diff --git "a/conferences_annotated/sentence_level/train.csv" "b/conferences_annotated/sentence_level/train.csv" new file mode 100644--- /dev/null +++ "b/conferences_annotated/sentence_level/train.csv" @@ -0,0 +1,981 @@ + sentence_id text position +379 iclr19_261_3_17 You should link to this literature (mostly in NLP) and contrast your task/model with theirs. NEG +976 midl19_51_2_16 5- Obtaining quantitative comparison results for staining accuracy is not feasible due to the reasons clearly defined by the authors. NEG +374 iclr19_261_3_11 Please provide variance measures on your results (within model configuration, across scene examples). NEG +1208 midl20_96_3_19 I am advising regulatory decision makers and do active research in clinical environments. NA +541 iclr20_1042_2_19 Similarly, the proposed rejection sampling scheme of OCDVAE is not consistent with the theory of VAEs and it's a post-hoc tweak that is not theoretically expected to provide a pdf of data with lower KL divergence to the true data pdf. NEG +801 iclr20_880_2_21 The training should be done by using the small network. NEG +683 iclr20_526_3_4 As others have found in the past, a variational approximation to the partition function contribution to the loss function (i.e. the negative phase) results in the loss of the variational lower bound on log likelihood and the connection between the resulting approximation and the log likelihood becomes unclear. NA +982 midl19_52_2_1 The authors test their algorithm on a dataset of 95 subjects for neuromuscular disease. NA +231 graph20_61_2_9 The proposed methodology of design and development relies on well established practices: eliciting requirements through focus groups, designing using action design research framework, implementation through agile development, evaluating the system through uncontrolled longitudinal studies and feedback sessions. POS +147 graph20_39_3_8 While identifying the uniqueness of each patients medical conditions and how/why they record information is important, I think this could be greatly shortened to the most pertinent points to demonstrate the differences. POS +1321 neuroai19_36_1_9 " As well as whether or how adversarial attacks (as framed) might have relevance to neuroscience.""" NEG +167 graph20_45_2_2 This approach preserves the readability of correlational patterns from the original PCP while making cluster assignments more obvious than alternatives relying on edge bundling and on just the use of line color. NA +72 graph20_29_3_40 This would be good to report, either way, even though only a small number of trials was removed overall. NEG +70 graph20_29_3_38 CLARITY Removing tap points that are further than a fixed distance away from the target center will likely affect W levels differently. NA +925 midl19_49_1_21 other comments: - The authors use 2D images to represent leaflet shapes, I'm concerned whether 2D photograph is precise enough. NEG +503 iclr19_938_3_8 MAAC does not consistently outperform baselines, and it is not clear how the stated explanations about the difference in performance apply to other problems. NEG +171 graph20_45_2_6 However, there is one key weakness which prevents me from being more positive with respect to acceptance: an evaluation of the proposed visualization in practical use through a user study is absent. NEG +130 graph20_39_2_12 Overall, the analysis lacks clarity, rigour and situated in the literature. NEG +168 graph20_45_2_3 The implementation of the proposed visualization requires tackling several interesting aspects including a scheme to connect lines between duplicated axes by drawing Hermite spline segments that preserve the line slopes at the axes and a layout optimization based on an A* algorithm to compute the shortest path ordering of duplicated axes. NA +264 iclr19_1091_1_2 The paper is easy to read, and seems technically sound. POS +507 iclr19_997_3_0 Summary This paper proposes an evolutionary-based method for the multi-objective neural architecture search, where the proposed method aims at minimizing two objectives: an error metric and the number of FLOPS. NA +708 iclr20_526_3_29 While I understand the stance taken by the authors that these methods leverage the tractability of the conditional distributions, these strategies are sufficiently general to be considered widely applicable and a true competitor for AdVIL. NEG +1187 midl20_90_2_8 The results are also very nice. POS +272 iclr19_1091_1_10 In the main text, no results are presented that warrant such a conclusion. NEG +799 iclr20_880_2_19 In fact, each composing matrix is initialized randomly. NA +420 iclr19_304_3_34 While for criterion 1 you define overfitting as 'above the diagonal line and underfitting as below the line, which is at least measurable depending on sample density of the randomization, such criteria are missing for C2 and C3. NEG +1102 midl20_119_2_3 Improvement on plaque detection is signification. POS +879 midl19_36_2_4 " Predicting the confidence map with fully convolutional networks was initially done by : ""Microscopy Cell Counting with Fully Convolutional Regression Networks"", W. Xie, J.A." NA +696 iclr20_526_3_17 I note that I am aware of the theoretical representation differences between directed and undirected models, I am wondering how these differences actually matter in practical applications at scale. NEG +1109 midl20_127_4_4 If a sonographer is able to acquire these images, they are also able to perform these measurements. NEG +42 graph20_29_3_10 " Perhaps worse, the paper immediately jumps from this patched-together explanation, straight to calling it a ""novel finding"", and then to suggesting design guidelines from it, as if it was now a proven fact." NEG +56 graph20_29_3_24 I might be wrong. NA +1151 midl20_71_1_8 Overall, the problem the paper tackles is critical, and the proposed network component is effective to some extent. POS +1002 midl19_52_2_21 The reason for high performance of the proposed method can be explained with the required number of parameters to train the method. NA +132 graph20_39_2_14 Lastly, in HCI, there is a movement towards ideas about participatory design, user-centred design, value-sensitive design and so on. NA +1193 midl20_96_3_4 Several experiments are proposed and results are presented. NA +977 midl19_51_2_17 It is necessary to provide more qualitative information regarding the staining results in addition to confirmation from two expert pathologists. NEG +38 graph20_29_3_6 The paper doesn't even acknowledge that this lack of success could simply be due to a lower external validity than the authors hoped for. NEG +306 iclr19_1333_1_4 " This makes it for me not possible to advice publication as is.""" NA +492 iclr19_866_1_25 In fact Mei et al. 2016 requires no human annotation or linguistic knowledge. NA +837 midl19_14_2_11 It is not clear from the explanation in Section 3.1 how the authors deal with the differences in resolution between DRIVE and IDRID data sets. NEG +927 midl19_49_1_23 Though this is not the issue to be considered in this work. NA +270 iclr19_1091_1_8 The discussion of the results reflects this, but the introduction and conclusion suggest otherwise. NEG +408 iclr19_304_3_22 Is that an assumption? NA +691 iclr20_526_3_12 Relevance and Significance: This paper is highly relevant to the ICLR community and -- to the extent that one believes that training and inference in MRFs is important -- also significant. POS +1233 neuroai19_23_1_2 The work is lacking a discussion of the most recent work in the similarity of visual processing in convnets to brain data, which incorporate recurrence into convnets (Nayebi et al. 2018, Kubilius et al. 2018 and 2019), thereby potentially allowing for similar behavior as a PredNet. NEG +166 graph20_45_2_1 The visualization represents clusters of data points in multivariate data by duplicating axes from the canonical PCP visualization to represent 2D subspaces of the multivariate data. NA +713 iclr20_526_3_34 More detail for this application of AdVIL would be nice. NEG +71 graph20_29_3_39 I imagine that more of these errors occurred in the W=10mm condition. NA +556 iclr20_1493_2_14 Interestingly, they also construct a dataset where they Bayes-optimal classifier is robust and neural networks *do* learn a robust classifier (adversarial squares sans label noise). NA +946 midl19_51_1_14 The contribution is therefore incremental, building on top of well-known techniques. NEG +971 midl19_51_2_11 4- The authors conclude that the despeckling NN is crucial to obtain realistic images, however, the results presented in Figures 8 and 9 do not provide enough information to support this conclusion. NEG +961 midl19_51_2_1 The aim for this work is to provide an image that is familiar to the pathologists such that it will remove the need for specific training for CM interpretation. NA +1132 midl20_56_4_14 In Table 3., the result of the proposed method is slightly higher than the CSM. NA +1318 neuroai19_36_1_6 " No trouble understanding the material or writing By focusing on the more biologically plausible ""feedback alignment"" networks, the paper does sit at the intersection of neuro and AI." POS +1098 midl20_108_3_14 " The work also raises some interesting points regarding multi-task training for pathology and with further work could be a good paper.""" POS +603 iclr20_2046_2_8 For example, what kind of additional benefit will it bring when integrating the priority queue into the MCTS algorithms? NEG +547 iclr20_1493_2_4 The contribution of the two datasets (the symmetric and asymetric CelebA) is, in my opinion, an extremely important contribution in studying adversarial robustness and on their own these datasets warrant further study. POS +668 iclr20_2157_3_15 " Is it just smoothing? """ NEG +316 iclr19_1399_1_9 The formalization that the authors proposed is basically the definition of curriculum learning. NEG +1256 neuroai19_26_1_12 " Authors could also add some context by considering related works in the computational neuroscience literature, e.g. Stroud et al. Nature Neurosciencevolume 21, pages 17741783 (2018) and pseudo-url (though the latter is very recent).""" NEG +1199 midl20_96_3_10 The task itself would imply that a deep network classifier is potentially an overkill. NA +4 graph20_25_2_4 The results show that although it took longer for participants to create their passwords with BendyPass, they were able to recall and enter them quicker with BendyPass than with PIN. NA +595 iclr20_2046_2_0 This paper proposes A*MCTS, which combines A* and MCTS with policy and value networks to prioritize the next state to be explored. NA +789 iclr20_880_2_9 Now, if internal matrices have more dimensions of the rank of the original matrix, the product of the internal matrices is exactly the original matrix. NA +748 iclr20_727_1_3 They show by means of extensive experiments on real as well as synthetic data that their approach is able to attain and often surpass state of the art predictive models which rely on parametric modelling of the intensity function. POS +1186 midl20_90_2_7 Contrast normalization yielded the best results for detecting meniscus tears, and layer normalization for detecting the remaining pathologies.The algorithm was explained very well. POS +900 midl19_41_1_0 To investigate whether a conditional mapping can be learned by a generative adversarial network to map CTP inputs to generated MR DWI that more clearly delineates hyperintense regions due to ischemic stroke. NA +642 iclr20_2094_1_16 Namely, the state representation is ambiguous: pseudo-formula is obviously not a boolean variable, but a boolean vector (where each component is associated with an item). NEG +1194 midl20_96_3_5 automatic patient data anonymity and data cleansing are important topics - the results look good with a big but (see below) - this is clearly an application paper, testing well known methods in a new scenario. POS +817 midl19_13_2_2 This is an important advantage for leveraging hundreds of recorded cases without having available segmentations. NA +1045 midl19_59_3_13 Not 100% clear if the IMM method used in the experiments is the method described in section 3.2 (alpha=1/T) ? NEG +1237 neuroai19_23_1_6 For this result to be convincing, I would like to see some reasons why the authors think PredNet is outperforming previous models. NA +898 midl19_40_3_16 Those three papers should be included in the state-of-the-art section: - Constrained convolutional neural networks for weakly supervised segmentation, Pathak et al., ICCV 2015 - DeepCut: Object Segmentation from Bounding Box Annotations using Convolutional Neural Networks, Rajchl et al., TMI, 2016 - Constrained-CNN losses for weakly supervised segmentation, Kervadec et al., MIDL 2018 Since the AJI and object-level Dice are not standard and introduced in other papers, it would be easier to put their formulation back in the paper, so the reader does not have to go look for it. NEG +452 iclr19_601_3_6 " Even if intuitively understandable, all parameters in equations should be explicitly described (e.g., h,w,H,W in eq.1)""" NEG +94 graph20_36_1_8 The system does not seem to follow a particular rationale. NEG +95 graph20_36_1_9 The fact that participants complained about the lack of information about syrup pouring reveals that this is more a trial and error approach than an informed design procedure. NEG +1149 midl20_71_1_6 If it's the latter one, is the convolution done with a 4D filter? NEG +839 midl19_14_2_13 The segmentation architecture does not use batch normalization. NA +534 iclr20_1042_2_12 This is not true in a beta-VAE. NEG +153 graph20_43_1_0 " This paper presents two variations of the standard Fitts' law study, to understand the effect of (1) a situation where targets initially appear with a given size (called the ""visual width"" in the paper) but are revealed to have a larger clickable size revealed once the cursor gets close (called the ""motor width"") or vice versa; and (2) different gaps between targets arranged side-by-side." NA +997 midl19_52_2_16 Did the authors considered to utilize complex valued networks for this task? NEG +585 iclr20_1724_2_2 The construction of the dataset focuses on demonstrating that compositional action classification and long-term temporal reasoning for action understanding and localization in videos are largely unsolved problems, and that frame aggregation-based methods on real video data in prior work datasets, have found relative success not because the tasks are easy but because of dataset bias issues. NA +861 midl19_14_2_35 The abstract should be improved. NEG +58 graph20_29_3_26 " p. 2) That seems quite a stretched ""contribution"", at least in the absence of actual data about how long designers do spend on testing width values today." NEG +463 iclr19_659_2_10 This problem is important for practical usage. NA +562 iclr20_1493_2_21 Discussion/interpretation of the results: - Sufficient vs necessary: While the experimental design and results are both of very high quality, I am slightly confused about the interpretation of the results. POS +61 graph20_29_3_29 These are differences between values that are already expressed in percents. NA +901 midl19_41_1_1 To perform image-to-image translation from multi-modal CT perfusion maps to di usion weighted MR outputs To make use of generated MR data inputs to perform ischemic stroke lesion segmentation. NA +24 graph20_25_2_25 Association for Computing Machinery, New York, NY, USA, 37643774. NA +1384 neuroai19_59_3_9 For instance, it is hard to see differences between the cue periods in the bottom two heatmaps, but differences may appear in some numerical measure of the average discriminability over these regions. NEG +878 midl19_36_2_3 In my view, the proposed methods are not completely novel, I think the authors are suggested to cite them, just name a few. NEG +244 graph20_61_2_22 Figure 4: I would suggest to split the figure into 2 rows (3.5 and 3.6) and annotate columns in black font over white paper background, instead of white font over blue application background: with a low zoom level on my PDF reader, I had first confused these annotations with potential widgets in the application. NEG +25 graph20_26_3_0 Thank you for submitting a revised version of this submission, and addressing concerns raised in the previous round of reviews. NA +222 graph20_61_2_0 This submission reports on the creation of a system to help medical residents and their reviewers to assess their learning using an information visualization dashboard, designed for and with them in a participatory process, deployed in their setting, and evaluated with them through a longitudinal study. NA +795 iclr20_880_2_15 The approximation act as the non-linear layers among linear layers. NA +105 graph20_36_1_19 It would have been a good start for a design rationale. NEG +652 iclr20_2094_1_26 Etc. (f) Even if the aforementioned issues are fixed, it seems that the framework is using many hyper-parameters (\gamma, \beta, \alpha_t, etc.) which are left unspecified. NEG +509 iclr19_997_3_2 In the exploration step, architectures are sampled by using genetic operators such as the crossover and the mutation. NA +131 graph20_39_2_13 With a few grammatical typos, it reads as a thread of different perspective, with little grounding in HCI and related field. NEG +1055 midl20_100_1_6 I suggest you either argue for the novelty or remove the claim from the paper. NEG +565 iclr20_1493_2_24 In fact, if real-world datasets end up being like the asymmetric dataset, then the results of this paper would actually indicate the *opposite* of the above statement. NEG +508 iclr19_997_3_1 The proposed method consists of an exploration step and an exploitation step. NA +1396 neuroai19_59_3_21 The paper in the process reveals some (expected) results about how spiking RNNs behave on a working memory task. POS +75 graph20_35_1_1 The first study explores how users respond to new node ideas suggested by the tool and whether that creates more detailed maps. NA +36 graph20_29_3_4 BLAMING AGE Honestly, I found it quite a weak argument to put the lack of generalization of the approach on age (p. 10). NEG +1230 neuroai19_2_2_20 " And emphasize that this only solves credit assignment for certain types of learning problems (at the moment).""" NEG +57 graph20_29_3_25 " by reducing the time and cost of conducting user studies, our model will let them focus on other important tasks such as visual design and backend system development, which will indirectly contribute to implementing better, novel UIs.""" NA +196 graph20_56_1_3 The approach is interesting and the use cases described demonstrate the technique well. POS +1333 neuroai19_37_3_11 Hardly what I'd call moderate effort. NEG +1332 neuroai19_37_3_10 " It is probable that revolutionary computational systems can be created in this way with only moderate expenditure of resources and effort"" Of course whole fields are working on this problem." NA +1075 midl20_100_1_26 Having said that, if the model predictions does not change, then AUC does not change. NA +477 iclr19_866_1_10 The trajectory encoder operates differently for goal-oriented vs. trajectory-oriented instructions, however it is not clear how a given instruction is identified as being goal- vs. trajectory-oriented. NEG +86 graph20_36_1_0 This paper presents a projection system to help unexperienced people to draw latte art on a cappuccino. NA +334 iclr19_242_2_10 The improvement on test errors does not look significant. NEG +899 midl19_40_3_17 " Replacing (a), (b), ... by Image, ground truth, ... in figures 2, 3, and 4 would improve readability. """ NA +1366 neuroai19_54_3_5 1) If I understand correctly, attribution is computed only for a single OSR stimulus video. NEG +781 iclr20_880_2_1 In fact, the major claim is that using a cascade of linear layers instead of a single layer can lead to better performance in deep neural networks. NA +1082 midl20_100_1_33 Finally, I would very much have liked to to see a frame from one of the videos. NEG +1087 midl20_108_3_3 The authors evaluated the quality of these representations on multiple tasks, illustrating the added benefit of their multi-task system and the utility of using multiple tasks to supervised the feature extraction. NA +538 iclr20_1042_2_16 I.e., there are two different probabilistic models modeling the same data in inconsistent ways and one or the other is used depending on the part of the system. NEG +666 iclr20_2157_3_13 I am also not clear on where the image attribution prior comes from for the image task. NEG +418 iclr19_304_3_32 My main concern here, besides the motivations that I did not fully understand (s.b. NEG +1125 midl20_56_4_3 The idea of learning convolution weights for different input image quality is novel. POS +1306 neuroai19_34_2_8 Generally, great paper. POS +1340 neuroai19_37_3_18 When it hears an incoming pattern of spikes that matches a pattern it knows, it responds with a spike of its own. NA +558 iclr20_1493_2_16 Excessive Invariance causes Adversarial Vulnerability (pseudo-url): Jacobsen et al offers an explanation for adversarial examples based on the fact that NNs are not sensitive to many task-relevant changes in inputs, which seems to tie in nicely to the discussion in this paper, as under the presented setup the Bayes-optimal classifier will certainly exploit (and be somewhat sensitive) to such changes. NA +1183 midl20_90_2_4 The method was tested on two different datasets, which is impressive. POS +1010 midl19_52_2_29 " c- Quantitative results can be mentioned in the abstract. """ NEG +717 iclr20_57_3_1 To solve this problem, the authors first applied distant supervision technique to harvest hard-negative training examples and then transform the original task to a multi-task learning problem by splitting the original labels to positive, hard-negative, and easy-negative examples. NA +1197 midl20_96_3_8 Why hasn't the semi-supervised paradigm be explored in more detail instead of only using a few biasing iterations with user input? NEG +1024 midl19_56_3_13 Their voxel resolution is only sligthly smaller than in this work (120x120x40), with a similar latent dimensionality (64D, here: 3*29=87). NA +1017 midl19_56_3_6 Authors could comment on how their model could be incorporated into (e.g. deep) segmentation approaches, because I do not see an immediate way to do that without requiring the (precise) image-based localization of mandible landmarks in a test volume. NA +1127 midl20_56_4_7 It conducts extensive experiments for three different settings and the results demonstrate the effectiveness of the proposed method.1). POS +1095 midl20_108_3_11 It would be helpful to put the results in context with all other methods such as automatic and semi-automatic methods. NEG +8 graph20_25_2_8 The paper is well written: the work is motivated well, the related work is mostly comprehensive, and the design and evaluation sections are clear and have enough detail for others to attempt to reproduce/replicate the study. POS +802 iclr20_880_2_22 If results are significantly different, then the authors can reject the hypothesis. NA +1200 midl20_96_3_11 Bluntly: surgical parts are predominantly red, non-surgical parts anything and blue/green. NA +1228 neuroai19_2_2_18 Seeing if these meta-learnt rules line up with previously characterized biological learning rules is particularly interesting. NEG +1085 midl20_108_3_1 The authors extended unsupervised NIC to a multi-task supervised system. NA +1054 midl20_100_1_5 Plenty of works combine autoencoders with LSTMs. NEG +74 graph20_35_1_0 This paper presents QCue, a tool to assist mind-mapping through suggested context related to existing nodes and through question that expand on less developed branches, including two studies, a detailed description of the algorithm design, and rater evaluation of their results. NA +739 iclr20_720_2_12 In comparison to past frameworks, the approach of this paper seems less theoretically motivated. NEG +522 iclr20_1042_2_0 This paper tackles the problem of catastrophic forgetting when data is organized in a large number of batches of data (tasks) that are sequentially made available. NA +335 iclr19_242_2_11 If given more computing resources, and under same timing constraint, we have many other methods to improve performance. NA +88 graph20_36_1_2 The results suggest that participants perform better with the system. NA +858 midl19_14_2_32 I would certainly accept the paper is this experiment were included and the results were convincing. NA +217 graph20_56_1_24 Why dont they include the feedback? NEG +886 midl19_40_3_4 The trained network is then fine tuned with a direct CRF loss, as in Tang et al. Evaluation is performed on two datasets in several configurations (with and without CRF loss, and variation on the labels used) ; showing the effects of the different parts of the method. NA +173 graph20_45_2_8 " The paper would have been significantly stronger if the expected benefits were measured in a practical scenario.""" NEG +926 midl19_49_1_22 3D scanner such as CT, MRI, optical scanner could be more suitable for this work? NEG +905 midl19_49_1_0 This paper presents a clustering method using deep autoencoder for aortic value shape clustering. NA +319 iclr19_1399_1_12 While these results are scientifically interesting, I don't expect it to be of practical use. POS +1216 neuroai19_2_2_6 Section 1 pitches the method as solving the credit assignment problem, citing problems with weight symmetry etc, that apply to many forms of learning. NA +693 iclr20_526_3_14 In most modeling situations, one would simply impose the directed graphical model directly and skip the formalization in terms of an MRF. NA +266 iclr19_1091_1_4 The contribution is minor, and the reasoning behind it could be better motivated. NEG +6 graph20_25_2_6 The main strength of the paper is the experimental user study design with users who are visually impaired. POS +138 graph20_39_2_20 " I would encourage the authors to situate the research questions into the broader literature and determine whether they fit into some of the well-established methods informing the designing of health-related technologies. """ NEG +163 graph20_43_1_10 Finally, I found the study results to be difficult to interpret, as many of the results subsections are ANOVA output with little interpretation and commentary to help the reader understand what was found. NEG +1148 midl20_71_1_5 2) what is the dimension of input, is it W D or H W D$ ? NEG +1232 neuroai19_23_1_1 However, the contribution of the authors does not appear to extend beyond combining existing data sets with existing network architectures. NEG +1013 midl19_56_3_2 The paper is written clearly. POS +1376 neuroai19_59_3_1 The importance is tempered by the findings only covering what is to be expected, and not pushing beyond this or describing a path to push beyond this. NEG +523 iclr20_1042_2_1 To avoid catastrophic forgetting, the authors learn a VAE that generates the training data (both inputs and labels) and retrain it using samples from the new task combined with samples generated from the VAE trained in the previous tasks (generative replay). NA +737 iclr20_720_2_10 Additionally, it is well known that Option-Critic approaches (when unregularized) tend to learn options that terminate every step [2]. NA +704 iclr20_526_3_25 The comparison to PCD-1 in Fig. 3 seems a bit unfair in that the learning curve ends at 8000 iterations, while PCD-1 continues to improve NLL. NEG +578 iclr20_1493_2_38 While completely alleviating this concern may once again be quite difficult/impossible, it could be significantly alleviated by generating training samples dynamically (at every iteration) instead of generating a dataset in one shot and training on it. NEG +665 iclr20_2157_3_12 So with that the paper positions itself not as a survey but as a method paper but lacks evidence that the method expected gradients performs better. NEG +19 graph20_25_2_19 In summary, this is an interesting paper that will contribute to the GI community. NA +779 iclr20_855_3_15 " 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. """ NEG +777 iclr20_855_3_13 The paper is written as if the conclusions could be extended to model-based methods in general. NA +673 iclr20_305_3_5 With this modeling step, the authors formulate an event-based policy gradient, which considers models for which goal to send to followers and when. NA +809 iclr20_934_1_6 The proposed method is very similar with the unsupervised GraphSAGE, which also optimizes Eq.(7). NEG +137 graph20_39_2_19 This makes the paper weak, lacking impactful significance, and thus leaning would not argue strongly towards acceptance. NEG +557 iclr20_1493_2_15 While I think the datasets presented in this work are much more interesting and certainly more realistic, this work should be put in context. POS +1142 midl20_70_4_4 " I only regret the fact that this is a short paper, and there is therefore not enough space for a more formal description and discussion of the methodology.""" NEG +702 iclr20_526_3_23 I am somewhat alarmed at the use of 100 updates of the joint model q(v,h) (K1 = 100) for every update of the other parameters. NEG +744 iclr20_720_2_20 " 2] ""When Waiting is not an Option: Learning Options with a Deliberation Cost"" Jean Harb, Pierre-Luc Bacon, Martin Klissarov, and Doina Precup." NA +782 iclr20_880_2_2 As the title reports, expanding layers seems to be the key to obtain extremely interesting results. NA +368 iclr19_261_3_3 This is a very interesting task and the dataset/models are a very useful contribution to the community. POS +1315 neuroai19_36_1_3 Adversarial attacks are artificial: attacker has access to gradient of the loss function. NA +672 iclr20_305_3_4 A `termination' menas that an agent should stop executing the previous selected action; the leader signals as such to the agent. NA +327 iclr19_242_2_1 The authors claim the proposed method has better generalization performance. NA +317 iclr19_1399_1_10 There is no novelty about this. NEG +1063 midl20_100_1_14 " If you do use it, you cannot argue that you learn from ""a small number of labeled samples"" as done in the final paragraph of the paper." NEG +1031 midl19_56_3_20 1] Oktay O, Ferrante E, Kamnitsas K, Heinrich M, Bai W, Caballero J, et al. Anatomically Constrained Neural Networks (ACNNs): Application to Cardiac Image Enhancement and Segmentation. NA +528 iclr20_1042_2_6 A normal flow is to first describe the model and what the involved variables mean, and then talk about what the loss for learning it should be, not the other way around. NA +526 iclr20_1042_2_4 Unfortunately, there are several things that left me unconvinced about this paper: 1) Presentation of the paper - Variables x, y, z are introduced and talked about without explanation. NEG +87 graph20_36_1_1 There is a user study comparing participants performance with the system, and with watching explanatory videos only. NA +479 iclr19_866_1_12 A contrastive loss would seemingly be more appropriate for learning the instruction-goal distance function. NEG +338 iclr19_242_2_14 The experiments are not strong. NEG +658 iclr20_2157_3_5 I think a few papers to have a look at are a survey article about graph based biasing pseudo-url as well as methods for using graph convolutions with biases based on graphs: pseudo-url and pseudo-url . NA +148 graph20_39_3_9 I would have also liked to see some of the images of the visualizations for myself. NEG +1189 midl20_96_3_0 The presented paper aims to label and remove irrelevant sequences from laparoscopic videos. NA +1289 neuroai19_32_1_15 So while an interesting connection they did not make clear where they substantively pursue it. NEG +1015 midl19_56_3_4 There are certain original aspects in this work (latent en-/decoding, inception-based decoder network, latent space interpolation, generalization to previously unseen shapes etc.), but the work may not be as original as authors suggest, since they may not be aware of a very similar work (see Cons), where some of the discussed concepts have already been proposed and explored. NEG +246 graph20_61_2_24 " Congratulations for opensourcing the code to potentially help other institutions with medical programs (""across Canada"", or beyond?)." NA +365 iclr19_261_3_0 This paper presents CoDraw, a grounded and goal-driven dialogue environment for collaborative drawing. NA +660 iclr20_2157_3_7 It is not clear which model is used in Figure 2. NEG +1051 midl20_100_1_2 The method is compared to five embryologists and results clearly shows that learning directly from the clinical outcome outperfoms embryologists by a large margin. POS +198 graph20_56_1_5 The basics of the technique are well-described: the user draws a shape that the system then selects matches for, based on two similarity metrics (one calculated by Pearson's coefficient and the other by a PCA algorithm). POS +1080 midl20_100_1_31 A mior nitpick: You define all abbreviations except for UBar. NEG +797 iclr20_880_2_17 If this does not lead to the same improvement, there should be a value in the expansion. NEG +404 iclr19_304_3_17 " What do you mean by ""easier to learn""?" NA +688 iclr20_526_3_9 Specifically, it states that the generator is minimizing a Jenson-Shannon divergence which has a fixed point at the true data density. NA +483 iclr19_866_1_16 Where do they come from? NEG +158 graph20_43_1_5 While I appreciate the overall motivation, I'm not sure if a Fitts' law study is the right approach for going about understanding the effects of these kinds of interfaces. NEG +315 iclr19_1399_1_8 It is difficult for me to accept it. NA +256 iclr19_1049_1_1 The method does not make major changes to the network structure, but by modifying the calculations in the network. NA +1014 midl19_56_3_3 Methods, materials and validation are of a sufficient quality. POS +1116 midl20_135_3_2 Some points to address are listed in the following: The early stopping is not clear. NEG +384 iclr19_261_3_25 Speaker-follower models for vision-and-language navigation. NA +177 graph20_53_2_4 " The system requires that the virtual objects are implemented in a way that they do not only present an outside facade but also contain primitives of its components not displayed on the outside (i.e., ""internal faces"")." NA +655 iclr20_2157_3_1 The structure of the paper is strange because it discusses attribution priors but then they are not used for the method. NEG +314 iclr19_1399_1_7 The authors claim the formalization of the problem to be one of their contributions. NA +357 iclr19_242_2_35 Like the authors said, they did not propose new data augmentation method, and their contribution is how to combine data augmentation with large-batch training. NA +298 iclr19_1291_3_11 For example: Is there any difference between the results of table 1, if we look at the cooperative setup? NEG +455 iclr19_659_2_2 Experimental results demonstrate that the proposed method can achieve better performance than non-ensemble one under the same training steps, and the decision space can also be stabilized. NA +543 iclr20_1493_2_0 This paper proposes studying adversarial examples from the perspective of Bayes-optimal classifiers. NA +1026 midl19_56_3_15 Compared to the proposed work, where latents represent clinically relevant mandible landmarks, an auto-encoder approach as in ACNN is more general: relevant landmarks as in the mandible cannot be identified for arbitrary anatomies, and a separate training of decoder and decoder as proposed here crucially depends on a semantically meaningful latent space with a supervised mapping to the dense representation (e.g. hand-labeled landmarks vs. voxel labelmaps). NEG +732 iclr20_720_2_5 Additionally, it is very much not clear why someone, for example, would select the approach of this paper in comparison to popular paradigms like Option-Critic and Feudal Networks. NEG +527 iclr20_1042_2_5 The graphical model or factorization assumptions are not even mentioned until after the loss has been defined. NEG +1167 midl20_77_4_14 " Section 3: combing should be combining """ NEG +159 graph20_43_1_6 Or, put in a different way, I'm not sure if the study results are all that valuable for designers (given that it's looking at 1D pointing), or whether this type of interface is common enough that it's useful to have a new Fitts' law formula to account for it. NEG +1104 midl20_119_2_5 " This will provide more insights or explanations.""" NA +360 iclr19_242_2_38 However, the authors quote a previous paper that use different data augmentation and (potentially) other experimental settings. NA +1229 neuroai19_2_2_19 Define the model more explicitly. NEG +202 graph20_56_1_9 I had to re read the paper back and forward to finally tease out what I think is the way it works. NEG +963 midl19_51_2_3 This will potentially bring us closer to rapid evaluation of lesions during surgical operation using fast CM. POS +674 iclr20_305_3_6 The authors compare this approach on 4 environments with M3RL, which also solves (extensions of) principal-agent problems. NA +112 graph20_36_1_26 This might have affected the metric, with no real impact on the perceived result. NA +771 iclr20_855_3_7 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. NA +307 iclr19_1399_1_0 In my opinion this paper is generally of good quality and clarity, modest originality and significance. POS +731 iclr20_720_2_4 The two improvements in section 3.2 seem quite low level and are only applicable to this particular approach to hierarchical RL. NEG +422 iclr19_304_3_36 You present a number for C2 in Section 5, but that is only applicable to the present data set (i.e. assuming that training accuracy is 1). NEG +684 iclr20_526_3_5 To deal with this issue, the authors argue (in Lemma 1) that the gradient of their approximate objective is at least in the same direction as the ELBO (lower bound) objective. NA +1034 midl19_59_3_2 Comparison to (unsupervised) domain adaptation methods would also have been interesting (e.g. gradient reversal (Ganin et al. 2014, Kamnitsas et al. 2016)). NEG +816 midl19_13_2_1 The method introduces a self attention mechanism using weakly supervised labels, thereby avoiding the need to use more exhaustive annotations such as segmentations. NA +459 iclr19_659_2_6 This idea is simple and works well. POS +1322 neuroai19_37_3_0 The paper provides a broadly useful synthesis of key differences between ANN and SNN approaches. POS +1261 neuroai19_29_1_4 The authors should have identified a task where networks trained on MNIST perform poorly, and then propose a different strategy or architecture. NEG +1079 midl20_100_1_30 If there is an issue with Tran et al you should state it clearly, if not, you should accept their results. NEG +928 midl19_49_1_24 The paper is not well organized. NEG +449 iclr19_601_3_3 Experiments are convincing. POS +841 midl19_14_2_15 The vessel segmentation performance is evaluated on the DRIVE data set. NA +80 graph20_35_1_6 The two studies are well-described and designed studies. POS +1363 neuroai19_54_3_2 3) Suggest testable hypotheses. NA +1118 midl20_135_3_4 It is not clear whether T1 and T2 is available for all cases (mostly) In Table 1, bold results are not always the best, this is very misleading. NEG +215 graph20_56_1_22 Having reviewed this approach with experts, the authors state that the experts did not get it, and so they choose to describe the system with a use-case method. NA +1381 neuroai19_59_3_6 The statistical tools are fairly well described and appear to be well-suited for illustrating the phenomena of interest. POS +30 graph20_26_3_5 However, I noted that there are several typos throughout the text, and I recommend a thorough editing pass for the camera ready. NEG +242 graph20_61_2_20 The choice for visualizing rotation schedules using an interval chart rather than a more space-consuming Gantt chart widespread in time/project management is smart. POS +1278 neuroai19_32_1_4 In the spirit of insight it would have been very nice to have a quantification of error with respect to parameters (priors on slow identity, fast form). NEG +641 iclr20_2094_1_15 d) The problem formulation is very unclear. NEG +209 graph20_56_1_16 The nice video provided was helpful in showing this technique. POS +720 iclr20_57_3_4 This implementation showed improvement of performance on both tasks. NA +435 iclr19_495_1_7 Also, in the experiments, it is said that one can combing normalizing flows with TRPO without describing the details. NEG +936 midl19_51_1_4 The general organization of the paper is sound This paper tackles a problem that is relevant to the whole medical community. POS +1308 neuroai19_34_2_10 Would have been great to include another Imagenet-trained architecture, since different architectures have widely varying macaque brain predictivity, and that of VGG16 is not particularly high (Schrimpf et al., 2018 BrainScore). NEG +1020 midl19_56_3_9 Further, there is always the chance that authors are not aware of every piece of related literature (in all of computer graphics), as it might be the case here. NEG +1273 neuroai19_3_3_8 " Also, it would be very interesting to use these models to predict situations that might trigger maladaptive behaviors, by finding scenarios in which the pathological behavior becomes optimal. """ NA +551 iclr20_1493_2_8 I outline these below. NA +544 iclr20_1493_2_1 They construct a pair of synthetic but somewhat realistic datasetsin one case, the Bayes-optimal classifier is *not* robust, demonstrating that the Bayes-optimal classifier may not be robust for real-world datasets. NA +918 midl19_49_1_14 The experiments measure the recon accuracy. NA +906 midl19_49_1_1 It is the first work to identify aortic value prosthesis types using a general representation learning technique. NA +746 iclr20_727_1_1 Instead of learning the conditional intensity for the point process, as is usually the case, the authors instead propose an elegant method based on Normalizing Flows to directly learn the probability distribution of the next time step. POS +403 iclr19_304_3_16 As in that case correlation in the data can be destroyed by the introduction of randomness making the data easier to learn. NA +1380 neuroai19_59_3_5 A comparison with Bellec et al. 2018, which looks at working memory tasks in spiking networks, would also have been appropriate. NEG +362 iclr19_242_2_40 Moreover, instead of showing the consistent benefits of large batch, the authors tune the batchsize as a hyperparameter for different experiments. NEG +271 iclr19_1091_1_9 The same problem also occurs for the conclusion about the robustness of SRL approaches. NEG +466 iclr19_659_2_13 I think more examples, such as in section 8.1, should be put in the main text. NEG +1245 neuroai19_26_1_1 The heavy lifting is seemingly done by well known architectures: default RNN & a feed-forward NN. NA +1027 midl19_56_3_16 In contrast, ACNN auto-encoders train their encoder and decoder in conjunction. NA +277 iclr19_1091_1_15 Due to the shared feature extractor, the contradictory objectives (and hence the need for tuning of the weights in the cost function) are still a potential problem. NEG +1343 neuroai19_37_3_22 " We require a new class of theories that dispose of the simplistic stimulus-driven encode/ transmit/decode doctrine. """ NA +1326 neuroai19_37_3_4 " The paper opens ""In recent years we have made significant progress identifying computational principles that underlie neural function." NA +829 midl19_14_2_3 The adversarial loss allows to leverage complementary data sets that do not have all the regions of interest segmented. NA +1067 midl20_100_1_18 You argue that including embryologists decisions in the prediction is an easier task. NA +1074 midl20_100_1_25 This holds for all the popular performance measures. NEG +442 iclr19_495_1_14 I wonder how good the results are if these more advanced versions are used. NA +299 iclr19_1291_3_12 Does their model outperform a model which has global communication with IR? NEG +954 midl19_51_1_22 Is there some reference for multiplicative residual connections? NEG +375 iclr19_261_3_12 Are the machinemachine pairs consistently performing well together? NEG +881 midl19_36_2_6 " The proposed localisation map is actually the result of distance transform, and has been initially used in : ""Counting in The Wild"", C. Arteta, V. Lempitsky, A. Zisserman, In ECCV 2016. """ NA +1044 midl19_59_3_12 " How is training till ""convergence"" (section 4.3) defined?" NA +1058 midl20_100_1_9 If you want your work applied in clinics, this is much more important than improving the results. NEG +851 midl19_14_2_25 The results for vessel segmentation in IDRID images do not look as accurate as those in the DRIVE data set. NEG +472 iclr19_866_1_5 STRENGTHS + Decoupling instruction-to-action mapping by introducing goals as a learned intermediate representation has advantages, particularly for goal-directed instructions. POS +1181 midl20_90_2_2 The normalization was followed by a BlurPool layer to solve the shift variance. NA +563 iclr20_1493_2_22 First, if my understanding of the paper is correct, the experiments show that (a) the Bayes-optimal classifier can be non-robust in real-world settings, and (b) even when the Bayes-optimal classifier is robust, NNs can learn a non-robust decision boundary. NA +1266 neuroai19_3_3_1 This model allows for more flexibility in modelling human behaviors in normal and pathological states. POS +388 iclr19_304_3_1 Instead of using a hold-out set they propose to randomly flip the labels of certain amounts of training data and inspect the corresponding 'accuracy vs. randomization curves. NA +22 graph20_25_2_23 Thumprint: Socially-Inclusive Local Group Authentication Through Shared Secret Knocks. NA +715 iclr20_526_3_36 " Given the comparison to PCD in the RBM setting, I am somewhat surprised that AdVIL is so competitive with VCD in the case of the DBM. """ POS +124 graph20_39_2_6 Second, I particularly appreciate the authors' use of different methods (focus group, interviews, and observation) but fail to see an understanding of the needed sensitivity towards participants with some form of a chronic condition. NEG +423 iclr19_304_3_37 Criterion 2 (b) is not clear. NEG +1253 neuroai19_26_1_9 So it fits well with the workshop theme. POS +1348 neuroai19_37_3_27 There was an absence of nuance. NEG +957 midl19_51_1_25 Is the math right? NEG +1397 neuroai19_59_3_22 The proof-of-concept work (among others) that this can be done with spiking RNN may inspire more work in this area. POS +142 graph20_39_3_3 I enjoyed the paper. POS +50 graph20_29_3_18 In the example given in p. 1 (choosing between 5 or 7-mm circular icons), it is unclear why the designer would need a model, or to know by how much a 7-mm icon would improve accuracy. NEG +1303 neuroai19_34_2_5 Figures exceptionally detailed and thoroughly labelled. POS +980 midl19_51_2_20 " This could potentially add a bias to the results presented here. """ NA +623 iclr20_2046_2_28 How would this affect the results? NA +1011 midl19_56_3_0 Summary: Authors present AnatomyGen, a CNN-based approach for mapping from low-dimensional anatomical landmark coordinates to a dense voxel representation and back, via separately trained decoder and encoder networks. NA +81 graph20_35_1_7 The level of detail in the algorithm description is a particular strength, giving a clear picture of how it works and why those choices were made. POS +703 iclr20_526_3_24 For larger scale domains, I fear this could become an important obstacle to effective model training. NEG +482 iclr19_866_1_15 How many are there? NEG +321 iclr19_1399_1_14 I realize that they were obtained with a simple network, however, showing improvements in this regime is not that convincing. NEG +1163 midl20_77_4_10 Maybe get rid of performing motions? NA +956 midl19_51_1_24 Can we prove that at least visually? NEG +1329 neuroai19_37_3_7 Arguably ACh and noradrenaline are more important for network states and dynamics, and equally important for plasticity as dopamine. NA +66 graph20_29_3_34 the error rate difference was |29 38| = 9%. NA +798 iclr20_880_2_18 3) the small improvement of the expanded network can be given by the different initialization. NA +1234 neuroai19_23_1_3 How would you expect those networks to perform when trained on unlabeled video data? NA +990 midl19_52_2_9 How does the temporal and spatial blocks work? NEG +1023 midl19_56_3_12 CNN-based shape modeling and latent space discovery and was realized for heart ventricle shapes with an auto-encoder, and integrated into Anatomically Constrained Neural Networks (ACNNs) [1]. NA +417 iclr19_304_3_31 But you state it as if those measures are actually correct, which you didnt show yet. NEG +632 iclr20_2094_1_6 There are various classes of BPPs, and it would be relevant to briefly present them. NA +1334 neuroai19_37_3_12 Claims of efficiency of more brain-like approaches compared to AI are disingenuous. NEG +64 graph20_29_3_32 However, that still makes a 10% prediction error quite high in my book, and worthy of contextualization. NA +139 graph20_39_3_0 This paper describes the exploration of designing data visualizations of daily medical records by patients, and what kinds of visualizations may assist providers in best keeping track with an patients medical status. NA +1123 midl20_56_4_1 A novel dynamic weight prediction model is proposed to learn to predict the kernel weights for each convolution based on different context settings. NA +968 midl19_51_2_8 2- It is not clear why the histology images were used for denoising network training. NEG +491 iclr19_866_1_24 The paper incorrectly references Mei et al. 2016 when stating that methods require a large amount of human supervision (data annotation) and/or linguistic knowledge. NEG +939 midl19_51_1_7 The quantitative results delivered by the de-speckling images, which seem to be computed using simulated realization of random speckle noise, look also convincing. POS +974 midl19_51_2_14 For instance, Figure 9 needs to use the same images presented in Figure 8 to provide enough support for the need of despeckling network. NEG +532 iclr20_1042_2_10 Tables and figures are inconveniently far from where they are referenced in the text. NEG +65 graph20_29_3_33 Perhaps I misunderstood something. NA +1351 neuroai19_53_1_1 This sheds new light on how artificial network algorithms might be implementable by the brain. POS +783 iclr20_880_2_3 Moreover, the proposed approach is extremely simple and it is well explained in Section 2 with equations (1) and (2). POS +1170 midl20_85_3_2 From table 1, it is clear that ECE is much lower for the proposed method. POS +296 iclr19_1291_3_9 Authors provide 3 baselines: 1) no communication, but IR 2) no communication, no IR 3) global communication, no IR (commNet) I think having a baseline that has global communication with IR can show the effect of selective communication better. NA +591 iclr20_1724_2_8 It is a well-argued, thoughtful dataset contribution that sets up a reasonable video understanding dataset. POS +419 iclr19_304_3_33 is the lack of measurable criteria. NEG +438 iclr19_495_1_10 BTW, in the Section 4.3, what does [-1, 1]^2 mean? NEG +261 iclr19_1049_1_6 " Minor Example 2: ""A"" -> ""AI"".""" NEG +724 iclr20_57_3_8 The objective function L3 is not well justified. NEG +1139 midl20_70_4_1 The novelty of the proposed framework is to take the label structure into account and to learn label dependencies, based on the idea of conditional learning in (Chen et al., 2019) and the lung disease hierarchy of the CheXpert dataset (Irvin and al., 2019). POS +281 iclr19_1091_1_19 Why is it worthwhile to study this task separately? NEG +626 iclr20_2094_1_0 This paper aims at solving geometric bin packing (2D or 3D) problems using a deep reinforcement learning framework. NA +876 midl19_36_2_1 The authors consider the problem of nuclei detection, and propose to decompose the task into three subtasks, trying to predict the confidence map, localization map and a weight map. NA +935 midl19_51_1_3 They present an architecture making use of two network, a de-noise/de-speckle network (trained independently on one of the two types of CM images used in this work) followed by a generative network (cycle gan). NA +1354 neuroai19_53_1_4 Given its technical details it was reasonably straightforward to follow. POS +188 graph20_53_2_15 I strongly recommend that this be moved to a subsection of the previous section, i.e., the Results section. NEG +1121 midl20_135_3_7 " obtained an F1-score of 0.68 -> 0.686? """ NA +251 graph20_61_2_29 2016. doi: 10.1109/TVCG.2015.2467613 - Papers from the IEEE VIS'16 Workshop: Logging Interactive Visualizations & Visualizing Interaction Logs pseudo-url DESIGN CHOICES AND INSIGHTS GAINED I found the design considerations to be mostly obvious and known to designers and developers of user interfaces and information visualization. NA +583 iclr20_1724_2_0 The paper introduces CATER: a synthetically generated dataset for video understanding tasks. NA +903 midl19_41_1_3 The authors had better compare segmentation result between CTP with orginal MRI and CTP with CGAN MRI. NEG +967 midl19_51_2_7 The authors should validate their selection of two step approach (NN + filter) compared to an end-to-end FCN (with an additional loss like TV) for the despeckling network. NEG +1180 midl20_90_2_1 The main contribution of the work was adding a normalization step to the network, and learning the affine transformation parameters during the training. NA +514 iclr19_997_3_7 Cons - The contribution of the proposed method is not clear to me. NEG +820 midl19_13_2_5 The different loss functions are all based on previously proposed approaches and exploited in this case for this dual background/foreground problem. NEG +706 iclr20_526_3_27 Perhaps PCD-1 results in performance that is far better than AdVIL. NEG +487 iclr19_866_1_20 It would be better to evaluate on one of the few common benchmarks for robot language understanding, e.g., the SAIL corpus, which considers trajectory-oriented instructions. NEG +663 iclr20_2157_3_10 " It is not clear if the paper is presenting ""expected gradients"" or existing attribution priors." NEG +273 iclr19_1091_1_11 The appendix includes some tests in this direction, but conclusions should not be based on material that is only available in the appendix. NEG +685 iclr20_526_3_6 The result is fairly obvious, but the conditions for validity have interesting consequences for the training algorithm, as it relates the approximation error to the norm of the gradient of the ELBO loss. POS +1159 midl20_77_4_6 Was the setup the same as in Gessert et al (2019), i.e. with a robot moving the object and mirrors moving the OCT FOV? NA +1281 neuroai19_32_1_7 Seemed broad and was unsupported by any citations and to my knowledge GANs and VAEs have been used specifically to find interpretable features. NA +448 iclr19_601_3_2 The paper is fairly well written and structured, and it seems technically sound. POS +411 iclr19_304_3_25 Page 4, Monotony. NA +458 iclr19_659_2_5 Using ensembles of Q-function can naturally reduce the variance of decisions, so it can speed up the training procedure for certain tasks. NA +1276 neuroai19_32_1_2 It is difficult to judge whether the new model is important because it has not been evaluated except by eye it does seem to reconstruct an image. NEG +1218 neuroai19_2_2_8 The efficiency of backprop should be mentioned in the intro if it is something this work is aiming to address. NEG +667 iclr20_2157_3_14 Where is this extra information? NEG +1307 neuroai19_34_2_9 Clear presentation of thorough work, exploring an important question. POS +149 graph20_39_3_10 Another concern I have is about the disparity between the emphasis on how each patients medical history (and in turn, visualization) is unique, and then the proposal of general design guidelines for creating patient visualizations. NEG +1353 neuroai19_53_1_3 It would have been nice to present a figure showing how e-prop yields eligibility traces resembling STDP, as this is one of the key connections of this work to biology. NEG +955 midl19_51_1_23 How do we know that the network is learning 1/F (inverse of speckle noise)? NEG +733 iclr20_720_2_6 " The authors mention that Feudal approaches ""employ different rewards for different levels of the hierarchy rather than optimizing a single objective for the entire model as we do.""" NA +218 graph20_56_1_25 Surely they found out useful information. NA +239 graph20_61_2_17 Data characterization is assorted with visibly clear understanding and explanation of the domain. POS +1293 neuroai19_32_1_19 They have some qualitative evaluation in images of filters but they could explore the parameter space to understand what led to these features. NA +725 iclr20_57_3_9 It would be important to see if the proposed method is also beneficial with the state of the art neural networks on the two applications. NEG +192 graph20_53_2_19 " General minor issues: - ""users authoring process"" -> ""users' authoring process""""" NA +951 midl19_51_1_19 I feel it would have been extremely interesting to evaluate the performance of those same clinicians (and others) diagnosing cancer using both H&E stained image and CM images of the same patient (or patient distributions) vs a control group. NA +191 graph20_53_2_18 Despite these weaknesses with regards to the study reporting and discussion, the paper is interesting and showcases good and novel work and I think the GI community would benefit from its presentation (albeit with some changes as suggested above). POS +1202 midl20_96_3_13 Do we really need a labelled ground truth here? NEG +828 midl19_14_2_2 This is important when processing these images, where anatomical and pathological structures usually share similar visual properties and lead to false positive detections (e.g. red lesions and vessels, or bright lesions and the optic disc). NA +832 midl19_14_2_6 The strategy proposed to tackle this issue is not novel as adversarial losses have been used before for image segmentations. NEG +644 iclr20_2094_1_18 We dont know if it is an episodic MDP (which is usually the case in DRL approaches to combinatorial optimization tasks). NA +984 midl19_52_2_3 Moreover, they have done some ablation studies to show the importance of the receptive field and temporal frames for MRF reconstruction. NA +1316 neuroai19_36_1_4 For FA networks, it's unclear why an attacker could not access true gradient, and be forced to use the approximate gradient. NEG +1356 neuroai19_53_1_6 Gives important new results about how eligibility traces can be used to approximate gradients when adequately combined with a learning signal. POS +16 graph20_25_2_16 This brings up another issue: the PIN baseline is the current de facto standard, but other baselines (e.g., physical PIN from the previous paragraph) would position the work better and help justify use of BendyPass very different and unfamiliar interaction modality. NEG +1275 neuroai19_32_1_1 They do not make direct comparisons to previous models or study quantitatively the results of the model with respect to its parameters. NEG +1124 midl20_56_4_2 Experiments show that the proposed method outperforms the model trained on the context-agnostic setting and acquires similar results to models trained by context-specific settings.1). NA +582 iclr20_1493_2_43 " I would be more than happy to significantly improve my score if these concerns can be addressed in the revision and corresponding rebuttal.""" NA +1327 neuroai19_37_3_5 " While not yet complete, we have sufficient evidence that a synthesis of these ideas could result in an understanding of how neural computation emerges from a combination of innate dynamics and plasticity"" What follows is a useful survey of a selection of ideas, by far not complete." NEG +983 midl19_52_2_2 They compare their method with two state of the art deep learning methods and illustrate superior performance on NRMSE, PSNR, SSIM and R2 metrics. NA +627 iclr20_2094_1_1 Namely, the framework is based on the actor-critic paradigm, and uses a conditional query learning model for performing composite actions (selections, rotations) in geometric bin packing. NA +1297 neuroai19_32_1_23 " This warranted some potentially interesting discussion though admittedly 4 pages isnt a lot of space.""" NA +323 iclr19_1399_1_16 I suggest checking the papers citing Bengio et al. (2009) to find lots of closely related papers. NA +1140 midl20_70_4_2 The method is then shown to significantly outperform the state-of-the-art methods of (Irvin and al., 2019; Allaouzi and Ahmed, 2019). POS +631 iclr20_2094_1_5 b) In the related work section, very little is said about Bin Packing Problems. NEG +1097 midl20_108_3_13 The method is well explained and the validation is strong with convincing results versus state of the art methods. POS +778 iclr20_855_3_14 Can we get the same conclusions on a different domain where other model-based methods have been successful; e.g. continuous control tasks? NA +1110 midl20_127_4_5 Such a system might speed up this process. NA +437 iclr19_495_1_9 The experiments also talk about 2D bandit problem, and again, without any descriptions. NEG +119 graph20_39_2_1 The authors utilise a range of methods in order to better understand the attitude and perspective of both participants to provide relevant and appropriate design insights for developing tools to support the visualisation of data collected during a clinical visit. NA +1086 midl20_108_3_2 A hard-parameters sharing network was presented with a shared, compressed representation branching out in task-specific networks. NA +790 iclr20_880_2_10 Whereas, if internal matrices have a number of dimensions lower than the rank of the original matrix, these matrices act as filters on features or feature combination. NA +341 iclr19_242_2_17 It is unclear what is the default batchsize for Imagenet. NEG +815 midl19_13_2_0 This paper presents a method for the instrument recognition task from laparoscopic images, using two generators and two discriminators to generate images which are then presented to the network to classify surgical gestures. NA +2 graph20_25_2_2 The evaluation consisted of two sessions (taking place one week apart) in which participants first created their passwords and then used them to sign in. NA +229 graph20_61_2_7 " I would suggest to use active voice instead of passive to clarify who contributed what (""The system was developed"", ""...was installed"")." NEG +1238 neuroai19_23_1_7 For example, is there something different about the feature maps that support this? NEG +1347 neuroai19_37_3_26 " Largely contradicts this one ""It is probable that revolutionary computational systems can be created in this way with only moderate expenditure of resources and effort"" I felt the paper could have done more to link with current state-of-the-art AI approaches." NEG +76 graph20_35_1_2 The second study expands on those findings to balance the depth and breadth of mind maps creation. NA +1291 neuroai19_32_1_17 The main place to improve is to have some quantitative analysis of the quality of their model perhaps MSE of image reconstruction. NA +1126 midl20_56_4_5 The method part is well-written and easy to understand. POS +371 iclr19_261_3_7 The humanhuman similarity score is pretty far above those of the best models, even though MTurkers are not optimized (and likely not as motivated as an NN) to solve this task. NEG +180 graph20_53_2_7 The paper further assessed the tool in an exploratory study looking at usability and induced workload, with promising results. POS +993 midl19_52_2_12 4- How does the specifics of the network architecture influence the performance? NEG +577 iclr20_1493_2_37 In particular, with such low-variance directions, at standard dataset sizes the distributions generated here are most likely statistically indistinguishable from their robust/non-robust counterparts (you can see hints of this in the fact that the CNN gets . NA +1158 midl20_77_4_5 However, one weakness of the paper was that the details of the experimental setup for data generation were not clear without following up the Gessert et al (2019) reference. NEG +1021 midl19_56_3_10 " Authors claim to introduce many concepts for the first time, such as the ""first demonstration that a deep generative architecture can generate high fidelity complex human anatomies in a [...] voxel space [from low-dimensional latents]""." NEG +1038 midl19_59_3_6 The medical decathlon (pseudo-url) would have provided easy access to more datasets and tasks. NA +1201 midl20_96_3_12 How would a generic linear classifier on the image histograms perform here, or perceptual hashing with a linear classifier on top? NA +1192 midl20_96_3_3 Iterative refinement is claimed to be semi-supervised learning. NA +895 midl19_40_3_13 It is improving the final performances, speeding up convergence, both ? NA +589 iclr20_1724_2_6 Finally, the localization task is challenging, especially when camera motion is introduced, with much space for improvement left for future work. NA +1046 midl19_59_3_14 " in section 5: ""Table 2 shows, that both IMM and T-IMM...""." NA +785 iclr20_880_2_5 However, in its present form, it is hard to understand why the claim is correct. NEG +1070 midl20_100_1_21 It is not obvious how to best get around this issue, since the first embryologist screening probably has false negatives, but you need to take it into account. NEG +311 iclr19_1399_1_4 The analysis of the results is quite insightful. POS +301 iclr19_1291_3_14 " Why is CommNet work worse than IRIC and IC in table 2?""" NEG +722 iclr20_57_3_6 The paper lacks insight about a principled way to label such examples, the costs associated with such labeling, and impacts of the labeling quality on accuracy. NEG +689 iclr20_526_3_10 In the current development, Theorem 1 only states that the optimization process will converge to the stationary points of the approximate ELBO objective (L1 in the paper's notation). NA +812 iclr20_934_1_10 Since the proposed method uses the multi-channel representation, how to set the number of channels pseudo-formula ? NEG +37 graph20_29_3_5 Age difference is one among many possible explanations, but one in which this paper rushes in nevertheless, at the expense of any other. NEG +353 iclr19_242_2_31 My main concern is that the benefit of this method is unclear. NEG +1283 neuroai19_32_1_9 Some development of the model could have been left to the references and didn't add much to their contribution (e.g. Taylor approximation to a Lie model) . NEG +714 iclr20_526_3_35 Also, I would like to see the test estimated NLL (via AIS) learning curves for VCD and AdVIL. NEG +236 graph20_61_2_14 " Also, before initiating collaborations, I would say that all parties must first be aware of each others contributions, so I would rephrase the reason as a ""lack of communication"" among them." NEG +1164 midl20_77_4_11 Section 2: In description of n-Path-CNN3D, extent should be extend Section 2, Dataset: For data generation, we consider various smooth curved trajectories with different motion magnitudes this is a bit vague, can you provide more information? NEG +729 iclr20_720_2_2 To me the proposed approach does not seem particularly novel and the idea that hierarchy can be useful for multi-task learning is also not new. NEG +1150 midl20_71_1_7 3) more details about the convGRU may be useful, for example its architecture. NEG +867 midl19_14_2_43 " 2] Maninis, Kevis-Kokitsi, et al. ""Deep retinal image understanding.""" NA +499 iclr19_938_3_4 Furthermore, the different baselines perform differently: there is no method that consistently performs well. NA +1152 midl20_71_1_9 " The conclusion is more like a validation for the usefulness of the temporal information, while technical novelty may not be very sufficient in this case.""" NEG +553 iclr20_1493_2_11 For example, a few very closely related works are as follows: - Adversarial examples are not Bugs, they are Features (pseudo-url): Ilyas et al (2019) demonstrate that adversarial perturbations are not in meaningless directions with respect to the data distribution, and in fact a classifier can be recovered from a labeled dataset of adversarial examples. NA +55 graph20_29_3_23 I doubt many designers would consider a clickable, 2.4-mm high font or icon on a touch screen in any case. NA +302 iclr19_1333_1_0 This paper proposes a new set of heuristics for learning a NN for generalising a set of NNs trained for more specific tasks. NA +794 iclr20_880_2_14 There are some possibilities, which have not been explored: 1) the performance improvement derives from the approximation induced by the representation of float or double in the matrices. NEG +286 iclr19_1091_1_24 " Why are the robotics priors not in Table 1?""" NEG +1064 midl20_100_1_15 In the discussion you almost exclusively focus on the work by Tran et al and why comparing with that work is unfair. NEG +219 graph20_56_1_26 It sounds like a classic case of theres nothing wrong with our system, just change the user. NEG +882 midl19_40_3_0 This paper attempt to do nuclei segmentation in a weakly supervised fashion, using point annotations. NA +387 iclr19_304_3_0 Overview: The authors aim at finding and investigating criteria that allow to determine whether a deep (convolutional) model overfits the training data without using a hold-out data set. NA +1022 midl19_56_3_11 However, I am aware of at least one work where such concepts have been proposed and explored already. NEG +390 iclr19_304_3_3 I have several issues with this work. NEG +1047 midl19_59_3_15 I guess this should actually be table 4. NEG +352 iclr19_242_2_30 I will keep my score and argue for the rejection of this paper. NA +193 graph20_56_1_0 The authors describe the design and implementation of a shape-based brushing technique targeted at selecting a particular type of data - trajectories. NA +235 graph20_61_2_13 The passive voice of the sentence does not help to identify who posited this reason: the authors of the submission or Vieira et al. [36]? NEG +1227 neuroai19_2_2_17 There are exiting directions in both AI and neuroscience this work could be take. NA +694 iclr20_526_3_15 I would appreciate a more forceful motivation of the relevance of MRFs rather than just stating it as a important model with applications. NEG +480 iclr19_866_1_13 The goal search process relies on a number of user-defined parameters - The nature of the instructions used for experimental evaluations is unclear. NEG +567 iclr20_1493_2_26 I believe a more measured conclusion (perhaps that we *need* more regularization methods, but even then we may not be able to get perfect robustness and accuracy) would better fit the strong results presented in the paper. NEG +116 graph20_36_1_30 The results presented in appendix do not seem so different, and I think the result will be even more similar with a little practice. NEG +1108 midl20_127_4_3 " This paper aims to solve the above problems by..."", but the authors use 2D ultrasound images made by a sonographer, so the system therefore does not solve these problems." NEG +653 iclr20_2094_1_27 " Under such circumstances, it is quite impossible to reproduce experiments. """ NEG +1169 midl20_85_3_1 To achieve this , the idea of introducing Dirichlet distribution after neural network is used from Evidential Deep Learning (EDL) paper. NA +1295 neuroai19_32_1_21 Weight sharing across shifted filters separates out feature and position yet many of their learned transformations are also translations. NA +757 iclr20_727_1_12 Minor point: - The extension of the method to Marked Temporal Point Processes in the Evaluation section seems out of place, esp. NEG +444 iclr19_495_1_16 Update: I feel the idea of this paper is straightforward, and the contribution is incremental. NEG +1089 midl20_108_3_5 The introduction and description of the state of the art, in addition to the main limitations of popular algorithms is very clear and interesting to read. POS +120 graph20_39_2_2 First, the authors attempted to identify a gap in the literature concerning how visualisation designs can support the review and analysis of user-generated data. NA +1072 midl20_100_1_23 The only way training size can influence AUC is by influencing the training of the model. NA +1198 midl20_96_3_9 A radical ablation study is clearly missing here. NEG +1048 midl19_59_3_16 " Figure 1 could have been a bit more clear """ NEG +62 graph20_29_3_30 In my experience, many pointing studies have error rates ranging from 0 to, say, 15%, perhaps more when the tasks or input devices make it particularly difficult. NA +743 iclr20_720_2_16 If this is the case, I feel like the empirical results are not novel enough to create value for the community and too tied to a particular approach to hierarchy which does not align with much of the past work on HRL. NEG +920 midl19_49_1_16 It is not convincing to claim that the clustering is correct since even a noise can be decoded into a normal image. NEG +453 iclr19_659_2_0 This paper proposes the deep reinforcement learning with ensembles of Q-functions. NA +515 iclr19_997_3_8 The proposed method is compared with the existing multi-objective methods in terms of classification accuracy, but if we focus on that point, the performance (i.e., error rate and FLOPs) of the proposed method is almost the same as those of the random search judging from Table 4. NEG +570 iclr20_1493_2_29 The RBF SVM, for small enough bandwidth can express any function and is convex, so no argument needs to be made about its ability to find the Bayes-optimal classifier. NA +223 graph20_61_2_1 Quality The methodology employed for conducting this research sources methods from diverse fields and is relevant. POS +1028 midl19_56_3_17 How do authors suggest to apply their approach to anatomies where it is impossible (in terms of feasibility and manual effort) to place a sufficiently large number of unique landmarks on the anatomy (e.g. smooth shapes, such as left ventricle in ACNN)? NEG +613 iclr20_2046_2_18 This is also the case for Theorems 2-4. NA +843 midl19_14_2_17 There are other existing data sets such as HRF (pseudo-url), CHASEDB1 (pseudo-url) and DR HAGIS (pseudo-url) with higher resolution images that are more representative of current imaging devices. NA +221 graph20_56_1_28 " Id like to see an inclusion of the user review. """ NA +1346 neuroai19_37_3_25 " We require a new class of theories that dispose of the simplistic stimulus-driven encode/ transmit/decode doctrine. """ NA +516 iclr19_997_3_9 It would be better to compare the proposed method to the existing multi-objective methods in terms of classification accuracy and other objectives. NEG +465 iclr19_659_2_12 Minor things: +The main idea is described too sketchily. NEG +52 graph20_29_3_20 I assume that strong design guidelines already exist for this? NA +1280 neuroai19_32_1_6 The statement that: GANs and VAE features are not typically interpretable. NA +485 iclr19_866_1_18 Similarly, what is the nature of the different action spaces? NEG +875 midl19_36_2_0 The paper is well-written, and easy to read and understand. POS +1331 neuroai19_37_3_9 Which leads me to a few concerns. NEG +818 midl19_13_2_3 Overall a clearly written paper, with nice visual results. POS +386 iclr19_261_3_28 " On learning to refer to things based on their discriminative properties. """ NA +155 graph20_43_1_2 Overall, I found the design of the study to be sound, as is the data analysis and modeling methodology. POS +1137 midl20_56_4_19 " Therefore I recommend the weak accept. """ NA +600 iclr20_2046_2_5 Experimental results show that the proposed algorithm outperform the MCTS algorithms. POS +1113 midl20_127_4_8 " The boxplot shows that six outliers are resolved by the AF-Net, so it can be debated if that is clinically relevant to reduce (6/435=)1.4% of the errors.""" NEG +248 graph20_61_2_26 Rendering in SVG with d3 might pose issues regarding accessibility, where efforts for compliance are left at the discretion of application developers rather than library developers. NEG +761 iclr20_76_2_2 Theresults quantify how smooth Gaussian data should be to avoid the curse of dimensionality, and indicate that for kernel learning the relevant dimension of the data should be defined in terms of how the distance between nearest data points depends on sample numbers. NA +870 midl19_25_3_0 The paper is well written and describes an interesting and relatively novel approach to solving multi-class classification in a clinical domain where overlap between classes is frequently a possibility. POS +855 midl19_14_2_29 Since the HRF data set contains images from normal, glaucomatous and diabetic retinopathy patients, I would suggest to use that one. NA +441 iclr19_495_1_13 However, there are more powerful variants of normalizing flows such as the Multiplicative Normalizing Flows or the Glow. NEG +1270 neuroai19_3_3_5 One needs to go see Appendix C to understand what the model used (SQL) consists in. NEG +1284 neuroai19_32_1_10 When they say steerable filter I was a little confused, do they just mean the basis vectors learned vary smoothly with respect to some affine transform parameter? NA +542 iclr20_1042_2_20 " 3) Experiments Finally, the experimental results do not look very compelling, it seems to be overall worse than the baselines in the two image datasets and slightly better in the audio dataset, so it's unclear that this approach is superior.""" NEG +1042 midl19_59_3_10 " The way table 2 is presented at the moment it seems like T-IMM is better than all methods also for ""100%""." NEG +701 iclr20_526_3_22 " What is meant by ""RBM loss"" in Fig. 2(d), I do not see this defined?" NEG +117 graph20_36_1_31 " In summary, the idea is interesting, but the design rationale is unclear, and it is unclear the results justify using this system.""" NEG +1312 neuroai19_36_1_0 Premise is that feedback alignment networks are also more robust to adversarial attacks. NA +628 iclr20_2094_1_2 Experiments are performed on several instances of 2D-BPP and 3D-BPP, Overall, bin packing problems are challenging tasks for DRL, and I would encourage the authors to pursue this research topic. NA +1393 neuroai19_59_3_18 Does limiting the synaptic time constants limit the intrinsic time constants, and if so by how much? NEG +206 graph20_56_1_13 I assuming - as one would consider the obvious choice - that directionality is taken from the direction of the sketched brush at the time the user draws it. NA +291 iclr19_1291_3_4 The paper is well written, easy to follow, and everything has been explained quite well. POS +1222 neuroai19_2_2_12 So I think the present work needs to be repitched slightly as solving credit assignment in an online/few shot learning setting. NEG +381 iclr19_261_3_19 Learning to follow navigational directions. NA +819 midl19_13_2_4 Mainly an incremental paper, proposing a combination of well established GAN-based networks to accomplish a classification task. NEG +972 midl19_51_2_12 For example, it is not clear what are the non-desirable artifacts, where are the eliminated nuclei and why the network has a harder time to learn. NEG +1069 midl20_100_1_20 In your case, you train on data that has already been filtered to only include positive decisions by embryologists, otherwise the eggs would not have been implanted. NA +200 graph20_56_1_7 There appears to be a set of small multiples for each of the two metrics. NA +596 iclr20_2046_2_1 It further establishes the sample complexity to determine optimal actions. NA +897 midl19_40_3_15 Section 2.3 should make the differences (if any) with Tang et al. explicit. NEG +1178 midl20_85_3_13 " Overall, the idea is fine. """ POS +854 midl19_14_2_28 That would be equivalent to assume that the new data set(s) does (do) not contain the annotations, and will allow to quantify the performance there. NA +1328 neuroai19_37_3_6 For example, many of the interactions between myriad excitatory and inhibitory types across brains regions and neuromodulators, of which dopamine is just one of several, is largely unknown. NEG +1389 neuroai19_59_3_14 " It seems that one of the main points of the work is that ""longer intrinsic timescales correspond to more stable coding"", but I didn't find that this point was made very convincingly." NEG +804 iclr20_934_1_0 This paper proposed a dual graph representation method to learn the representation of nodes in a graph. NA +1061 midl20_100_1_12 You do not report results for the embryologist trained LSTM, so what do you use this LSTM for? NEG +981 midl19_52_2_0 This paper proposes to use a CNN architecture to reconstruct MR Fingerprinting parametric maps. NA +792 iclr20_880_2_12 Hence, without non-linear functions, where is the added value of the method? NEG +769 iclr20_855_3_5 The first is the presentation of the empirical results. NA +1174 midl20_85_3_6 Now, it is difficult to connect use of prior and improvement in ECE. NEG +813 iclr20_934_1_11 How does this parameter affect the performance? NEG +1115 midl20_135_3_1 It is an interesting idea and the quality is overall rather good for an abstract paper. POS +505 iclr19_938_3_10 It is unclear how the model actually operates and uses attention during execution. NEG +891 midl19_40_3_9 " How resilient is the method to ""forgotten"" nuclei ; i.e. nucleus without a point in the labels ?" NEG +348 iclr19_242_2_24 It contradicts with the authors other explanation. NEG +1037 midl19_59_3_5 When used on another dataset they do not show gains anymore. NA +1141 midl20_70_4_3 The paper reads well and the methodology seems to be interesting. POS +1392 neuroai19_59_3_17 How does this relate to their synaptic time constants? NEG +340 iclr19_242_2_16 In figure 1 (b), the results of M=4,8,16,32 are very similar, and it looks unstable. NEG +786 iclr20_880_2_6 In fact, the model presented in the paper has a major obscure point. NEG +615 iclr20_2046_2_20 However, it would be better to have some discussion earlier right after these theorems are presented. NEG +154 graph20_43_1_1 Models are fit which account for these differences, on both new data gathered from 12 participants, and data sets gathered from several past studies. NA +1292 neuroai19_32_1_18 Then this evaluation could be used to study impacts of the parameters of their model which could then lead to neural hypotheses. NA +1263 neuroai19_29_1_6 The question of how the brain and artificial network can perform relational reasoning is critical in both fields, since many believe that it may be one of the primary ingredients of intelligence. NA +77 graph20_35_1_3 Both studies compare the new mind mapping tool to digital options without computer assistance. NA +1387 neuroai19_59_3_12 " For instance, the claim of ""stronger cue-specific differences across the cue stimulus window"" between fast and slow intrinsic timescale neurons in the RNN model isn't clearly supported by the heatmap in Figure 3 -- the cue-specific differences for the short instrinsic timescale group to me appears to be at least as great as that of the long intrinsic timescale group within the cue stimulus window." NEG +287 iclr19_1291_3_0 This work is an extension to the work of Sukbaatar et al. (2016) with two main differences: 1) Selective communication: agents are able to decide whether they want to communicate. NA +322 iclr19_1399_1_15 Even the results with the VGG network are very far from the best available models. NEG +686 iclr20_526_3_7 I have a minor issue with the discussion (in the last paragraph of sec. 3.2) stating that the theoretical statement of the proposed objective relies on a much weaker assumption than the nonparametric assumption made in the theoretical justification of GANs. NEG +1076 midl20_100_1_27 Maybe you meant the size of the test set? NA +1252 neuroai19_26_1_8 The paper takes a crudely 'neuroscience inspired' concept (though, admittedly it could simply be 'task structure' inspired) and builds a simple model from it, which it benchmarks on a appropriately designed simplest-working-example. NA +913 midl19_49_1_8 One major concern is whether the results are reliable: 1. NEG +1267 neuroai19_3_3_2 Although innovative and promising, the work is quite preliminary and would benefit from comparison and validation with real human behavior. NEG +1188 midl20_90_2_9 " However, if different models were trained for predicting each parameter, not only training but also prediction would not be efficient.""" NEG +82 graph20_35_1_8 One small point that could be clarified is why a between subjects design was chosen over a counterbalanced within subjects. NEG +512 iclr19_997_3_5 Pros - The performance of the proposed method is better than the existing multi-objective architecture search methods in the object classification task. POS +1395 neuroai19_59_3_20 The authors use an artificial network model to shed light on the biological mechanisms enabling and shaping working memory in the brain. NA +964 midl19_51_2_4 2- Two step approach combining despeckling and generative networks are reasonable for the task. POS +464 iclr19_659_2_11 However, the authors didnt show these results in the paper. NEG +1128 midl20_56_4_8 Opposite to the Method part, it's hard to read the abstract and introduction. NEG +359 iclr19_242_2_37 I have suggested the authors to compare with stronger baselines to demonstrate the benefits. NA +1367 neuroai19_54_3_6 Is the attribution analysis stable for different stimulus frequencies? NA +929 midl19_49_1_25 Details of training should be more clearly written. NEG +1135 midl20_56_4_17 Results show the effectiveness of the proposed method. POS +333 iclr19_242_2_9 Even provided more computing resources, the proposed method is not faster than small batch training. NEG +10 graph20_25_2_10 The paper never justifies why Bend Passwords [33] is the best design to adapt for users who are visually impaired. NEG +738 iclr20_720_2_11 So, if you generically apply Option-Critic, it would in fact be possible to disentangle the inductive bias of hierarchy from the inductive bias of temporal abstraction by using options that always terminate. NA +470 iclr19_866_1_3 Such a modular approach has the advantage that the instruction-to-goal and goal-to-policy mappings can be trained separately and, in principle, allow for swapping in different modules. NA +831 midl19_14_2_5 The contribution is original in the sense that complementing data sets is a really challenging task, difficult to address with current available solutions. POS +32 graph20_29_3_0 Through four studies, this paper proposes to lift a theoretical limitation in the application range of the Dual Gaussian Distribution Model, namely that it could also work when touch acquisition occurs from a touchscreen to that same touchscreen. NA +150 graph20_39_3_11 It seemed that the initial statement was that general guidelines were not useful because of the uniqueness at each patient. NA +413 iclr19_304_3_27 Although you didnt show anything but only state assumptions or claims (which may be reasonable but are not backed up here). NEG +580 iclr20_1493_2_41 A suggestion rather than a concern and not impacting my current score: but it would be very interesting to see what happens for robustly trained classifiers on the symmetric and asymmetric datasets. NA +382 iclr19_261_3_21 Learning Symmetric Collaborative Dialogue Agents with Dynamic Knowledge Graph Embeddings. NA +793 iclr20_880_2_13 How the proposed method can have better results. NA +1077 midl20_100_1_28 In that case, it is the ratio of positive/negative that is relevant. NA +376 iclr19_261_3_13 Are the humans? NEG +1324 neuroai19_37_3_2 Its an opinion piece. NA +614 iclr20_2046_2_19 The authors give some concrete examples in Section 6.2 for these bounds. POS +1330 neuroai19_37_3_8 The dynamics of neuromodulation is largely unknown. NA +378 iclr19_261_3_16 Framing: there is a lot of work in collaborative / multi-agent dialogue models which you have missed see refs below to start. NEG +89 graph20_36_1_3 This is overall an interesting idea of interactive system supporting skill acquisition. POS +692 iclr20_526_3_13 One this last point, it seems ironic to me that the proposed strategy for training the MRF is through the use of three separate directed graphical models (an encoder q(h | x), a decoder and a VAE to model the approximate prior over the latents h). NA +601 iclr20_2046_2_6 Cons: However, there are several issues that should be addressed including the presentation of the paper: The algorithm seeks to combine A* search with MCTS (combined with policy and value networks), and is shown to outperform the baseline MCTS method. NEG +213 graph20_56_1_20 One would expect that trying some combination would be an obvious step, especially given the unclear feedback from the expert review. NEG +833 midl19_14_2_7 However, it is the first time that it is applied for complementing data sets and have some interesting modifications that certainly ensures novelty in the proposal. POS +332 iclr19_242_2_8 It is unclear to me what is the benefit of the proposed method. NEG +133 graph20_39_2_15 A utilisation of these perspectives in framing the research ideas would have done more good to the paper than proposing a new design space for visualisation of user-generated data. NEG +346 iclr19_242_2_22 I fail to understand the the authors augmentation. NEG +555 iclr20_1493_2_13 A Discussion of Adversarial Examples are not Bugs they are Features (pseudo-url): Nakkiran (2019) actually constructs a dataset (called adversarial squares) where the Bayes-optimal classifier is robust but neural networks learn a non-robust classifier due to label noise and overfitting. NA +7 graph20_25_2_7 It is particularly important to evaluate technology with target stakeholders. POS +28 graph20_26_3_3 In particular, clarifications around the motivation behind the path tracing task, and additional related work that have utilized path tracing to determine endpoints (e.g., [17], [18]) and to mark or detect features along a path (e.g., [66]) were helpful in positioning the contributions of this work in relation to prior work. POS +320 iclr19_1399_1_13 The results in Figure 3 are very far from the state of the art. NEG +1225 neuroai19_2_2_15 In understanding the model, it would be useful to more explicitly define the model. NEG +187 graph20_53_2_14 For some reason, the actual qualitative aspects of the study are then reported as a subsection in the discussion (6.3 - Comment Observations). NA +1206 midl20_96_3_17 Writing, experimental setup and methodological proposals need to be improved and condensed. NEG +634 iclr20_2094_1_8 Again, a brief discussion about those results would be relevant. NEG +552 iclr20_1493_2_10 Prior work: the paper seems to ignore a plethora of prior work around studying adversarial robustness and understanding its roots. NEG +304 iclr19_1333_1_2 The issue of model selection (clearly the main issue here) is not addressed. NEG +31 graph20_26_3_6 " For example, page 3: HoloLense -> Hololens. """ NA +1083 midl20_100_1_34 " I am aware of the page limitation, so maybe MIDL should allow an extra page solely for an image of the raw data.""" NA +796 iclr20_880_2_16 2) the real improvement seems to be given by the initialization which has been obtained by using the non-linear counterpart of the expansion; to investigate whether this is the case, the model should be compared with a compact model where the initialization is obtained by using the linear product of the non-linear counterpart of the expanded network. NEG +1157 midl20_77_4_4 The discussion of the results reveals findings that may well be of interest to others. POS +174 graph20_53_2_1 is a tool for authoring object component behaviour within VR. NA +766 iclr20_855_3_2 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). NA +245 graph20_61_2_23 " For further inspiration on visualization for comparing (resident) profiles, I'd suggest to browse other works by Plaisant et al. in addition to [29]: pseudo-url pseudo-url IMPLEMENTATION DETAILS The implementation details report on constraints that may be too project-specific (with occurrences of ""project"" or ""the University"") and would gain to be generalized." NEG +1205 midl20_96_3_16 There will be domain shift problems for the simple methods but same is true for the presented method. NA +145 graph20_39_3_6 There are a few comments I have about the paper that I describe below. NEG +90 graph20_36_1_4 The system remains simple. POS +682 iclr20_526_3_3 That said, it does seems like a fairly creative combination of existing approaches. NEG +310 iclr19_1399_1_3 The method does show some modest improvements in the experiments provided by the authors. POS +93 graph20_36_1_7 What are the design choices? NEG +1156 midl20_77_4_3 The methods employed seem reasonable and quantitative evaluation is performed to compare them. POS +924 midl19_49_1_20 It is a self-contradictory statement. NA +1173 midl20_85_3_5 Also, I would be convinced that the variance would increase for out of distribution test samples because you used a prior that enforced uncertainty of all labels. NEG +1035 midl19_59_3_3 Method only evaluated on one dataset (BRATS). NEG +243 graph20_61_2_21 The decisions on color scales adjustments to highlight under-performance while shadowing over-performance on EPA count per rotation is well motivated by contextual needs. POS +214 graph20_56_1_21 The last point leads me to what I see as *the* major weakness of the paper. NA +998 midl19_52_2_17 6- The quantitative results are yielded using multiple segmentation masks due to MR physics related concerns. NEG +853 midl19_14_2_27 It would be interesting to simulate such an experiment by taking an additional data set with vessel annotations (e.g., some of those that I suggested before, HRF, CHASEDB1 or DR HAGIS) and evaluate the performance there, without using any of their images for training. NEG +425 iclr19_304_3_39 Does that mean that you assume that whenever the training accuracy drops lower than that of the model without regularization, it starts to underfit? NA +1221 neuroai19_2_2_11 The credit assignment problem exists in these cases also. NEG +226 graph20_61_2_4 Signifiance The system has been designed and developed and evaluated so that it ended up being useful to domain experts (medical residents and their reviewers). POS +1133 midl20_56_4_15 There can be more discussion here.The authors propose a framework to utilize one model under different acquisition context scenarios. NA +0 graph20_25_2_0 The submission presents evaluation of BendyPass, a prototype based on Bend Passwords design [33], with visually impaired people. NA +401 iclr19_304_3_14 What do you mean by the data being independent? NEG +84 graph20_35_1_10 The results are individually compelling, but what does it mean all together? POS +162 graph20_43_1_9 It seems awkward to use such a similar term here, when C-D manipulation is not the focus. NEG +114 graph20_36_1_28 What is the objective: people's perception or a metric? NEG +670 iclr20_305_3_2 The leader is modeled as a semi-MDP with event-based policy gradients and modules to model/predict followers' actions. NA +1062 midl20_100_1_13 If you dont use it, remove it from the section. NEG +889 midl19_40_3_7 Few questions: - Since the method is quite simple and elegant, I expect it could be adapted to other tasks. POS +625 iclr20_2046_2_30 " In the first paragraph of Section 6.2, there is a typo: V*=V_{l*}=\eta should be V*-V_{l*}=\eta ?""" NEG +952 midl19_51_1_20 A lengthy study, I agree, but a necessity in light of other recent works highlighting how dangerous is to use GANs for this kind of tasks. NEG +807 iclr20_934_1_3 Overall, the idea is presented clearly and the writing is well structured. POS +258 iclr19_1049_1_3 The method is flexible and different entities correspond to different rules. POS +624 iclr20_2046_2_29 In fact, the proof of the theorems could be moved to appendices. NA +826 midl19_14_2_0 The authors present a deep learning method for fundus image analysis based on a fully convolutional neural network architecture trained with an adversarial loss. NA +234 graph20_61_2_12 " One reason for this gap seems to be the lack of collaboration among the developers, end-users and visualization experts.""" NA +135 graph20_39_2_17 There is the question of how the data and the proposed guidelines might bring about some implications for design (Dourish, 2006) and practice. NA +616 iclr20_2046_2_21 The experimental results are carried out under the very simplified settings for both the proposed algorithm and the baseline MCTS. NEG +836 midl19_14_2_10 Taking this into account, I would suggest the authors to incorporate at least one paragraph in Related works (Section 2) describing the current existing approaches to do that. NA +842 midl19_14_2_16 Despite the fact that this set has been the standard for evaluating blood vessel segmentation algorithms since 2004, the resolution of the images is extremelly different from the current ones. NEG +959 midl19_51_1_27 " It is necessary to run a study to confirm that in a similar way that CM images were confirmed having diagnostic value and could therefore be used instead of H&E stained images.""" NEG +592 iclr20_1724_2_9 The authors recognize that since the dataset is synthetically generated it is not necessarily predictive of how methods would perform with real-world data, but still it can serve a useful and complementary role similar to the one CLEVR has served in image understanding. POS +343 iclr19_242_2_19 The baselines are fairly weak, the authors did not compare with any other method. NEG +1146 midl20_71_1_3 The improvement gained by the proposed method validates the effectiveness of recurrent units, and the most significant gain is from the false positive rates. NA +293 iclr19_1291_3_6 The comparison between their model with three baselines was extensive; they reported the mean and variance over different runs. POS +364 iclr19_242_2_42 " I think it could at least be improved for clarity. """ NEG +178 graph20_53_2_5 This is briefly addressed in the limitations, but I would have found some discussion of this aspect very helpful, especially earlier when introducing the research motivation. NEG +659 iclr20_2157_3_6 Some of these should serve as baselines. NEG +996 midl19_52_2_15 Does the order of concatenation influence the results? NEG +14 graph20_25_2_14 Other designs exist (e.g., work by Das et al. (2017) is just one example. NA +165 graph20_45_2_0 The paper proposes a new visualization scheme that combines the properties of scatterplots and parallel coordinates plots (PCPs): the Cluster-Flow Parallel Coordinates Plot (CF-PCP). NA +59 graph20_29_3_27 AMOUNT OF ERROR Throughout the paper, prediction errors (additive) up to 10% are described as small, and that is surprising (5% in Exp 1, 10% in Exp 2, 7% in Exp 3, 10% in Exp 4). NEG +709 iclr20_526_3_30 With respect to Deep Boltzmann Machine (DBM), I would prefer to see quantitative comparisons against published results. NEG +676 iclr20_305_3_11 Supporting arguments The approach seems sound and conceptually related to a multi-agent generalization of STRAW pseudo-url, where a planner predicts / commits to an action-plan for a single agent. POS +780 iclr20_880_2_0 This paper is extremely interesting and quite surprising. POS +424 iclr19_304_3_38 " I neither understand ""As the accuracy curve is also monotone decreasing with increasing regularization we will also detect the convexity by a steep drop in accuracy as depicted by the marked point in the Figure 1(b)"" nor do I understand ""accuracy over regularization curve (plotted in log-log space) is constant""?" NEG +292 iclr19_1291_3_5 The experiments are competent in the sense that the authors ran their model in four different environments (predator and prey, traffic junction, StarCraft explore, and StarCraft combat). POS +1065 midl20_100_1_16 Instead, you should have made the comparison and highlighted the differences clearly. NEG +1242 neuroai19_23_1_11 It does not seem like predictive coding is the main thing going on in V1 (Stringer et al., Science 2019), so Id be curious how the authors think that should be taken into account in the future. NA +329 iclr19_242_2_4 The proposed method is very simple. NEG +429 iclr19_495_1_1 Although the concept of normalizing flow is simple, and it has been applied to other models such as VAE, there seems no work on applying it for policy optimization. POS +1012 midl19_56_3_1 The decoder network is made possible by a newly proposed architecture that is based on inception-like transpose convolutional blocks. NA +79 graph20_35_1_5 Overall, this paper is an interesting exploration of a novel area of computer supported brainstorming. POS +1052 midl20_100_1_3 The main weakness of the paper is in the methods section. NEG +937 midl19_51_1_5 It has the potential to improve pathology and cancer diagnosis by making it simpler and quicker The results of this work look visually convincing. POS +834 midl19_14_2_8 The paper is well written and organized, with minor details to address in this matter (see CONS). POS +749 iclr20_727_1_4 The writers have put their contributions in context well and the presentation of the paper itself is very clear. POS +890 midl19_40_3_8 Do you have any ideas in mind ? NA +254 graph20_61_2_32 " Audio quality of the voice over could be improved with a proper microphone and recording settings. """ NA +1361 neuroai19_54_3_0 The authors state three high-level improvements they want to make to CNN-based models of neural systems: 1 & 2) Capturing computational mechanisms and extracting conceptual insights. NA +136 graph20_39_2_18 Although the issues of implication for design has been misunderstood and widely misrepresented, what the proposed design guideline sought to point to might be regarded as some form of outlining implications for a design practice that is minimal and non-representative. NEG +852 midl19_14_2_26 However, since IDRID does not have vessel annotations, it is not possible to quantify the performance there. NA +723 iclr20_57_3_7 The experiments are not making a convincing case that similar improvements could be obtained on a larger class of problems. NEG +859 midl19_14_2_33 It is not clear if the values for the existing methods in Table 2 correspond to the winning teams of the IDRID challenge. NEG +969 midl19_51_2_9 Even though it is mentioned by the authors that these images resemble to noisy RCM, this should be either referenced or shown. NEG +496 iclr19_938_3_1 They used an attention mechanism over agent policies as an input to a central value function. NA +950 midl19_51_1_18 " The main contribution of the paper is scarcely justified by the statement ""...they confirmed that the images were similar to those in routine""." NEG +805 iclr20_934_1_1 In particular, it learns the embedding of paired nodes simultaneously for multiple times, and use the mean values as the final representation. NA +129 graph20_39_2_11 The analysis of the patient's interview provided a bigger picture of the different perspectives, and which makes the different factors more relational and understandable. NA +337 iclr19_242_2_13 Or apply distributed knowledge distillation like in (Anil 2018 Large scale distributed neural network training through online distillation) 3. NA +1388 neuroai19_59_3_13 I would be curious to know if making the input weaker or only giving it to a random subset of neurons makes this phenomenon more apparent. NEG +608 iclr20_2046_2_13 For example, in line 8 of Algorithm 2, why only the top 3 child nodes are added to the queue? NEG +195 graph20_56_1_2 The authors do an excellent job of describing the problem and grounding the approach in previous work. POS +909 midl19_49_1_4 The entire workflow is quite clear and complete. POS +584 iclr20_1724_2_1 " The dataset is an extension of CLEVR using simple motions of primitive 3D objects to produce videos of primitive actions (e.g. pick and place a cube), compositional actions (e.g. ""cone is rotated during the sliding of the sphere""), and finally a 3D object localization tasks (i.e. where is the ""snitch"" object at the end of the video)." NA +103 graph20_36_1_17 But it also makes me think about the actual difficulty of performing such art (I never tried myself). NA +1285 neuroai19_32_1_11 Their statement of the novelty of their method: (1) allowing each feature to have its own transformation was not clear. NEG +1299 neuroai19_34_2_1 The present paper makes the important case that random networks should be included as a matter of course in DCNN modelling projects, and sounds a note of caution about the field's temptation to over-interpret the particular features learned by high-performing trained networks. NA +719 iclr20_57_3_3 This authors evaluted their approach on two tasks: Text Classification and Sequence Labeling. NA +884 midl19_40_3_2 The idea is to generate two labels maps from the points: a Voronoi partitioning for the first one, and a clustering between foreground, background and neutral classes for the second. NA +776 iclr20_855_3_12 But models can also be used for value function estimation (Model Based Value Expansion) and reducing gradient variance(using pathwise derivatives). NA +849 midl19_14_2_23 It would also be interesting to analyze the differences in a qualitative way, as in Fig. 3 (b). NEG +894 midl19_40_3_12 Since there is so much dissimilarity between ImageNet and the target domains, I expect it to be mostly a glorified edge detector. NA +227 graph20_61_2_5 I advocate for accepting this submission. NA +1309 neuroai19_34_2_11 I'm not a big fan of the asterisks in Figures 3A and 3B used to indicate the best layers in various model tests. NEG +126 graph20_39_2_8 We need more detail to determine whether what the data suggest reflect the subjective perspective of the different users that participated in the study. NEG +405 iclr19_304_3_19 Better training error? NA +775 iclr20_855_3_11 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. NA +821 midl19_13_2_6 The presented evaluation is limited, with training done on only 8 datasets, which in this particular case is a limitation due to the importance of presenting the networks with different backgrounds from various surgical sites and perspectives during surgery. NEG +711 iclr20_526_3_32 It seems as though, in the application of AdVIL to the DBM, the authors are exploiting the structure of the model in how they define their sampling procedure. NEG +1313 neuroai19_36_1_1 " The authors show because the ""gradient"" in the feedback pathway is a rough approximation, it is hard to use this gradient to train an adversarial attack." NA +1060 midl20_100_1_11 As I read it, UBar is the same LSTM just trained on clinical outcomes. NA +303 iclr19_1333_1_1 This particular recipe might be reasonable, but the semi-formal flavour is distracting. NEG +267 iclr19_1091_1_5 The most important point of critique is that the conclusion that the split representation is the best is at best premature. NEG +545 iclr20_1493_2_2 In the other case, the Bayes-optimal classifier is robust, but neural networks fail to learn the robust decision boundary. NA +212 graph20_56_1_19 I found it odd that at the authors retained both metrics, delivering different results, without trying some blended version that might reduce complexity for the user. NEG +896 midl19_40_3_14 Minor improvements for the camera ready version, in no particular order: Tang et al. 2018 was actually published at ECCV 2018, the bibliographic entry should be updated. NA +916 midl19_49_1_11 Please compare to other representation learning methods such as sparse coding (e.g. spherical K-means, dictionary learning), dimension reduction (e.g. PCA, t-sne). NA +675 iclr20_305_3_8 Decision (accept or reject) with one or two key reasons for this choice. NA +814 iclr20_934_1_13 " Some unsupervised network embedding baseline methods, such as DeepWalk and Node2Vec, should be included into the experiment section. """ NEG +342 iclr19_242_2_18 In Table 1, the proposed method tuned M as a hyperparameter. NA +549 iclr20_1493_2_6 The paper also definitively proves that there are realistic datasets where the Bayes-optimal classifier is non-robust, which goes against quite a bit of conventional wisdom in the field and opens up many new paths for research. POS +966 midl19_51_2_6 Error measures presented in Table 1 needs to help readers to identify the benefit of the proposed neural network. NEG +295 iclr19_1291_3_8 Right now, the authors explain a bit about the model performance in Starcraft combat, but I found the explanation confusing. NEG +695 iclr20_526_3_16 What is unique about the MRF formalism that -- for practical applications -- could not be effectively captured in a directed graphical model? NEG +73 graph20_29_3_41 " Fig. 12 should also show the actual success rates measured in these studies.""" NA +605 iclr20_2046_2_10 These discussions are critical to understand the merit of the proposed algorithms. NA +104 graph20_36_1_18 I expected more discussion on this point in the paper. NEG +645 iclr20_2094_1_19 Also, it seems that the MDP is specified for a single instance of 3D-BPP. NA +1365 neuroai19_54_3_4 The technical aspects of the paper seem correct, though I have some higher-level conceptual concerns. POS +1040 midl19_59_3_8 " It shows that for ""100%"" T-IMM actually is not significantly better than most of the other initialization strategies." NEG +439 iclr19_495_1_11 I have seen {-1, 1}^2, but not [-1, 1]^2). NA +1019 midl19_56_3_8 First, it is not fully clear where this number 3 comes from, and second, the quality of the work speaks for itself. NEG +91 graph20_36_1_5 This will not be a revolution, but it might be of interest. NEG +1112 midl20_127_4_7 The authors also do not include a Section with a discussion. NEG +1217 neuroai19_2_2_7 But the related work in Section 2 then goes on to talk about the efficiency of backprop for solving online learning and few-shot learning tasks. NA +908 midl19_49_1_3 Clustering of aortic value prosthesis shapes has a high contribution to personalized medicine. POS +102 graph20_36_1_16 I wish there was a condition with these schematics only. NEG +238 graph20_61_2_16 Requirement analysis was conducted through focus groups including active participation of domain experts (including involving them in sketching their desired features for data presentation). POS +1203 midl20_96_3_14 Can't simple heuristics perform at least as well? NEG +942 midl19_51_1_10 The study has potential and could have interesting applications in clinical settings. POS +938 midl19_51_1_6 Both the de-speckle network and the GAN appear to deliver very good results, at least at first glance. POS +5 graph20_25_2_5 This submission contributes new knowledge about how users who are visually impaired can enter passwords. POS +1262 neuroai19_29_1_5 Overall the writing is relatively clear, but it would have been beneficial to describe the hypotheses more explicitly, e.g. what neural activity would be expected for a place, grid, or concept representation with respect to MNIST. NEG +965 midl19_51_2_5 3- Qualitative stained image results look promising Cons: 1- Median filter is used after the despeckling network, however it is not clear the added benefit of using median filter in despeckling process. NEG +18 graph20_25_2_18 Thus, ideally the evaluation would compare other ways that participants can enter PIN passwords. NEG +121 graph20_39_2_3 What is missing is a clear articulation of the research problem and question within the literature provided. NEG +661 iclr20_2157_3_8 It is also not clear from the literature if these models are really working so I think these results should be presented in a more detail. NEG +1172 midl20_85_3_4 It is not clear why calibration is reported and not simple measures of uncertainty like variance or entropy? NEG +274 iclr19_1091_1_12 Furthermore, even the tests in the appendix are not comprehensive enough to to warrant the conclusion as written. NEG +433 iclr19_495_1_5 For example, normalizing flows are defined in Section 4, and then it is directly claimed that normalizing flows can be applied to policy optimization, without giving details on how it is actually applied, e.g., what is the objective function? NEG +811 iclr20_934_1_8 Thus, the novelty is incremental. NEG +1300 neuroai19_34_2_2 Comprehensive data measurement and modelling pipeline. POS +1314 neuroai19_36_1_2 The basic premise is very strange. NEG +181 graph20_53_2_8 This consisted of a small user study (N=16) featuring qualitative and quantitative measures. NA +184 graph20_53_2_11 " The quantitative data is described as ""qualitative"" for some reason, even when referring to barplots in Figure 9." NA +406 iclr19_304_3_20 I dont understand the assumptions. NEG +865 midl19_14_2_39 " iii) Explanation of the method... [1] Zhao, Yitian, et al. ""Automated Vessel Segmentation Using Infinite Perimeter Active Contour Model with Hybrid Region Information with Application to Retinal Images.""" NA +520 iclr19_997_3_13 Please elaborate on the procedure and settings of the Bayesian network used in this paper. NEG +1003 midl19_52_2_22 Please elaborate on this. NA +380 iclr19_261_3_18 References Vogel & Jurafsky (2010). NA +427 iclr19_304_3_41 " In my view, this evaluation of the (vague) criteria is not fit for showing their possible merit. """ NEG +78 graph20_35_1_4 They find that QCue produces more balanced and detailed mind maps and that some mind mapping tasks may be better suited to this type of computer intervention than others. NA +1009 midl19_52_2_28 term in Fig.2. NA +1394 neuroai19_59_3_19 The same type of comments apply to the second part of the results, which demonstrates that a task that doesn't require working memory results in neurons with shorter intrinsic timescales compared to the working memory task. NA +607 iclr20_2046_2_12 Many design choices for the algorithms are not clearly explained. NEG +1236 neuroai19_23_1_5 Does PredNet outperform other user-submitted models? NA +1385 neuroai19_59_3_10 The technical details are presented clearly on the whole. POS +1268 neuroai19_3_3_3 No comparison with human data. NEG +1057 midl20_100_1_8 These must be provided in a supplement to allow reproducability. NEG +1091 midl20_108_3_7 The decision to supervised the feature extraction in a multi-task setting is good and makes sense. POS +1101 midl20_119_2_2 The motivation and methodology are well explained with proper reference works. POS +1129 midl20_56_4_9 Some typo problems lie here. NEG +1223 neuroai19_2_2_13 Or discuss how it can be extended to more general learning problems. NEG +958 midl19_51_1_26 It is necessary to prove that the generated images retain their important diagnostic value. NEG +559 iclr20_1493_2_17 Adversarially robust generalization requires more data (pseudo-url): Schmidt et al show a setup where many more samples are required for adversarial robustness than for standard classification error. NA +846 midl19_14_2_20 I would suggest to include the F1-score and the area under the Precision/Recall curve, instead, which have been used already in other studies (see [1] and [2], for example, or Orlando et al. 2017 in the submitted draft). NEG +690 iclr20_526_3_11 Clarity: I found the paper to be very well written with a clear exposition of the material and sound development of the technical details. POS +1350 neuroai19_53_1_0 The authors consider how biologically motivated synaptic eligibility traces can be used for backpropagation-like learning, in particular by approximating local gradient computations in recurrent neural networks. NA +1298 neuroai19_34_2_0 The surprisingly high power of randomly weighted DCNNs is a point that has popped up a couple of times in recent human fMRI / MEG work. NA +436 iclr19_495_1_8 I can't get how exactly normalizing flows + TRPO works. NEG +490 iclr19_866_1_23 I wouldn't consider the results reported in Section 4.5 to be ablative studies. NEG +1390 neuroai19_59_3_15 " The work would have benefited from a discussion of the implications of longer intrinsic timescale neurons retaining task-relevant information for longer -- in particular, this finding feels a bit ""trivial"" without the case being made for why this should push understanding in the field." NEG +1277 neuroai19_32_1_3 They show images of a single reconstruction but no quantification of reconstruction quality or comparison to previous methods. NEG +69 graph20_29_3_37 Second, 29% and 38% error seems alarmingly high. NEG +808 iclr20_934_1_4 But the novelty is limited. NEG +1184 midl20_90_2_5 The results of the model was compared also to the state of the art.From the following sentence, I understand that for each pathology, a different model was trained. NA +11 graph20_25_2_11 There are many other potential designs out there and the paper does not fully explore the potential design space before picking Bend Passwords [33]. NEG +144 graph20_39_3_5 The writing is clear and the paper is easy to read. POS +917 midl19_49_1_13 This study did not give a gold-standard for shape clustering (though it could be difficult). NEG +639 iclr20_2094_1_13 A. Khan has also found approximation algorithms for the 3D Knapsack problem with rotations. NA +176 graph20_53_2_3 The tool is a very useful and novel contribution, although I have some questions about the validity of the use case scenario. NEG +994 midl19_52_2_13 Why do the authors reuse the input of a temporal block to its output and how does this influence the performance? NEG +440 iclr19_495_1_12 It seems that the authors only use the basic normalizing flow structures studied in Rezende&Mohamed (2015) and Dinh et al (2016). NEG +521 iclr19_997_3_14 " It would be better to provide discussions of recent neural architecture search methods solving the single-objective problem. """ NEG +1378 neuroai19_59_3_3 In particular, the setting of synaptic decay constants is an important detail in a paper about working memory. NA +1004 midl19_52_2_23 8-The lack of scalability and the requirement of computational time is highlighted in the introduction and abstract. NA +143 graph20_39_3_4 It is a qualitatively-driven paper, but I believe it provides much insight into what providers would like in patient visualizes, and takes into account how patients already record their information. POS +753 iclr20_727_1_8 Another, relatively small point which the authors glance over is the matter of efficient training. NEG +519 iclr19_997_3_12 For example, what is the drawbacks of the number of parameters, what is the advantages of FLOPs for multi-objective optimization? NEG +822 midl19_13_2_7 Indeed the critical factor is not to capture the instrument's appearance but rather model how variable the anatomical environment is. NA +428 iclr19_495_1_0 This paper generalizes basic policy gradient methods by replacing the original Gaussian or Gaussian mixture policy with a normalizing flow policy, which is defined by a sequence of invertible transformations from a base policy. NA +612 iclr20_2046_2_17 The authors need to give more discussion and explanation about it. NEG +1032 midl19_59_3_0 Transfer learning and dealing with small datasets is an important area of research - The paper proposes a novel method, enabling pretraining on several different tasks instead of only one dataset (e.g. ImageNet) like done most of the times - Results show clear performance increase on small datasets - Proper experiment setup and validation - Clearly written and comprehensible - Code is openly available - Little comparison to other state-of-the-art methods for transfer learning. POS +156 graph20_43_1_3 I also think that the overall motivation of understanding whether interfaces with distinct visual and motor widths (to use the paper's terms) is interesting. POS +539 iclr20_1042_2_17 As an example, q(z) could be arbitrarily multimodal as far as the encoder is concerned, but the Weibull seems to force one mode per class. NA +461 iclr19_659_2_8 My main concern about the paper is the time cost. NA +576 iclr20_1493_2_36 It is unclear if what is lacking from the NN is explicit regularization, or just more data. NA +255 iclr19_1049_1_0 This work proposes a variant of the column network based on the injection of human guidance. NA +979 midl19_51_2_19 6- I suggest the authors to use train validation and test split or a cross-validation, since the results presented here are from a validation set without a test set. NEG +1185 midl20_90_2_6 If this is true, the model is not efficient. NEG +1006 midl19_52_2_25 I believe the computational time can be added for each method in Table 1. NEG +560 iclr20_1493_2_18 And it seems to have very relevant connections to your work. NA +208 graph20_56_1_15 IN fact, the whole way the user draws the shape is poorly described. NEG +910 midl19_49_1_5 The introduction part is a little misleading for me. NEG +1 graph20_25_2_1 The prototype is a simplified version of Bend Passwords [33] geared towards users who are visually impaired. NA +269 iclr19_1091_1_7 Other than that, the different approaches tested all work well in different tasks. POS +265 iclr19_1091_1_3 However, the conclusions do not directly follow from the results, so should be made more precise. NEG +914 midl19_49_1_9 The experiments shown in Table 1 compare several different network settings. NA +230 graph20_61_2_8 INTRODUCTION The motivation and context is sound, with references on how information visualization and dashboards support learning analytics or educational data visualization. POS +1375 neuroai19_59_3_0 The question of how networks maintain memory over long timescales is a longstanding and important one, and to my knowledge this question hasn't been thoroughly explored in spiking, trained recurrent neural networks (RNN). NA +1119 midl20_135_3_5 It is strange that the T1, T2 generalize well to the validation set but not to the test. NEG +1370 neuroai19_54_3_9 The flow/high-level organization of the paper works well. POS +864 midl19_14_2_38 ii) Learning to leverage the information of complementary data sets is a challenging task. NA +1066 midl20_100_1_17 What is interesting is not who is better, but how, and how well, the task can be solved. NEG +873 midl19_25_3_3 Only a single (large) dataset is used, while there are many publicly available datasets that could be included for additional experiments. NEG +1246 neuroai19_26_1_2 While it does not seemingly add anything conceptual, the exact implementation is arguably new. NEG +398 iclr19_304_3_11 You mention complexity of data and model several times in the paper but never define what you mean by that. NEG +1271 neuroai19_3_3_6 The work has promising implications for computational psychiatry, but probably not for RL at this point. POS +45 graph20_29_3_13 " That, in turn, makes it quite difficult to understand the counter-argument developed in this paper---and especially since ""The evidence comes from a study by Bi et al."" (p. 4), which makes one wonder why Bi et al. put that ""limitation"" up in the first place." NEG +161 graph20_43_1_8 " In addition to the above concerns about the contribution of the paper, the term ""motor size"" is already used in Blanch et al.'s CHI 2004 work to refer to the situation where the control-display gain is manipulated to create objects with a higher or lower size in motor space as compared to their visual space on screen, work which is not cited in this paper." NEG +991 midl19_52_2_10 They seem to work in different dimensions of the signals. NA +727 iclr20_720_2_0 While this paper has some interesting experiments. POS +629 iclr20_2094_1_3 Unfortunately, I believe that the current manuscript is at a too early stage for being accepted at ICLR, due to the following reasons: (a) The paper is littered with spelling/grammar mistakes (just take the second sentence: With the developing -> development). NEG +288 iclr19_1291_3_1 2) Individualized reward: Agents receive individual rewards; therefore, agents are aware of their contribution towards the goal. NA +1059 midl20_100_1_10 In the methods section you describe training an autoencoder on unlabeled data, then training an LSTM using autoencoder embeding and embryologist grades. NA +1359 neuroai19_53_1_9 It also would have been nice to comment on the relationship of this work to unsupervised (e.g. Hebbian-based) learning rules. NEG +575 iclr20_1493_2_35 This concern does not make the contribution of the symmetric dataset less valuable, but a discussion of such caveats would help further elucidate the similarities and differences of this setup from real datasets. NEG +850 midl19_14_2_24 The authors of [2] provided a website with all the results on the DRIVE database (pseudo-url), so their segmentations could be taken from there. NA +911 midl19_49_1_6 The authors emphasize that the objective is to cluster the geometric shape of leaflets, and it is hard to represent the shapes in high-dimensional space (last paragraph of introduction). NA +1254 neuroai19_26_1_10 I'd say a fairly 'standard' work for the setting. POS +1264 neuroai19_29_1_7 " Its also critical to understanding the function of the hippocampus and entorhinal cortex in humans.""" NA +932 midl19_51_1_0 The paper presents an approach to aid interpretation of pathology images coming from confocal microscopes (CM images). NA +1100 midl20_119_2_1 It's shown that such self-expressiveness constraint can help to preserve subtle structures during image translation, which is critical for medical tasks, such as plaque detection. NA +1207 midl20_96_3_18 I have been working in this field for many years and published papers about these topics. NA +247 graph20_61_2_25 The responsive design choice is great for multiple device access with various form factors. POS +476 iclr19_866_1_9 This is in contrast to semantic parsing and symbol grounding models, which exploit the compositionality of language to generalize to new instructions. NA +85 graph20_35_1_11 " This research is well-written and a good contribution to the area of brainstorming, and it would be interesting to get more of a complete sense of the results.""" POS +469 iclr19_866_1_2 The second module is responsible for mapping goals from this embedding space to control policies. NA +225 graph20_61_2_3 Originality The review of related work is varied across relative disciplines and well positioned. POS +774 iclr20_855_3_10 The second has to do with the interpretation of the results. NA +331 iclr19_242_2_6 It looks to me the better generalization comes from more complicated data augmentation, not from the proposed large batch training. NEG +233 graph20_61_2_11 RELATED WORK The related work is well balanced with a review on visualization dashboards and visualization in medical training with references from diverse related research communities. POS +588 iclr20_1724_2_5 The compositional action classification task is harder and shows that incorporating LSTMs for temporal reasoning leads to non-trivial performance improvements over frame averaging. NA +475 iclr19_866_1_8 The goal-policy mapping approach would presumably restrict the robot to goals experienced during training, preventing generalization to new goals. NEG +599 iclr20_2046_2_4 And it combines A* search with MCTS to improve the performance over the traditional MCTS approaches based on UCT or PUCT tree policies. POS +421 iclr19_304_3_35 Instead, you present vague of sharp drops and two modes but do not present rigorous definitions. NEG +182 graph20_53_2_9 The latter assessed usability (SUS) and workload (NASA TLX) and custom miscellaneous items. NA +988 midl19_52_2_7 2- The extensive tests on a real dataset instead of phantom cases is definitely a strength of the paper. POS +1000 midl19_52_2_19 Are the results on the entire parametric maps in line with the current results? NEG +49 graph20_29_3_17 The described examples feel rather artificial. NEG +1305 neuroai19_34_2_7 Mostly neuroscientific, but addresses the important topic of how models from machine learning can best be used in neuro research. POS +1029 midl19_56_3_18 " Authors suggest that their solution ""is not constrained by statistical modes of variation"", as e.g. by PCA-based SSM methods." NA +529 iclr20_1042_2_7 " Text contradicting the equation: ""In order to balance the individual loss terms, we normalize according to dimensions and weight the KL divergence with a constant of 0.1""." NEG +1258 neuroai19_29_1_1 The work of Hill et al. (2019) very clearly addresses these questions by devising tasks that require generalization across domains, showing how training regime is sufficient to overcome the difficulties of these tasks, even in shallow networks. NA +486 iclr19_866_1_19 The domains considered for experimental evaluation are particularly simple. NEG +497 iclr19_938_3_2 Authors compare their approach with COMA (discrete actions and counterfactual (semi-centralized) baseline) and MADDPG (also uses centralized value function and continuous actions) MAAC is evaluated on two 2d cooperative environments, Treasure Collection and Rover Tower. NA +106 graph20_36_1_20 The experiment procedure give little details about participants background. NEG +443 iclr19_495_1_15 Maybe they can uniformly outperform Gaussian policy? NA +646 iclr20_2094_1_20 But this looks wrong since it should include the distribution of all instances of 3D-BPP. NA +1036 midl19_59_3_4 " Often new methods are manually ""overfitted"" to one dataset." NA +726 iclr20_57_3_10 " Table 3 (text classification result) does not list baselines.""" NEG +190 graph20_53_2_17 The paper does discuss limitations, but I think that this section should also address the fact that the study was largely preliminary / exploratory in nature; there was no comparison condition, nor a discussion of what a baseline condition might look like for this context. NEG +282 iclr19_1091_1_20 The GTC metric is not very well established (yet). NEG +29 graph20_26_3_4 I am satisfied with the changes in the modified manuscript, and changing am my recommendation to accept. POS +169 graph20_45_2_4 The results are demonstrated on several example datasets and contrasted against visualizations using traditional PCP and scatterplots. NA +1001 midl19_52_2_20 7- What is the number of parameters required for each method in Table 1? NEG +1204 midl20_96_3_15 Assessing in-focus will even get rid of blurred frames and frames as discussed in the Appendix. NA +100 graph20_36_1_14 Are there other patters with features not presented in these three? NEG +197 graph20_56_1_4 However, the paper is weakened by several writing and organizational aspects, and by an odd off-hand report of user feedback. NEG +495 iclr19_938_3_0 Summary Authors present a decentralized policy, centralized value function approach (MAAC) to multi-agent learning. NA +647 iclr20_2094_1_21 e) The Actor-Critic framework, coupled with a conditional query learning algorithm, is unfortunately unintelligible due to the fact that many notations are left unspecified. NEG +987 midl19_52_2_6 1- This paper is well written and the message is clear to the reader. POS +44 graph20_29_3_12 " UNLIMITING"" I found it quite hard to understand the point of Bi et al. for rejecting screen-to-screen pointing, at least the way it is explained in this paper." NEG +710 iclr20_526_3_31 Here again, MNIST would be a useful dataset. NEG +373 iclr19_261_3_9 Have you tried baselines like these? NA +678 iclr20_305_3_14 Make it clear that these points are here to help, and not necessarily part of your decision assessment. NA +566 iclr20_1493_2_25 It is unclear on what basis one can say that real-world datasets are more like the symmetric case or the asymmetric case. NEG +847 midl19_14_2_21 The method in [2] should be included in the comparison of vessel segmentation algorithms. NEG +460 iclr19_659_2_7 The main contribution is it provides a way to reduce the number of interactions with the environment. NA +1155 midl20_77_4_2 On the positive side, the extension of the Gessert model to motion forecasting seems like a useful one. POS +39 graph20_29_3_7 " As the authors state themselves p. 9, ""A common way to check external validity is to apply obtained parameters to data from different participants.""" NA +434 iclr19_495_1_6 and why one needs to compute gradients of the entropy (Section 4.1)? NA +973 midl19_51_2_13 The authors should provide support to these conclusions. NEG +1239 neuroai19_23_1_8 What precisely about predictive coding makes the similarity to brain data expected? NEG +1007 midl19_52_2_26 Minor suggestions a- Some recent work on using the complex-valued neural networks (Virtue Patrick et al., arxiv), geometry of deep learning (Golbabaee et al., arxiv)and recurrent neural networks (Oksuz et al.,arxiv) for MRF dictionary matching can be mentioned in the literature review with their strengths and weakneses. NA +456 iclr19_659_2_3 This paper is well-written. POS +260 iclr19_1049_1_5 Experiments have shown that the convergence speed and results are improved, but not significant. NEG +35 graph20_29_3_3 I also have a number of concerns that I would like to see addressed in a revision. NEG +827 midl19_14_2_1 The method allows to detect a series of relevant anatomical/pathological structures in fundus pictures (such as the retinal vessels, the optic disc, hemorrhages, microaneurysms and soft/hard exudates). NA +309 iclr19_1399_1_2 Hyperparameters were honestly optimized. POS +276 iclr19_1091_1_14 Because the parts of the state that are needed for multiple different prediction tasks (reconstruction, inverse model, etc.) need to be in the final state representation multiple times. NA +636 iclr20_2094_1_10 thesis 2015; Christensen et. al. Computer Science Review 2017). NA +654 iclr20_2157_3_0 The paper presents expected gradients which is a method which looks at a difference from a baseline defined by the training data. NA +622 iclr20_2046_2_27 In practice, this is not true because even at the leaf node the value could still be estimated by an inaccurate value network (e.g., AlphaGo or AlphaZero). NEG +844 midl19_14_2_18 I would suggest to incorporate results on at least one of these data sets to better understand the behavior of the algorithm on these images. NA +393 iclr19_304_3_6 Because of that, the experimental evaluation remains vague as well, as the criteria are tested on one data set by visual inspection. NEG +107 graph20_36_1_21 How did authors ensure homogeneity of the groups? NEG +349 iclr19_242_2_26 In section 4.2, I fail to understand why the proposed method can affect the norm of gradient. NEG +524 iclr20_1042_2_2 In this way, there's no need to store all past data and even the first learned batch keeps being refreshed and should not be forgotten. NA +784 iclr20_880_2_4 This paper can have a tremendous impact in the research in deep networks if results are well explained. POS +47 graph20_29_3_15 12), - and for some reason that makes it ok to consider that screen-to-screen pointing is compatible with Bi et al.'s model (which does not consider A). NA +203 graph20_56_1_10 Overall, the writing and the organization of the paper suffered from similar issues. NEG +21 graph20_25_2_21 REFERENCES Sauvik Das, Gierad Laput, Chris Harrison, and Jason I. Hong. NA +402 iclr19_304_3_15 Independent and identically distributed? NA +1107 midl20_127_4_2 " Main problem: The authors mention ""the AF are sonographer dependent, and its accuracy depends on the sonographer's experience." NEG +1154 midl20_77_4_1 The models are variants of that proposed in Gessert et al (2019), which is here extended in different ways to perform motion forecasting/prediction using a sequence of OCT volumes, rather than motion estimation between 2 OCT volumes. NA +989 midl19_52_2_8 3- The description of the network architecture is not clear for the reader. NEG +741 iclr20_720_2_14 While the experiments show the value of hierarchy, they do not show the value of this particular method of creating hierarchy. NEG +1310 neuroai19_34_2_12 It doesn't provide any additional information to the data lines themselves, and it leads the reader to expect these indicate statistically significant comparisons. NEG +1147 midl20_71_1_4 Meanwhile, a few clarifications may be necessary: 1) in term of runtime, does the addition of GRUs take much more training time and memory comparing to the concatenation of 3D volumes? NEG +33 graph20_29_3_1 This paper is well written and shows good experiment design and consistent analyses. POS +787 iclr20_880_2_7 Equation (1) and (2) are extremely clear. POS +923 midl19_49_1_19 and Hausdorff distance to measure the recon accuracy between original image and reconstructed image. NA +1043 midl19_59_3_11 But the higher performance is not significant. NEG +1345 neuroai19_37_3_24 " I feel this statement: ""Our challenge is to understand how this occurs." NA +1243 neuroai19_23_1_12 " Typo line 24 Moreover, we show that as (we) train the model Typo line 87 Second, the model does not rely on labeled data and learn(s)""" NA +579 iclr20_1493_2_39 It would be very interesting to see whether these results differ at all from the one-shot approach here. NA +481 iclr19_866_1_14 Are they free-form instructions? NEG +931 midl19_49_1_27 " All architectures listed in Table 1 should be stated clearly in experiments section not only in method section. """ NEG +356 iclr19_242_2_34 Moreover, the proposed method also use N times more augmented samples. NA +1248 neuroai19_26_1_4 I think a more persuasive bench marking could be done. NEG +759 iclr20_76_2_0 In order to rationalize the existence of non-trivial exponents that can be independent of the specific kernel used, this paper introduces the Teacher-Student framework for kernels. NA +92 graph20_36_1_6 To begin with, there is little details about the design rationale. NEG +1175 midl20_85_3_8 What is the experimental setup? NEG +1341 neuroai19_37_3_20 Repeat this process recursively tens to trillions of times, and suddenly you have a brain controlling a body in the world or doing something else equally clever. NA +803 iclr20_880_2_23 " If the authors can reject (1), (2) and (3), they should find a plausible explaination why performance improves in their experiments.""" NEG +745 iclr20_727_1_0 The authors propose a method for learning models for discrete events happening in continuous time by modelling the process as a temporal point process. NA +279 iclr19_1091_1_17 Please indicate why these tasks are chosen. NA +764 iclr20_855_3_0 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. NA +604 iclr20_2046_2_9 How could it improve over the traditional tree policy (e.g., UCT) for the selection step in MCTS? NEG +1383 neuroai19_59_3_8 For instance, while the heatmaps in Figure 3 provide visual evidence for their claims (except see my comments below), the work could have benefitted from a quantification of this evidence. NEG +400 iclr19_304_3_13 Page 4, Assumption. NA +257 iclr19_1049_1_2 Human knowledge is embodied in a defined rule formula. NA +944 midl19_51_1_12 I understand that the available space is limited and therefore it's difficult to bring in the paper all the information that would be necessary, but the introduction should be extended to include previous work both in terms of DL and medical research. NEG +1096 midl20_108_3_12 Moreover, is there is a reason you did not validate on all TUPAC16 tasks?The is well written paper with a clear description of the state of the art and the reasoning behind the presented method. POS +1219 neuroai19_2_2_9 While much human learning may be more naturally cast as online learning, not all of it is. NA +1161 midl20_77_4_8 Also, can the authors comment on what the accuracy requirement is for motion tracking in OCT? NA +494 iclr19_866_1_27 " There are several grammatical errors - The captions for Figures 3 and 4 are copied from Figure 1.""" NEG +772 iclr20_855_3_8 Unless a comparison can be made with the same amounts of experience, I don't see how Figure 2 can be interpreted objectively. NEG +550 iclr20_1493_2_7 However, there are a few (in my opinion) critical concerns that currently bar me from strongly recommending acceptance of the paper. NA +201 graph20_56_1_8 One main weakness of the paper is manifested here: I found the description of the bins, and how they are calculated, quite confusing. NEG +312 iclr19_1399_1_5 Weaknesses: - The experiments are done on CIFAR-10, CIFAR-100 and subsets of CIFAR-100. NA +290 iclr19_1291_3_3 The authors also claim these two extensions enable their model to converge faster and better. NA +1213 neuroai19_2_2_3 Either way this is important work, with many interesting future directions. POS +763 iclr20_76_2_4 " Therefore, the efficacy of the proposed model can not be well demontrated.""" NEG +1136 midl20_56_4_18 But the writing needs to be improved. NEG +949 midl19_51_1_17 Unfortunately the authors didn't report indications in this sense in their paper. NEG +183 graph20_53_2_10 Some issues in the study reporting: - What was the scale range for the prior experience questions? NEG +1176 midl20_85_3_9 Did you train on some other dataset and test on skin lesion dataset? NEG +1349 neuroai19_37_3_28 " While it covers important ground, I think the arguments need more refinement and focus before they can inspire productive discussion.""" NEG +1068 midl20_100_1_19 I am not convinced. NEG +1030 midl19_56_3_19 " While I agree that the linear latent space assumption of PCA is too simplistic and the global effect of PCA latents on the whole shape often undesirable, the ordering of latents according to ""percent of variance explained"" is actually desirable in terms of interpretability." NEG +250 graph20_61_2_28 " I would suggest the following references to inform analysis of user logs: - H. Guo, S. R. Gomez, C. Ziemkiewicz and D. H. Laidlaw, ""A Case Study Using Visualization Interaction Logs and Insight Metrics to Understand How Analysts Arrive at Insights,"" in IEEE Transactions on Visualization and Computer Graphics, vol. 22, no. 1, pp. 51-60, 31 Jan." NEG +823 midl19_13_2_8 A more complete evaluation with different surgical scenarios would be needed to demonstrate this feature. NA +1099 midl20_119_2_0 This paper proposes to add a self-expressiveness regularization term to learn a union of subspaces for image-to-image translation in medical domain. NA +940 midl19_51_1_8 I agree with the authors statement in the end of the paper where they say they could train both GAN and de-speckle network end to end. NA +391 iclr19_304_3_4 Foremost, the presented criteria are actually not real criteria (expect maybe C1) but rather general guidelines to visually inspect 'accuracy over randomization curves. NEG +1092 midl20_108_3_8 Multi-task learning can extract a shared representation that is generalisable and this is evidenced in the results in the TUPAC16 set. NA +1191 midl20_96_3_2 Motivation is based on anonymisation and data cleansing. NA +904 midl19_41_1_4 " The gain using CGAN MRI looks marginal, which would be better to apply ablation study. """ NEG +730 iclr20_720_2_3 While it is possible that I am missing something, I have tried going through the paper a few times and the contribution is not immediately obvious. NEG +554 iclr20_1493_2_12 While not in conflict with this work, it does closely relate and discuss many of the same issues discussed in this work, so relating them would be fruitful. NEG +366 iclr19_261_3_1 The authors argue convincingly that an interactive and grounded evaluation environment helps us better measure how well NLG/NLU agents actually understand and use their language rather than evaluating against arbitrary ground-truth examples of what humans say, we can evaluate the objective end-to-end performance of a system in a well-specified nonlinguistic task. NA +345 iclr19_242_2_21 For theorem 1, it is hard to say how much the theoretical analysis based on linear approximation near global minimizer would help understand the behavior of SGD. NA +535 iclr20_1042_2_13 The weighting of the KL that the authors introduce is going to bias the learned generator towards the high probability regions. NA +1209 midl20_96_3_20 " I am advocating open data access and reproducible research.""" NA +1372 neuroai19_54_3_11 e.g. what was the nonlinearity used in the model CNN? NEG +919 midl19_49_1_15 However, recon accuracy highly depends on decoder network. NA +211 graph20_56_1_18 These critical areas of confusion around how the process actually unfolds from start to finish should have been more clearly described. NEG +502 iclr19_938_3_7 The centralized nature is also semantically improbable, as the observations might be high-dimensional in nature, so exchanging these between agents becomes impractical with complex problems. NEG +537 iclr20_1042_2_15 A Weibull distribution is used to model the same data, again, in a different way. NA +1220 neuroai19_2_2_10 There may be much interest in how we learn from so few samples in certain settings, but we also learn some relationships/tasks in a classical associationist manner which is well modeled by 'slow' gradient-descent like learning (e.g. Rescorla Wagner). NA +175 graph20_53_2_2 With this, users can select part of a VR object, assign an animation behaviour, and preview it. NA