sentence_id text position 397 iclr19_304_3_10 In the current version, you seem to define overfitting on-the-fly while defining your criteria. NA 888 midl19_40_3_6 The authors also compare the annotation time between points, bounding boxes and full supervision, which really highlight the impact of their method (x10 speedup). POS 1369 neuroai19_54_3_8 " 2) I agree with a concern raised by reviewer 3: It's difficult to see a 1-layer network as a ""mechanistic explanation"" of a 3-layer network." NEG 872 midl19_25_3_2 The authors could spend a little more effort on explaining the intuition behind conditional versus unconditional labels and the advantages of each. NEG 635 iclr20_2094_1_9 Notably, several classes of geometric bin packing problems admit polynomial-time approximation algorithms (for extended surveys about this topic, see e.g. Arindam Khans Ph.D. NA 146 graph20_39_3_7 The description of each patient drags on a little long, and much of it does not become useful after in the later sections, since particular medical history is not referenced in later sections. NEG 1257 neuroai19_29_1_0 While the question of how neural networks may act over concept space is important, I dont think the approach used by the authors correctly adress this question. POS 60 graph20_29_3_28 To the best of my understanding, these are not percentages of prediction error (e.g. going from 50 to 55 is a 10% increase), which would be more ok. NA 113 graph20_36_1_27 The discussion mentions participants who felt the drawing were similar while the metric showed they were not. NEG 128 graph20_39_2_10 How representative is it, whats the bigger picture, can it be generalised to other not known scenarios? NEG 358 iclr19_242_2_36 However, I am not convinced by the experiments that the good performance is from the proposed method, not from the N times more augmented samples. NEG 1342 neuroai19_37_3_21 Our challenge is to understand how this occurs. NA 1304 neuroai19_34_2_6 Methods described clearly and in good detail. POS 127 graph20_39_2_9 Thirdly, from the discussion of the findings, quotes appear unpacked. NEG 750 iclr20_727_1_5 " Though the final proof is in the pudding, and the addition of the VAE to model the base distribution yields promising results, the only justification for it in the paper is to create a more ""expressive"" model." NEG 1018 midl19_56_3_7 " I would recommend weakening or at least toning down certain ""marketing"" claims like ""3 times finer than the highest resolution ever investigated in the domain of voxel-based shape generation"", or ""the finest resolution ever achieved among voxel-based models in computer graphics""." NEG 399 iclr19_304_3_12 Detailed: Page 3, last paragraph: Why did you not use bias terms in your model? NA 54 graph20_29_3_22 " The level of detail argued here seems quite artificial, e.g. ""If designers want a hyperlink to have a 77% success rate""." NEG 262 iclr19_1091_1_0 This paper discusses State Representation Learning for RL from camera images. NA 15 graph20_25_2_15 Thus, the paper should better position the proposed design/prototype within this design space. NEG 1134 midl20_56_4_16 The method is novel with extensive experiments. POS 1241 neuroai19_23_1_10 The question of how the visual world is represented in the brain is an essential question in neuroscience as well as for building successful machine learning techniques for artificial vision. NA 758 iclr20_727_1_13 " after setting up the expectation that the marks will not be modelled initially, up till footnote 2 on page 7.""" NEG 1240 neuroai19_23_1_9 Results were presented quite clearly, although datasets and methods rely entirely on previously published work, such that digging into previous work on PredNet and the Algonauts project was necessary for a full understanding. NEG 617 iclr20_2046_2_22 In fact, it is performed under the exact assumption where the theoretical analysis is done for the A*MCTS. NA 671 iclr20_305_3_3 " The leader sends messages to followers, an ""event"" is a pair (timestep, message of leader to a follower)." NA 204 graph20_56_1_11 A similar problem occurred with a critical aspect of the brushing technique: direction. NEG 751 iclr20_727_1_6 There are multiple ways of increasing the expressiveness of the underlying distribution: moving from RNNs to GRU or LSTMs, increasing the hierarchical depth of the recurrence by stacking the layers, increasing the size of the hidden state, more layers before the output layer, etc. A convincing justification behind using a VAE for the task seems to be missing. NEG 662 iclr20_2157_3_9 As I understand it, real improvements in predicting clinical variables has not been shown to be reproducible so this would be a significant claim of this paper. NEG 791 iclr20_880_2_11 Since the authors are using inner matrices with a number of dimensions higher than the number of dimensions of the original matrix, there is no approximation and, then, no selection of features or feature combinations. NA 633 iclr20_2094_1_7 Moreover, BPPs have been extensively studied in theoretical computer science, with various approximation results. NA 189 graph20_53_2_16 The actual discussion of the results unfortunately is very limited (especially because large parts of it consist of qualitative reporting), and are mostly a summary, rather than a contextualization of the results within existing work, or statements on implications of the results. NEG 1358 neuroai19_53_1_8 One part that would have been nice to clarify is the relative role of random feedback vs eligibility traces in successful network performance. NEG 734 iclr20_720_2_7 Why reward decomposition at the lower levels is a problem instead of a feature isn't totally clear, but this criticism does not apply to Option-Critic models. NEG 232 graph20_61_2_10 Obtained results are supported with clear metrics. POS 765 iclr20_855_3_1 The 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). NA 252 graph20_61_2_30 LIMITATIONS AND FUTURE WORK The limitations are mainly focused on the specificity of project requirements to one University in Canada, the small sample size of participants to evaluations. NA 1165 midl20_77_4_12 How were these trajectories formed? NA 930 midl19_49_1_26 The hyper-parameters of autoencoder and the recon decoder should be more clearly stated for reproducibility. NEG 1364 neuroai19_54_3_3 I agree these are good goals, and I think some progress is made, but that progress seems somewhat limited in scope. POS 328 iclr19_242_2_2 I think it is an interesting idea, but the current draft does not provide sufficient support. NEG 454 iclr19_659_2_1 Its main idea is updating multiple Q-functions, instead of one, with independently sampled experience replay memory, then take the action selected by the ensemble. NA 705 iclr20_526_3_26 I would like to see this curve extended until we start to see signs of overfitting. NEG 1215 neuroai19_2_2_5 The theoretical results stated are nice to have. POS 788 iclr20_880_2_8 Without non-linear functions, equations (1) and (2) describe a classical matrix factorization like Principal Component Analysis. NA 336 iclr19_242_2_12 For example, a simple thing to do is t0 separately train networks with standard setting and then ensemble trained networks. NA 12 graph20_25_2_12 For example, an equally feasible alternative is a design that uses a small physical numerical keyboard that users can carry with them and enter passwords even from their pockets (the haptic feedback that such a keyboard would enable would allow such interaction). NA 370 iclr19_261_3_6 Im not sure how impressed I should be by these results. NEG 1311 neuroai19_34_2_13 " Typo page 4 line 158: ""pray"" >> ""prey""""" NA 504 iclr19_938_3_9 Authors do not visualize the attention (as is common in previous work involving attention in e.g., NLP). NEG 1399 neuroai19_59_3_24 Looking into the nuances of the explored phenomena may provide new information for the field. NA 67 graph20_29_3_35 " Similarly, their 2D tasks showed only small differences in error rate, up to 2% at most.""" NA 1379 neuroai19_59_3_4 A short discussion of other training algorithms (such as surrogate gradient or surrogate loss methods) and why the given one was chosen instead would have been helpful. NEG 83 graph20_35_1_9 Finally, the discussion would benefit from some more general discussion, before the limitations, on the overall findings and what they mean for mind mapping and similar applications moving forward. NEG 721 iclr20_57_3_5 Strenghts: + the paper proposes a reasonable way to try to improve accuracy by identifying hard-negative examples + the paper is well written, but it would benefit from another round of proofreading for grammar and clarity Weaknesses: - performance of the proposed method highly depends on labels of hard-negative examples. NEG 679 iclr20_526_3_0 This paper presents a black-box style learning algorithm for Markov Random Fields (MRF). NA 318 iclr19_1399_1_11 The proposed method introduces a lot of complexity for very small gains. NEG 325 iclr19_1399_1_18 " Experiments with ImageNet or some other large data set would be advisable to increase significance of this work. """ NEG 1244 neuroai19_26_1_0 The proposed model is essentially a constrained/specific parameterisation within the broader class of 'context dependent' models. NA 27 graph20_26_3_2 pseudo-url The submitted modifications show a marked improvement in the exposition of the work. POS 199 graph20_56_1_6 As these two metrics deliver different candidates, the resulting set of trajectories is provided to the user in a set of small multiples illustrating the selected trajectories and sorted by similarity; the user can refine selections, although it was not clear how. NEG 856 midl19_14_2_30 A similar experiment can be made using other data sets with red/bright lesions (e.g. e-ophtha, pseudo-url) or optic disc annotations (e.g. REFUGE database, pseudo-url). NA 46 graph20_29_3_14 One example, in the last paragraph before EXPERIMENTS (p. 4), a point is made that goes like this: - a lack of effect might be due to A values that are too close to each other, - even if A should in fact have an effect according to some model (Eq. NA 960 midl19_51_2_0 The authors combine DL and computer vision methods to digitally stain confocal microscopy images to generate H&E like images. NA 363 iclr19_242_2_41 Regarding the theoretical part, I still do not follow the authors' explanation. NEG 468 iclr19_866_1_1 The first of two modules is responsible for learning a goal embedding of a given instruction using a learned distance function. NA 51 graph20_29_3_19 It seems that this sort of design issues can be solved using threshold values under which users simply cannot accurately acquire a target. NEG 530 iclr20_1042_2_8 But equation (2) shows a loss with no weighting. NA 207 graph20_56_1_14 But this is not clear. NEG 907 midl19_49_1_2 This work has a remarkable clinical value. POS 34 graph20_29_3_2 However I found the theoretical argument to use the DGDM in screen-to-screen pointing quite hard to follow, even though it is the main point of this article. NEG 385 iclr19_261_3_27 The red one! NA 1179 midl20_90_2_0 In this work, the authors purposed a new deep neural network architecture for detecting injuries/abnormalities in the knee. NA 40 graph20_29_3_8 Checking can also come up negative, and that is ok. NA 63 graph20_29_3_31 2-mm targets on a touch device could definitely count as difficult. NA 164 graph20_43_1_11 " Based on the above, I feel the paper is marginally below the acceptance threshold.""" NA 1302 neuroai19_34_2_4 Very well written. POS 1249 neuroai19_26_1_5 Perhaps compare to reference models [11] or [10] rather than a 'vanilla' RNN, as this amounts to not using any prior information about the task (which, by construction, we 'know' is useful). NEG 561 iclr20_1493_2_19 In general this list is not comprehensive either: there are many relevant connections to the robustness-accuracy tradeoff (pseudo-url, pseudo-url), and other works. NA 735 iclr20_720_2_8 " For Option-Critic models the authors claim that ""Rather than the additional inductive bias of temporal abstraction, we focus on the investigation of composition as type of hierarchy in the context of single and multitask learning while demonstrating the strength of hierarchical composition to lie in domains with strong variation in the objectives such as in multitask domains.""" NA 1162 midl20_77_4_9 Other specific suggestions: Section 2: region of interest (ROI) performing motions does not make sense to me. NEG 573 iclr20_1493_2_33 This is in stark contrast to real datasets, where there seem to be many different ways to perfectly separate say, dogs from cats, and the variance of the data seems to be very heavily concentrated in a small subset of directions. NEG 810 iclr20_934_1_7 The difference is that the proposed method learns a multi-channel representation and uses the attention technique to aggregate the multi-channel representation. NA 228 graph20_61_2_6 ABSTRACT Abstract provides information that is ideally expected: one sentence of context, summary of contribution, explanation of system and methodology. POS 457 iclr19_659_2_4 The main ideas and claims are clearly expressed. POS 1391 neuroai19_59_3_16 I think the interesting part may be in quantifying just how much of a difference there is between short and long timescale neurons -- for instance, does task-relevant information in both neuron groups fall off in a way that can be well predicted by their intrinsic time constants? NEG 1362 neuroai19_54_3_1 " Operationally, I'm not quite sure how these are different, so, to me this goal is roughly ""be explainable"", and progress towards it could be measured e.g. in MDLs." NA 259 iclr19_1049_1_4 However, the form of knowledge is limited and simple. NEG 1081 midl20_100_1_32 It is fine that you give your method a name (although I personally dislike it), but a bit weird not to explain it. NEG 862 midl19_14_2_36 The first 10 lines contains too much wording for a statement that should be much easier to explain. NEG 500 iclr19_938_3_5 Pro - MAAC is a simple combination of attention and a centralized value function approach. POS 409 iclr19_304_3_23 And how do you use that later? NEG 806 iclr20_934_1_2 The experimental result demonstrates some improvement over existing methods. NA 922 midl19_49_1_18 However, authors use Jaccard coef. NA 354 iclr19_242_2_32 The main baseline that has been compared is the standard small-batch training. NA 762 iclr20_76_2_3 The paper is well written, tghe major issue of this paper is the lack of comparison with other previous methods. NEG 1296 neuroai19_32_1_22 Is this an issue of spatial scale? NA 506 iclr19_938_3_11 " Reproducibility - It seems straightforward to implement this method, but I encourage open-sourcing the authors' implementation.""" NA 351 iclr19_242_2_29 after rebuttal ==================== I appreciate the authors' response, but I do not think the rebuttal addressed my concerns. NEG 416 iclr19_304_3_30 What you actually do here is you present 3 different general criteria that could potentially detect overfitting on label-randomized training sets. NA 760 iclr20_76_2_1 In this scheme, a Teacher generates data according to a Gaussian random field, and a Student learns them via kernel regression. NA 118 graph20_39_2_0 The paper explores the possibilities of reviewing and visualising patient-generated data from a range of stakeholders consisting mainly of healthcare providers and patients. NA 581 iclr20_1493_2_42 Overall, this paper is a very promising step in studying adversarial robustness, but concerns about discussion of prior work, discussion of experimental setup, and conclusions drawn, currently bar me from recommending acceptance. POS 1171 midl20_85_3_3 However, I have following concerns: 1. NA 431 iclr19_495_1_3 However, I find the paper written in a way assuming readers very familiar with related concept and algorithms in reinforcement learning. NEG 945 midl19_51_1_13 This paper still represent a niche application of a more general DL technique that has been already used for a large number of similar applications. NEG 447 iclr19_601_3_1 It is meant to generate overall images as image slice sequences with memory and computation economy by using a Multidimensional Upscaling method. NA 1103 midl20_119_2_4 Comment: It would nice if the authors could also show some visualisations of the latent space, with comparisons between with and without the constraint. NEG 430 iclr19_495_1_2 Thus I think this method is itself interesting. POS 194 graph20_56_1_1 " These are notoriously difficult to select directly due to issues of occlusion and the ""hairball"" effect when there are many trajectories intertwines, as is the case with eye tracking, network, or flight trails data." NA 1190 midl20_96_3_1 This is done with manual labelling and a ResNet-18. NA 740 iclr20_720_2_13 It certainly does not seem justified to me to just assume this framework and disregard past successful approaches even as a comparison. NEG 934 midl19_51_1_2 The authors propose to use a cycle-GAN to shift the distribution of CM images towards more standard H&E images which are easier to interpret. NA 283 iclr19_1091_1_21 Please provide some extra information on how it is calculated. NA 210 graph20_56_1_17 However, the video alludes to something not mentioned in the paper about directionality: only the Pearson algorithm identifies direction, and even from the video it was not clear how the user selected it. NEG 602 iclr20_2046_2_7 However, it does not clearly explain the key insights of why it could perform better. NEG 1301 neuroai19_34_2_3 Use of the same spatial transformer model with an interchangeable bank of input features is elegant. POS 697 iclr20_526_3_18 Experiments: The authors show the empirical advantages offered by the proposed method over the existing literature. POS 638 iclr20_2094_1_12 Note that the 2D Knapsack problem with rotations admits a 3/2 + \epsilon - approximation algorithm (Galvez et. al., FOCS 2017). NA 111 graph20_36_1_25 But in the other condition, participants could have perform just as well, with a slight rotation or translation. NEG 278 iclr19_1091_1_16 Minor points: - The choice for these tasks is not motivated well. NEG 999 midl19_52_2_18 Are the results on Table 1 heavily dependent on use of these masks? NEG 1337 neuroai19_37_3_15 Its more a series of statements than a cleverly woven argument. NEG 432 iclr19_495_1_4 Thus although one can get the general idea on how the method works, it might be difficult to get a deeper understanding on some details. NEG 1056 midl20_100_1_7 The methods section lacks details for reproducing the work. NEG 237 graph20_61_2_15 APPLICATION BACKGROUND This section conveniently introduces domain-specific terms and thus contributes to make the paper standalone in understanding the context. POS 313 iclr19_1399_1_6 These were good data sets a few years ago and still are good data sets to test the code and sanity of the idea, but concluding anything strong based on the results obtained with them is not a good idea. NEG 1377 neuroai19_59_3_2 The work would benefit from more detailed discussion of the training algorithm that provides some indication that the results aren't unduly sensitive to these details. NEG 41 graph20_29_3_9 These results remain valid, even if the proposed approach is not as context-independent as hoped. NEG 712 iclr20_526_3_33 Is that the case? NA 1265 neuroai19_3_3_0 This intriguing study proposes to modify the classical Q-learning paradigm by splitting the reward into two streams with different parameters, one for positive rewards and one for negative rewards. NA 533 iclr20_1042_2_11 2) Theoretical inconsistencies Although the system might work overall, two things seem to be technically incorrect: - The decoder and classifier are expected to approximate the distribution of training data according to the authors (for valid generative replay). NEG 611 iclr20_2046_2_16 In particular, the probability in the second term of Theorem 1 is hard to parse. NEG 43 graph20_29_3_11 I think this part needs to be drastically shortened or even removed, in favor of a more realistic discussion about generalization---and possible lack thereof. NEG 125 graph20_39_2_7 Its well known that 'chronic conditions might take a different form and thus interpreted within a particular context; this makes the contribution of the paper marginal, as one would expect a clear articulation of how the method is chosen to fit into the context of the wider literature on similar issues and ultimately the nature of the study participants. NEG 1177 midl20_85_3_11 " Last line of section 1: ""it can distinguish distributional versus data uncertainties""." NA 868 midl19_14_2_44 International Conference on Medical Image Computing and Computer-Assisted Intervention. NA 205 graph20_56_1_12 The authors state directionality is a critical advantage of their brushing technique, but never actually stipulate how direction is specified in the original share definition. NEG 985 midl19_52_2_4 I believe the experiments are thorough and well designed to back the claims of the paper. POS 1382 neuroai19_59_3_7 I feel that more tools should have been used to further support or push the results. NEG 569 iclr20_1493_2_28 The paper justifies the adversarial vulnerability of the Linear SVM by arguing that the Bayes-optimal classifier is not in the Linear SVM hypothesis class, which makes sense. POS 1374 neuroai19_54_3_13 " I believe this paper is addressing questions that many of the workshop attendees will find interesting.""" POS 484 iclr19_866_1_17 How different are the familiar and unfamiliar instructions? NEG 446 iclr19_601_3_0 Authors propose a decoder arquitecture model named Subscale Pixel Network. NA 700 iclr20_526_3_21 Its absence seems like a serious omission. NEG 995 midl19_52_2_14 5- How is the complex component of the signal concatenated into a channel ? NEG 621 iclr20_2046_2_26 Other comments: It is assumed that the noise of value and policy network is zero at the leaf node. NA 1247 neuroai19_26_1_3 The model description is nice and clear. POS 718 iclr20_57_3_2 The authors consider using 3 different objective functions: L1, the original cross entropy loss; L2, capturing the shared features in positive and hard-negative examples as regularizer of L1 by introducing a new label z; L3, a three-class classification objective using softmax. NA 747 iclr20_727_1_2 " To further increase the expressive power of the normalizing flow, they propose using a VAE to learn the underlying input to the ""Flow Module""." NA 699 iclr20_526_3_20 MNIST, in particular, is a well studied dataset that many readers will be able to easily interpret. NA 1049 midl20_100_1_0 Overall, the quality of the paper is fair. POS 540 iclr20_1042_2_18 But regardless of this, both models are inconsistent.) NEG 462 iclr19_659_2_9 Since the method requires updating multiple Q-functions, it may cost much more time for each RL time step, so Im not sure whether the ensemble method can outperform the non-ensemble one within the same time period. NEG 1182 midl20_90_2_3 The paper is written very well, the implementation details are provided to help reproducing the results. POS 728 iclr20_720_2_1 I am quite confused about what exactly the author are claiming is the core contribution of their work. NEG 1025 midl19_56_3_14 " Smooth shape interpolation by traversal of the latent space was also demonstrated, and some of their latents also corresponded to reasonable variations in anatomical shape, without being ""restricted"" to statistical modes of variation as discussed here." NA 825 midl19_13_2_10 " It would be interesting to have the author's point of you on the less than optimal results, and how they plan to improve it.""" NA 9 graph20_25_2_9 However, there are two main weaknesses: 1) the submission narrowly focuses on bend passwords, and 2) the evaluation compares BendyPass against only one baseline. NEG 941 midl19_51_1_9 I think this joint training might result in even better outcomes. NA 324 iclr19_1399_1_17 In summary, it is not a bad paper, but the experimental results are not sufficient to conclude that much. NEG 648 iclr20_2094_1_22 For example, in Eq (1) what are the dimensions K and V? NEG 1195 midl20_96_3_6 No effort has been made to fuse the proposed pipeline into a medical-image analysis specific methodological contribution. NEG 610 iclr20_2046_2_15 It does not give the explicit relations of the sample complexity with respect to different quantities in the algorithms. NEG 179 graph20_53_2_6 " How likely are designers of 3D objects to include such ""internal faces""; is this common?" NA 848 midl19_14_2_22 To the best of my knowledge, it has the highest performance in the DRIVE data set compared to several other techniques. NA 1251 neuroai19_26_1_7 Paper is clear and quite readable. POS 263 iclr19_1091_1_1 Specifically, it proposes to use a state representation consisting of 2 (or 3) parts that are trained separately on different aspects of the relevant state: reward prediction, image reconstruction and (inverse) model learning. NA 1259 neuroai19_29_1_2 I dont see how the current work adds more clarity to this research direction. NEG 1138 midl20_70_4_0 This paper presents a multi-label classification framework based on deep convolutional neural networks (CNNs) for diagnosing the presence of 14 common thoracic diseases and observations in X-rays images. NA 970 midl19_51_2_10 3- Please provide an evidence to support the positive effect of choosing an augmentation of size 512x512 after 50 epochs in Section 3.2. NEG 1336 neuroai19_37_3_14 " Sure neuromorphic systems are coming, but not definitely not with moderate expenditure of resources and effort"" While it covers important ground, I think the arguments need more refinement and focus before they can inspire productive discussion." NEG 1269 neuroai19_3_3_4 The figures are hard to parse because of the very short captions. NEG 518 iclr19_997_3_11 Please provide more details about this point. NA 123 graph20_39_2_5 Although sections 2 attempts to situate the research question into the context of varied perspectives, a better justification of the stake for the field would have been made clearer had it being the section doesn't read as if its an analysis of prior data, and not of related works. POS 1260 neuroai19_29_1_3 The main point relies purely on a visual representation of the top PCs of the penultimate layer of a CNN, which I believe is insufficient. NEG 350 iclr19_242_2_28 Related works: Smith et al. 2018 Don't Decay the Learning Rate, Increase the Batch Size. NA 1106 midl20_127_4_1 The authors show that the AF-Net is more robust compared to the U-Net and M-Net for AFV measurement. NA 294 iclr19_1291_3_7 " I have some concerns regarding their method and the experiments which are brought up in the following: Method: In a non-fully-cooperative environment, sharing hidden state entirely as the only option for communicate is not very reasonable; I think something like sending a message is a better option and more realistic (e.g., something like the work of Mordatch & Abbeel, 2017) Experiment: The experiment ""StarCraft explore"" is similar to predator-prey; therefore, instead of explaining StarCraft explore, I would like to see how the model works in StarCraft combat." NEG 478 iclr19_866_1_11 While there are advantages to training the modules separately, there is a risk that they are reasoning over different portions of the goal space. NEG 1090 midl20_108_3_6 The experiments are clearly explained and the results are well presented. POS 651 iclr20_2094_1_25 Also in the algorithm, what are l_i, w_i and h_i? NEG 716 iclr20_57_3_0 This paper is aimed at tackling a general issue in NLP: Hard-negative training data (negative but very similar to positive) can easily confuse standard NLP model. NA 1131 midl20_56_4_12 The authors do not compare the inference speed of the proposed method with others. NEG 1294 neuroai19_32_1_20 One of their stated novel contribution was that their filters were convolutional but they do not discuss the potential connection convolutional filters have to transformation of features which seemed like a gap. NEG 1357 neuroai19_53_1_7 While eligibility traces have received some attention in neuroscience their relevance to learning has not been thoroughly explored, so this paper makes a welcome contribution that fits well within the workshop goals. POS 912 midl19_49_1_7 I'm concerned that this would make the readers misunderstand the data are shape-models (point cloud dataset) before the description of dataset in Sec. 2. NEG 467 iclr19_866_1_0 The paper proposes a modular approach to the problem of mapping instructions to robot actions. NA 489 iclr19_866_1_22 The paper initially states that this distance function is computed from learned embeddings of human demonstrations, however these are presumably instructions rather than demonstrations. NEG 677 iclr20_305_3_13 Additional feedback with the aim to improve the paper. NA 687 iclr20_526_3_8 While I agree with the statement as such, the GAN development makes a stronger statement about the nature of the learning trajectory. NEG 857 midl19_14_2_31 I think this is a key experiment, really necessary to validate if the method is performing well or not. NEG 637 iclr20_2094_1_11 c) According to the problem formulation and the experiments, it seems that the authors are studying a restricted subclass of 2D/3D bin packing problems: there is only one bin, so (it seems that) the authors are dealing with geometric knapsack problems (with rotations). NA 134 graph20_39_2_16 From the guidelines outlined in section 9, it is hard to pinpoint new learning that the paper provides to the visualisation of subsequent design practices apart from restating well-known design insights. NEG 1093 midl20_108_3_9 Good and convincing results when compared to competing methods * Strong validation * It is a shame that the Kaplan-Meier estimator was not repeated for all baselines to further illustrate the strength of the multi-task features * There are many more TUPAC16 results [pseudo-url. NEG 1005 midl19_52_2_24 However, no quantitative comparisons are provided. NEG 185 graph20_53_2_12 " Finally we saw a high rating for the perception of realism and feelings of immersion in the environment (Q10) ( = 5.88, = 0.78).""" NA 643 iclr20_2094_1_17 Nothing is said about actions and transitions and rewards (we have to read the AC framework in order to get a clue of these components). NEG 1111 midl20_127_4_6 Note: the abstract is not included in the PDF. NEG 1317 neuroai19_36_1_5 Overall the technical aspects of this paper seem sound. POS 619 iclr20_2046_2_24 It is not clear whether such assumptions hold for practical problems. NEG 297 iclr19_1291_3_10 There are some questions in the experiment section that have not been addressed very well. NEG 275 iclr19_1091_1_13 " The second point is the motivation of the split approach: it seems in direct contradiction with the ""disentangled"" and ""compact"" demands the authors pose." NEG 498 iclr19_938_3_3 MAAC outperforms baselines on TC, but not on RT. NA 992 midl19_52_2_11 Even though the authors explain the details in the text I believe an additional illustration in each block (maybe in Appendix) might be helpful to reproduce the method in the paper for further research. NEG 887 midl19_40_3_5 The best combination (both labels + CRF) are close or on par with full supervision. NA 975 midl19_51_2_15 In addition, images representing eliminated nuclei using noisy RCM images should be presented with their counterpart using despeckling network. NEG 598 iclr20_2046_2_3 Pros: This paper presents the first study of tree search for optimal actions in the presence of pretrained value and policy networks. POS 474 iclr19_866_1_7 WEAKNESSES - The algorithmic contribution is relatively minor, while the technical merits of the approach are questionable. NEG 1250 neuroai19_26_1_6 Also perhaps report results from one of the 2 (mentioned) more complex benchmarks. NEG 1386 neuroai19_59_3_11 However, I feel that the work lacked clarity when it came to interpretation of the results. NEG 866 midl19_14_2_42 Imaging 34.9 (2015): 1797-1807. NA 1166 midl20_77_4_13 How big were the ROIs? NA 650 iclr20_2094_1_24 In the algorithm what is n_{gae}? NEG 874 midl19_25_3_4 " No public implementation of the method is provided, which would be a nice extra""" NEG 1145 midl20_71_1_2 It's compared to an earlier method which uses a 3D network and time-point concatenation and reports improvement in Dice scores, false positive rates and true positive rate. NA 943 midl19_51_1_11 One issue, from a purely organizational standpoint, is the fact that information about previous work is either omitted or scattered around the text. NEG 597 iclr20_2046_2_2 Experimental results validate the theoretical analysis and demonstrate the effectiveness of A*MCTS over benchmark MCTS algorithms with value and policy networks. POS 280 iclr19_1091_1_18 It seems the robot arm task is very similar to the navigation task, due to robot arm's end effector being position controlled directly. NEG 1122 midl20_56_4_0 The authors propose a framework to utilize one model under different acquisition context scenarios. NA 1371 neuroai19_54_3_10 Explanations are mostly complete, though some details are missing. POS 680 iclr20_526_3_1 The approach doubles down on the variational approach with variational approximations for both the positive phase and negative phase of the log likelihood objective function. NA 108 graph20_36_1_22 Last, I would like to talk about the results. NA 1114 midl20_135_3_0 This paper proposes a pulmonary nodule malignancy classification based on the temporal evolution of 3D CT scans analyzed by 3D CNNs. NA 736 iclr20_720_2_9 First of all, I should point out that [1] looked at applying Option-Critic in a many task setting and found both that there was an advantage to hierarchy and an advantage to added depth of hierarchy. NA 367 iclr19_261_3_2 They collect a novel dataset in this grounded and goal-driven communication paradigm, define a success metric for the collaborative drawing task, and present models for maximizing that metric. NA 1117 midl20_135_3_3 Specify that it is on the validation set if so, and clarify these points: number of epochs was set to 150, early stopping to 10 epochs Why is this clipping used? NA 564 iclr20_1493_2_23 " In particular, (b) indicates that it may be *necessary* to design regularization methods that steer NNs towards the correct decision boundaryit says nothing about whether these regularization methods will be *sufficient*, which the paper seems to suggest, e.g. in the abstract ""our results suggest that adversarial vulnerability is not an unavoidable consequence of machine learning in high dimensions, and may often be a result of suboptimal training methods used in current practice.""" NEG 172 graph20_45_2_7 The benefits of the visualization are only demonstrated through qualitative results. NEG 511 iclr19_997_3_4 The proposed method is evaluated on object classification and object alignment tasks. NA 451 iclr19_601_3_5 Figure 5 is referenced in the main text after figure 6. NEG 450 iclr19_601_3_4 Some minor issues: Figure 2 is not referenced anywhere in the main text. NEG 755 iclr20_727_1_10 I believe that since the model proposed by the authors allows easy back-propagation, their model ought to be easy and fast to train as well. NA 122 graph20_39_2_4 It read like some form of a haphazard account of few studies that point to the relevance of tracking and visualising patient data in order to inform better health decisions, and ultimately a better lifestyle. NEG 1226 neuroai19_2_2_16 For instance, how is the b at line 63 related to the activation x_i and ReLU at lines 75 and 76? NA 3 graph20_25_2_3 The experiment compared BendyPass with standard PIN security feature on touchscreen devices. NA 824 midl19_13_2_9 Quantitative assessment is fairly limited, and yielding underwhelming results compared to individual networks (ex. CycleGAN). NEG 1153 midl20_77_4_0 This paper evaluates 5 different models for motion tracking in 4D OCT. NA 800 iclr20_880_2_20 The product of a series of randomly initialized matrices can lead to a matrix that is initialized with a different distribution where, eventually, components are not i.i.d.. To show that this is not relevant, the authors should organize an experiment where the original matrix (in the small network) is initialized with the dot product of the composing matrices. NEG 396 iclr19_304_3_9 Detailed remarks: General: A proper definition or at least a somewhat better notion of overfitting would have benefitted the paper. NEG 1071 midl20_100_1_22 Your statement about AUCs and training sizes is either obviously correct or obviously wrong, depending on interpretation. NA 536 iclr20_1042_2_14 This is not a sound mechanism to achieve an as-faithful-as-possible (limited by the expressiveness of the encoder-decoder architectures) approximation to the training data. NEG 892 midl19_40_3_10 Could it be extended to work with only a fraction of the nuclei annotated ? NEG 921 midl19_49_1_17 In the last paragraph of the introduction, authors say 'it is hard to define a feasible metric describing the similarity of the valve shape in general.'. NA 531 iclr20_1042_2_9 I'm assuming the text is correct, but then a beta should be added to the equation in front of the KL divergence. NA 1279 neuroai19_32_1_5 If it had been evaluated and its efficacy varied in an interesting way with respect to the parameters of the model this could be a potentially important model to understand why the nervous system trades off between object identity associated features, transformation features, and speed. NA 308 iclr19_1399_1_1 Strengths: - The experiments are very thorough. POS 830 midl19_14_2_4 Thus, it is not necessary to have all the classes annotated in all the images but to have the labels at least in some of them. NA 1319 neuroai19_36_1_7 However at present, adversarial attacks likely have much larger relevance to AI than neuro. NEG 96 graph20_36_1_10 There is no clue about scalability neither. NEG 97 graph20_36_1_11 To which extent the system supports other patterns? NEG 1053 midl20_100_1_4 The methodological novelty seems insignificant. NEG 1120 midl20_135_3_6 Can you comment? NA 471 iclr19_866_1_4 The paper evaluates the method in various simulated domains and compares against RL and IL baselines. NA 101 graph20_36_1_15 Looking at table 1 makes me think these instructions are quite clear on how to make these 3 patterns. NA 330 iclr19_242_2_5 In this case, I would expect the authors provide more intuitive explanations. NEG 767 iclr20_855_3_3 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. NA 1033 midl19_59_3_1 Only compared to IMM which is very similar to the proposed T-IMM. NEG 99 graph20_36_1_13 This inevitably has an effect on syrup pouring. NA 1094 midl20_108_3_10 pseudo-url] yet the presented method is benchmarked only against 3. NEG 305 iclr19_1333_1_3 A quite severe issue with this report is that the authors don't report relevant learning results from before (+-) 2009, and empirical comparisons are only given w.r.t. other recent heuristics. NEG 594 iclr20_1724_2_11 " As this direction (of increased resolution to make the problem less artificial) is likely to be important, a brief discussion of this finding from the main paper text would be appropriate - p3 resiliance -> resilience - p4 objects is moved -> object is moved - p6 actions itself -> actions themselves; builds upon -> build upon - p7 looses all -> loses all; suited our -> suited to our; render's camera parameters -> render camera parameters; to solve it -> to solve the problem - p8 (Xiong, b;a) and (Xiong, b) -> these references are missing the year; models needs to -> models need to - p9 phenomenon -> phenomena; the the videos -> the videos; these observation -> these observations; of next -> of the next; in real world -> in the real world """ NA 933 midl19_51_1_1 The clinical value of CM images has been highlighted in previous work, but although effective towards the goal of detecting the presence of cancer, these images are hard to interpret by humans. NA 224 graph20_61_2_2 Clarity The presentation is very clear, with pertinent textual and visual explanations. POS 216 graph20_56_1_23 In fact, this reads as if the feedback from the experts was so bad that they did not want to describe it. NA 1339 neuroai19_37_3_17 " For example ... ""A neuron simply sits and listens." NA 838 midl19_14_2_12 " It would be interesting to know that aspect, as it is crucial to allow the network to learn to ""transfer"" its own ability for detecting a new region from one data set to another." NA 1235 neuroai19_23_1_4 It would have been useful to put these in context of the results of the algonauts contest, which pitched supervised methods such as Alexnet against user-submitted content. NA 1231 neuroai19_23_1_0 I believe the concept of using predictive coding and unlabeled video data to train convnets is a great idea. POS 773 iclr20_855_3_9 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. POS 609 iclr20_2046_2_14 The complexity bound in Theorem 1 is hard to understand. NEG 377 iclr19_261_3_14 Depending on those variance numbers you might also consider doing a statistical test to argue that the auxiliary loss function and and RL fine-tuning offer certain improvement over the Scene2seq base model. NEG 1288 neuroai19_32_1_14 It was not clear though where they experimentally varied/tested this prior in their algorithm. NEG 752 iclr20_727_1_7 Also, using the VAE for a predictive task is a little unusual. NEG 517 iclr19_997_3_10 This paper argues that the choice of the number of parameters is sub-optimal and ineffective in terms of computational complexity. NA 669 iclr20_305_3_1 Summary The authors apply MARL to principal-agent / mechanism design problems where selfish agents need to be incentivized to coordinate towards a leader's (collective) goal. NA 1224 neuroai19_2_2_14 The submission is pretty clear. POS 372 iclr19_261_3_8 You might be able to convince me more if you had a stronger baseline e.g. a bag-of-words Drawer model which works off of the average of the word embeddings in a scripted Teller input. NEG 285 iclr19_1091_1_23 How would rotating the measurement frame of the ground-truth influence the results? NEG 488 iclr19_866_1_21 The paper provides insufficient details regarding the RL and IL baselines, making it impossible to judge their merits. NEG