on the Librispeech model seems interesting, but the paper does not contain sufficient information to understand and assess the experimental set-up or the results.\\n\\nIn summary, while I believe that the generalizations proposed in this paper have potential, in its present form the manuscript doesn't describe clearly and accurately enough how the framework can be applied to recover previous theoretical results or perform certification in practice.\\n\\n----------\\n\\nI acknowledge I have read the authors' response. As it addresses my main concerns I have changed my rating from \\\"weak reject\\\" to \\\"weak accept\\\".\"}" ] }