Datasets:

boda
/

review_evaluation_main_10

paper_id stringlengths 10 19	venue stringclasses 15 values	focused_review stringlengths 7 9.67k	point stringlengths 55 634
NIPS_2019_899	NIPS_2019	Weakness: - Latent language seems to add a more complex intermediate problem. You are now introducing text understanding which might be a harder problem. Do we really need text? why not use a program for guidance? Surely, a program is more expressive than macro-action and interpretable and you dont have language understanding challenges. Maybe a clever data collection strategy can collect programs. - The problem is solved using supervised learning without any exploration based learning. This makes me wonder how easy the setup is. Did you try comparing against agents that are trained without text but use reinforcement learning? The environment must have some reward (score, number of enemies killed etc.). Of course, once you consider exploration it is not clear how accurate your instruction model would be. Maybe this is a limitation of this direction? Questions and Other Comments - Why would any instruction be repeated at all? I am trying to understand the purpose of using one-hot vector encoding for instructions. How many instructions occur more than once? - Did you evaluate against rule-based bots? How is the performance against another model trained using the same strategy (i.e. self play). - I believe authors first generate a text using the instruction model and then re-encode the text using an encoder. Why not do something like this: create an instruction embedding g(f(s)) from state encoding f(s). Pass this instruction encoding directly to the executor (as opposed to generated text). Then you can add an auxiliary objective which will try to bring g(f(s)) closer to the gold instruction encoding. This might work better as you are not adding discretization in between which can fail due to a single wrong decoding (say as based on a tie). - Equation on line 214 should have different state s for each time "i". Otherwise your state representation is not changing while you take new actions. - Many missing citations (see Tellex et al., AAAI 2011, Tellex et al., RSS 2014, Chaplot et al., AAAI 2017, Bahdanau et al., ICLR 2017, Misra et al., EMNLP 2018, Mirowski et al., 2019, Chen et al., CVPR 2019) etc.	- Many missing citations (see Tellex et al., AAAI 2011, Tellex et al., RSS 2014, Chaplot et al., AAAI 2017, Bahdanau et al., ICLR 2017, Misra et al., EMNLP 2018, Mirowski et al., 2019, Chen et al., CVPR 2019) etc.
NIPS_2018_710	NIPS_2018	- My general reservation about this paper is that while it was helpful in clarifying my own understanding of BN, a lot of the conclusions are consistent with folk wisdom understanding of BN (e.g. well-conditioned optimization), and the experimental results were not particularly surprising. Questions: - Taking Sharp Minima Hypothesis at face value, Eq 2 suggests that increasing gradient variance improves generalization. This is consistent with the theme that decreasing LR or decreasing minibatch size make generalization worse. Can you comment on how to reconcile this claim with the body of work in black-box optimization (REBAR, RELAX, VIMCO, Reinforcement Learning) suggesting that reducing variance of gradient estimation improves generalization & final performance? - Let's suppose that higher SGD variance (eq 2) == better generalization. BN decreases intra-unit gradient variance (Fig 2, left) but increases intra-minibatch variance (Fig 4, right). When it is applied to a network that converges for some pair \alpha and B, it seems to generalize slightly worse (Fig 1, right). According to the explanations presented by this paper, this would imply that BN decreased M slightly. For what unnormalized architectures does the application of BN increase SGD variance, and for what unnormalized architectures does BN actually decrease SGD variance? (requiring LR to be increased to compensate?) How do inter-layer gradient variance and inter-minibatch gradient variance impact on generalization? - For an unnormalized network, is it possible to converge AND generalize well by simply using a small learning rate with a small batch size? Does this perform comparably to using batch norm? - While Section 4 was well-written, the argument that BN decreases exponential condition number is not new; this situation has been analyzed in the case of training deep networks using orthogonal weights (https://arxiv.org/pdf/1511.06464.pdf, https://arxiv.org/pdf/1806.05393.pdf), and the exploding / vanishing gradient problem (Hochreiter et al.). On the subject of novelty, does this paper make a stronger claim than existing folk wisdom that BN makes optimization well-conditioned?	- My general reservation about this paper is that while it was helpful in clarifying my own understanding of BN, a lot of the conclusions are consistent with folk wisdom understanding of BN (e.g. well-conditioned optimization), and the experimental results were not particularly surprising. Questions:
dapU3n7yfp	ICLR_2024	- Lack of baselines: The attack algorithm is based on HotFlip (2018), which is a bit old and less effective than the recently proposed baselines. I am wondering if the authors have compared with adversarial attack baselines proposed more recently such as Seq2sick [1], which shows better optimization effectiveness for text-to-text generation tasks. - Lack of defense models (detoxified models): While I appreciate the authors’ efforts in comparing different pretrained models, it would also be interesting to evaluate against different defense approaches/detoxification approaches, such as [2,3,4], and confirm whether the attack is still effective. - Validity of toxicity evaluation: The authors mention that their evaluation setup “can be applied to bridge the evaluation of toxicity in different PLMs”, and “speculate that the success rate of ASRA attack might be positively correlated with the toxicity of language models.” However, I do not see any evidence about whether the ASR here can be a good proxy to reflect model toxicity. Given that the model toxicity evaluation is conducted by evaluating model responses with a lot of different inputs and contexts, the setup of this paper is to evaluate model responses given “unnatural” prompts. I am thus suspicious about whether the test above can give an accurate evaluation of model toxicity. - There is an important concern regarding the potential misuse of this work by malicious users to bypass safety controls of LLMs and elicit model toxicity. I believe it would be valuable for the paper to include a discussion of this aspect. [1] Seq2Sick: Evaluating the Robustness of Sequence-to-Sequence Models with Adversarial Examples. (AAAI 2020) [2] Plug and Play Language Models: A Simple Approach to Controlled Text Generation (ICLR 2020) [3] DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts (ACL 2021) [4] Exploring the Limits of Domain-Adaptive Training for Detoxifying Large-Scale Language Models (NeurIPS 2022)	- Lack of defense models (detoxified models): While I appreciate the authors’ efforts in comparing different pretrained models, it would also be interesting to evaluate against different defense approaches/detoxification approaches, such as [2,3,4], and confirm whether the attack is still effective.
ICLR_2023_3305	ICLR_2023	1.The review of related work on uncertainty in meta-learning is small and it is difficult to locate the main contribution of this paper. 2.The Ood detection only extends from classification to regression, which is not very innovative. 3.The experiments only focus on the comparison between various models proposed by authors and lack of comparison with other methods.	2.The Ood detection only extends from classification to regression, which is not very innovative.
NIPS_2020_1507	NIPS_2020	1. The proposed method is based on a pre-defined causal graph, which has limitations if the causal graph is unavailable. In the experimental results sections, the authors only showed the results with the graph constructed by the PC algorithm. It is not clear how the way of graph construction affects the final results. 2. The optimization details for the objective function is missed. 3. This paper lacks the complexity analysis for the proposed method. 4. Only validating the proposed method on one real dataset (i.e., Adult) cannot guarantee its applicability in the wide spectrum of real-world applications. 5. The assumption about the mutual independency of the exogenous variables are too strong to be satisfied in real-world applications.	1. The proposed method is based on a pre-defined causal graph, which has limitations if the causal graph is unavailable. In the experimental results sections, the authors only showed the results with the graph constructed by the PC algorithm. It is not clear how the way of graph construction affects the final results.
oIwoBDsJJI	ICLR_2024	1. The graph Foster distance is a direct application of the optimal transport problem on the graph Foster distributions. 2. Compared with the Fused Gromov-Wasserstein Distance (FGW), the improvement in the computation time and the classification accuracy for the graph Foster distance in the experiments is very marginal.	2. Compared with the Fused Gromov-Wasserstein Distance (FGW), the improvement in the computation time and the classification accuracy for the graph Foster distance in the experiments is very marginal.
3LdaPmAnji	EMNLP_2023	1. Experimental results do not show the effectiveness of fine-grained classes. 2. Pherpas lacks Inter-annotator Agreement to better demonstrate the quality of EDeR.	1. Experimental results do not show the effectiveness of fine-grained classes.
NIPS_2017_349	NIPS_2017	- The paper is not self contained Understandable given the NIPS format, but the supplementary is necessary to understand large parts of the main paper and allow reproducibility. I also hereby request the authors to release the source code of their experiments to allow reproduction of their results. - Use of deep-reinforcement learning is not well motivated The problem domain seems simple enough that a linear approximation would have likely sufficed? The network is fairly small and isn't "deep" either. - > We argue that such a mechanism is more realistic because it has an effect within the game itself, not just on the scores This is probably the most unclear part. It's not clear to me why the paper considers one to be more realistic than the other rather than just modeling different incentives? Probably not enough space in the paper but actual comparison of learning dynamics when the opportunity costs are modeled as penalties instead. As economists say: incentives matter. However, if the intention was to explicitly avoid such explicit incentives, as they _would_ affect the model-free reinforcement learning algorithm, then those reasons should be clearly stated. - Unclear whether bringing connections to human cognition makes sense As the authors themselves state that the problem is fairly reductionist and does not allow for mechanisms like bargaining and negotiation that humans use, it's unclear what the authors mean by ``Perhaps the interaction between cognitively basic adaptation mechanisms and the structure of the CPR itself has more of an effect on whether self-organization will fail or succeed than previously appreciated.'' It would be fairly surprising if any behavioral economist trying to study this problem would ignore either of these things and needs more citation for comparison against "previously appreciated". * Minor comments Line 16: > [18] found them... Consider using \citeauthor{} ? Line 167: > be the N -th agentâs should be i-th agent? Figure 3: Clarify what the `fillcolor` implies and how many runs were the results averaged over? Figure 4: Is not self contained and refers to Fig. 6 which is in the supplementary. The figure is understandably large and hard to fit in the main paper, but at least consider clarifying that it's in the supplementary (as you have clarified for other figures from the supplementary mentioned in the main paper). Figure 5: - Consider increasing the axes margins? Markers at 0 and 12 are cut off. - Increase space between the main caption and sub-caption. Line 299: From Fig 5b, it's not clear that \|R\|=7 is the maximum. To my eyes, 6 seems higher.	- Consider increasing the axes margins? Markers at 0 and 12 are cut off.
ACL_2017_333_review	ACL_2017	The criticisms are very minor: - It would be best to report ROUGE F-Score for all three datasets. The reasons for reporting recall on one are understandable (the summaries are all the same length), but in that case you could simply report both recall and F-Score. - The Related Work should come earlier in the paper. - The paper could use some discussion of the context of the work, e.g. how the summaries / compressions are intended to be used, or why they are needed. - General Discussion: - ROUGE is fine for this paper, but ultimately you would want human evaluations of these compressions, e.g. on readability and coherence metrics, or an extrinsic evaluation.	- It would be best to report ROUGE F-Score for all three datasets. The reasons for reporting recall on one are understandable (the summaries are all the same length), but in that case you could simply report both recall and F-Score.
NIPS_2016_208	NIPS_2016	1. The novelty is a little weak. It is not clear what's the significant difference and advantage compared to NCA [6] and "Small codes and large image databases for recognition", A. Torralba et al., 2008, which used NCA in deep learning. 2. In the experiment of face recognition, some state-of-the art references are missing, such as Baidu' work "Targeting Ultimate Accuracy: Face Recognition via Deep Embedding", http://vis-www.cs.umass.edu/lfw/results.html#baidu. In that work, the triplet loss is also used and it reported the result trained on the dataset containing 9K identities and 450K images, which is similar with Webface. The VRF can achieve 98.65% on LFW which is better than the result in Table 3 in this paper.	1. The novelty is a little weak. It is not clear what's the significant difference and advantage compared to NCA [6] and "Small codes and large image databases for recognition", A. Torralba et al., 2008, which used NCA in deep learning.
NIPS_2021_1872	NIPS_2021	Weakness: 1. The background on linear GCNs may not be clear. Why do we need to study linear GCN? Most of current models are non-linear. 2. The difference between with self loop and without self loop? 3. The authors used the step-size to be T/K. Why exactly T/K? can we use larger or smaller than T/K?	2. The difference between with self loop and without self loop?
NIPS_2021_1907	NIPS_2021	There is little improvement empirically. Furthermore, it is unclear if the gains in this paper are due solely to the confidence widths or if the design of the algorithm is important too. For the empirical study, it is unclear how the other experiments would perform if they had access to the same confidence widths presented in this work. This may make the algorithmic comparison fairer since the differences in performance would be solely due to the sampling procedures. Also, (and I am torn on this since the setup is nice and clear) it is worth noting that the authors are most of the way through page 5 before any results are presented. Other comments and questions: - Does theorem 1 hold for an adaptive sequence of x_n’s or a fixed sequence? The theorem just seems to specify a set of (x,y)’s that have been collected. Ie, is this a truly anytime result or for a fixed sequence? In the case of a linear kernel, the gap in the confidence widths between an anytime and fixed confidence bound is O(\sqrt(d)) which behaves like O(sqrt(\gamma_n)) in that setting. I guess that the algorithm is using these as an adaptive sequence which is maybe okay from a Bayesian perspective. - Same question for Thm 2 - For the result in remark 2, do other works get the same factor of d since log(N^d) = dlog(N)? This work is tighter in terms of \sqrt(\gamma) but is the d dependence the same? - Why is MVR the right sampling objective? - Regarding the statement in Section 6 about simple and cumulative regret bounds, it is somewhat expected that the cumulative regret is linear if you do this well on simple regret as your objective is largely one of exploration. Take for example the SE kernel as the variance \sigma -> 0. In this setting, we recover standard multiarmed bandits where http://sbubeck.com/ALT09_BMS.pdf for instance show that there cannot be an algorithm that is simultaneously optimal in both simple and cumulative regret. Minor comments: - Make sure that the colors chosen for the plots are colorblind friendly. There are a variety of palettes in python for this. - Some of the axes in the plots in the main body and especially Appendix G are hard to read. The authors do a good job discussing the limitations of their work, though more consideration should be given to potential negative societal impacts than simply saying “our work is theoretical, therefore we can do no wrong.”	- Does theorem 1 hold for an adaptive sequence of x_n’s or a fixed sequence? The theorem just seems to specify a set of (x,y)’s that have been collected. Ie, is this a truly anytime result or for a fixed sequence? In the case of a linear kernel, the gap in the confidence widths between an anytime and fixed confidence bound is O(\sqrt(d)) which behaves like O(sqrt(\gamma_n)) in that setting. I guess that the algorithm is using these as an adaptive sequence which is maybe okay from a Bayesian perspective.
NIPS_2020_773	NIPS_2020	* It's not clear how the Kalman Filtering perspective provides any new insight. Both the global query-specific prior and frequency capping are trivial to specify in the standard attention framework. The Kalman Filtering perspective seems like an unnecessary distraction from what is in reality two simple modifications to a standard attention mechanism. It's especially confusing since a Kalman filter is traditionally specified as a sequential mechanism, more similar to an RNN than a Transformer. Section 3.5 doesn't sufficiently address these questions. For example, is the difference between expectation and estimation important in this setting? I currently view the KF perspective as a negative that distracts from the core modeling changes. * There are only two benchmark results. While the improvements are statistically significant, it's unclear whether they are nontrivial improvements. The authors need to provide more context here, especially for the real-world system. For example, is a +4.4% CTRgain big or small for this system?	* There are only two benchmark results. While the improvements are statistically significant, it's unclear whether they are nontrivial improvements. The authors need to provide more context here, especially for the real-world system. For example, is a +4.4% CTRgain big or small for this system?
NIPS_2017_320	NIPS_2017	#ERROR!	- I believe in section 2 it could be made clearer when gradients are calculated by a solver and when not.
bxltAqTJe2	EMNLP_2023	1. I strongly recommend that the authors validate the quality of the GFC API output to ensure the accuracy of the ground truth. This step is crucial in ensuring the reliability of the findings. 2. Since the ground truth output can be assess on the Internet, which may lead to test data leakage problem. I encourage authors to discuss the data leakage problem and its implications in this work. 3. The novelty of the proposed CACN method appears limited, as it seems to be a direct combination of CoT and Reverse check worthiness. Additionally, the claim normalization task could be seen as a form of text augmentation or summarization. 4. The description of the baseline is not sufficiently clear. For instance, it is unclear whether the finetuning refers to full-parameter or parameter-efficient prompt tuning. Clarifying this aspect would enhance the understanding of the experimental setup. 5. It is essential to ensure fairness in the comparisons made. For example, the same method should be applied to different models, and the same base model should be evaluated with different prompt paradigms. This would provide a more comprehensive and unbiased analysis. 6. Table 4 lacks results for the proposed CACN method. It is important to include these results to provide a complete evaluation of the proposed approach under 0 and few shot setting. 7. The performance degradation in few-shot learning compared to 0-shot learning requires further explanation. I recommend that the authors examine the few-shot learning exemplars and provide additional insights in this aspect.	2. Since the ground truth output can be assess on the Internet, which may lead to test data leakage problem. I encourage authors to discuss the data leakage problem and its implications in this work.
ARR_2022_63_review	ARR_2022	1. There are some other contemporary state-of-the-art models, the authors can consider citing and including them for an extensive comparison. 2. It will be good to see some analysis and insights on different combinations of pre-training datasets introduced in Table 1. Here are some questions: 1. Since some of the sub-tasks, like dialogue state tracking, require a fixed format of the output, if the model generation is incomplete or in an incorrect format, how can we tackle this issue? 2. The dialogue multi-task pre-training introduced in this work is quite different from the original language modeling (LM) pre-training scheme of backbones like T5. Thus I was curious about why not pre-train the language backbone on the dialogue samples first with the LM scheme, then conduct the multi-task pre-training? Will this bring some further improvement? 3. It will be good to see some results and analysis on the lengthy dialogue samples. For instance, will the performance drop on the lengthy dialogues?	1. Since some of the sub-tasks, like dialogue state tracking, require a fixed format of the output, if the model generation is incomplete or in an incorrect format, how can we tackle this issue?
NIPS_2020_1491	NIPS_2020	1. The algorithms require some prior knowledge of the problem such as the number of tasks and switches, time horizon, and the full-information feedback, which is due to the binary loss. 2. I think the authors need a further survey in the contextual bandit with switching regret. Here the task index resembles the context. [Luo et al, 2018] would be close to this paper. Moreover, the idea of "meta-experts" can also be seen in [Wu et al, 2019]. [Luo et al, 2018] Luo, Haipeng, et al. "Efficient contextual bandits in non-stationary worlds." Conference On Learning Theory. 2018. [Wu et al, 2019] Wu, Yi-Shan, Po-An Wang, and Chi-Jen Lu. "Lifelong Optimization with Low Regret." The 22nd International Conference on Artificial Intelligence and Statistics. 2019.	2. I think the authors need a further survey in the contextual bandit with switching regret. Here the task index resembles the context. [Luo et al, 2018] would be close to this paper. Moreover, the idea of "meta-experts" can also be seen in [Wu et al, 2019]. [Luo et al, 2018] Luo, Haipeng, et al. "Efficient contextual bandits in non-stationary worlds." Conference On Learning Theory. 2018. [Wu et al, 2019] Wu, Yi-Shan, Po-An Wang, and Chi-Jen Lu. "Lifelong Optimization with Low Regret." The 22nd International Conference on Artificial Intelligence and Statistics. 2019.
ICLR_2022_3267	ICLR_2022	Weakness: Rigorousness: This paper is a purely theoretical work in my opinion (though it has numerical simulations). Unfortunately, I did not find the claims in the paper is rigorous enough for a theoretical work. Specifically, the paper made several strong claims without rigorous mathematical analysis. For example, in section 2, the paper tries to explain the transfer phenomenon of adversarial examples from the lens of holomorphic optimal Bayes classifier. But there is no theorem/proof or detailed analysis, making the claim not convincing at all. Another example, in section 3, the paper lists the primal and dual problem for learning the maximum margin classifier, but there is again no rigorous mathematical derivation. Although the paper states that its claims are evident from the definition, as a reader, I did not find it evident at all. To sum up, the problem of the current presentation of this paper is that I cannot tell which parts are rigorous theorems and which parts are just intuitions. Therefore, I am very concerned about the rigorousness of this paper, and I do not think it qualifies as a theoretical work. I recommend the authors to 1) state clearly which parts are rigorous math and which are just intuitions/illustrations 2) summarize all the rigorous parts into mathematical theorems, 3) then give rigorous proofs.	1) state clearly which parts are rigorous math and which are just intuitions/illustrations
NIPS_2017_53	NIPS_2017	Weakness 1. When discussing related work it is crucial to mention related work on modular networks for VQA such as [A], otherwise the introduction right now seems to paint a picture that no one does modular architectures for VQA. 2. Given that the paper uses a billinear layer to combine representations, it should mention in related work the rich line of work in VQA, starting with [B] which uses billinear pooling for learning joint question image representations. Right now the manner in which things are presented a novice reader might think this is the first application of billinear operations for question answering (based on reading till the related work section). Billinear pooling is compared to later. 3. L151: Would be interesting to have some sort of a group norm in the final part of the model (g, Fig. 1) to encourage disentanglement further. 4. It is very interesting that the approach does not use an LSTM to encode the question. This is similar to the work on a simple baseline for VQA [C] which also uses a bag of words representation. 5. () Sec. 4.2 it is not clear how the question is being used to learn an attention on the image feature since the description under Sec. 4.2 does not match with the equation in the section. Speficially the equation does not have any term for r^q which is the question representation. Would be good to clarify. Also it is not clear what \sigma means in the equation. Does it mean the sigmoid activation? If so, multiplying two sigmoid activations (with the \alpha_v computation seems to do) might be ill conditioned and numerically unstable. 6. () Is the object detection based attention being performed on the image or on some convolutional feature map V \in R^{FxWxH}? Would be good to clarify. Is some sort of rescaling done based on the receptive field to figure out which image regions belong correspond to which spatial locations in the feature map? 7. () L254: Trimming the questions after the first 10 seems like an odd design choice, especially since the question model is just a bag of words (so it is not expensive to encode longer sequences). 8. L290: it would be good to clarify how the implemented billinear layer is different from other approaches which do billinear pooling. Is the major difference the dimensionality of embeddings? How is the billinear layer swapped out with the hadarmard product and MCB approaches? Is the compression of the representations using Equation. (3) still done in this case? Minor Points: - L122: Assuming that we are multiplying in equation (1) by a dense projection matrix, it is unclear how the resulting matrix is expected to be sparse (arenât we mutliplying by a nicely-conditioned matrix to make sure everything is dense?). - Likewise, unclear why the attended image should be sparse. I can see this would happen if we did attention after the ReLU but if sparsity is an issue why not do it after the ReLU? Perliminary Evaluation The paper is a really nice contribution towards leveraging traditional vision tasks for visual question answering. Major points and clarifications for the rebuttal are marked with a (). [A] Andreas, Jacob, Marcus Rohrbach, Trevor Darrell, and Dan Klein. 2015. âNeural Module Networks.â arXiv [cs.CV]. arXiv. http://arxiv.org/abs/1511.02799. [B] Fukui, Akira, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. 2016. âMultimodal Compact Bilinear Pooling for Visual Question Answering and Visual Grounding.â arXiv [cs.CV]. arXiv. http://arxiv.org/abs/1606.01847. [C] Zhou, Bolei, Yuandong Tian, Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. 2015. âSimple Baseline for Visual Question Answering.â arXiv [cs.CV]. arXiv. http://arxiv.org/abs/1512.02167.	2. Given that the paper uses a billinear layer to combine representations, it should mention in related work the rich line of work in VQA, starting with [B] which uses billinear pooling for learning joint question image representations. Right now the manner in which things are presented a novice reader might think this is the first application of billinear operations for question answering (based on reading till the related work section). Billinear pooling is compared to later.
NIPS_2019_634	NIPS_2019	see section 5 ("improvements") below. Originality: while the methods are not particularly novel (autoregressive and masked language modelling pretraining have both been used before for ELMo and BERT; this work extends these objectives to the multi-lingual case), the performance gains on all four tasks are still very impressive. - Quality: This paper's contributions are mostly empirical. The empirical results are strong, and the methodology is sound and explained in sufficient technical details. - Clarity: The paper is well-written, makes the connections with the relevant earlier work, and includes important details that can facilitate reproducibility (e.g. the learning rate, number of layers, etc.). - Significance: The empirical results constitute a new state of the art and are important to drive progress in the field. ---------- Update after authors' response: the response clearly addressed most of my concerns. I look forward to the addition of supervised MT experiments on other languages (beyond the relatively small Romanian-English dataset) on subsequent versions of the paper. I maintain my initial assessment that this is a strong submission with impressive empirical results, which would be useful for the community. I maintain my final recommendation of "8".	- Quality: This paper's contributions are mostly empirical. The empirical results are strong, and the methodology is sound and explained in sufficient technical details.
NIPS_2022_768	NIPS_2022	. W1: The presentation can be improved. There is no overview of the approach to explain the components, and a few components and concepts appear without much prior context. For example, "encoder" appears without where it is exactly being used. Same for "topic aggregation". "count vector" was used only once without definition. W2: This paper does not provide the related work description for the existing ETM this work is based on in the related work section. It also doesn't distinguish that in the method section. Thus, it is hard to evaluate the novel contribution of this paper compared to the existing approach. For example, much of the framework comes from the SawTooth paper, but this paper failed to include or summarize the overall structure. W3: Many parts of the paper are not justified or explained enough. For example, how modeling the relation with a Bernoulli distribution complement Phi modeling the relations between layers in SawTooth, or why the proposed approach is better was not explained. How is the adaptive structure different from just fine-tuning the parameters? The paper quickly attributes the improved performance to the joint tree likelihood, but how exactly is it better than the other approaches such as SawTooth or TopicNet? What information does it capture that SawTooth and TopicNet do not?	.W1: The presentation can be improved. There is no overview of the approach to explain the components, and a few components and concepts appear without much prior context. For example, "encoder" appears without where it is exactly being used. Same for "topic aggregation". "count vector" was used only once without definition.
ICLR_2022_1919	ICLR_2022	The proposed Plug-In inversion method is introduced very late in the paper (at the end of Section 3.4) even though it consists only of combining the augmentation and search space restriction techniques provided in the previous sections (3.1 - 3.3). It would have been much clearer for the reader if this fact was fully described earlier in Section 3 (for example, by moving Section 3.4 earlier in the paper), instead of delaying the presentation of the full method. My greatest concern is the fact that I am not sure whether the claim of the authors that the method can be applied to multiple networks without tuning is fully supported by the experimental results. More specifically, in Figures 7 and 8 the method is applied in a variety of networks, resulting in images which, while intelligible, are of varying quality. As such, to fully support the argument of the lack of need of extensive hyperparameter tuning, one would also need to show that using the same regularizer (for example, TV, which can be applied to any model) with the same hyperparameter leads to more extensive degradation, when applied to different models. I believe that a small example to demonstrate this motivation would greatly improve the paper. The above point is made even more unclear by the fact that one of the models considered by the authors does use a TV regularizer (which leads to drastic improvement in image quality). The ColorShift augmentation proposed by the authors is, unless I am mistaken, a variant of color jittering (in the sense that the adjustment is made directly to the pixels of the image, rather than on hue, saturation, contrast etc.). Given that applying some form of adjustment to the color of the image is a form of data augmentation which has been used in prior work (Krizhevsky et al., 2012, section 4.1), I am not sure if this data augmentation is as novel as the authors claim (although I do appreciate the qualitative analysis of its effects in the context of inversion). Questions: - The hyperparameters for ColorShift are fixed to a = b = 1 , and if I understand correctly this comes from a qualitative analysis of the results in a few simple experiments. Is this correct? Minor comments/typos: - As a minor comment related to the above, I believe the authors should indicate in the captions of the figures which model the respective images come from. - There is a space missing in the second-to-last line of page 6. - The caption in Figure 10 in the appendix is incomplete. References: Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems, 25, 1097-1105.	- As a minor comment related to the above, I believe the authors should indicate in the captions of the figures which model the respective images come from.
NIPS_2018_840	NIPS_2018	1. It is confusing to me what the exact goal of this paper is. Are we claiming the multi-prototype model is superior to other binary classification models (such as linear SVM, kNN, etc.) in terms of interpretability? Why do we have two sets of baselines for higher-dimensional and lower-dimensional data? 2. In Figure 3, for the baselines on the left hand side, what if we sparsify the trained models to reduce the number of selected features and compare accuracy to the proposed model? 3. Since the parameter for sparsity constraint has to be manually picked, can the authors provide any experimental results on the sensitivity of this parameter? Similar issue arises when picking the number of prototypes. Update after Author's Feedback: All my concerns are addressed by the authors's additional results. I'm changing my score based on that.	3. Since the parameter for sparsity constraint has to be manually picked, can the authors provide any experimental results on the sensitivity of this parameter? Similar issue arises when picking the number of prototypes. Update after Author's Feedback: All my concerns are addressed by the authors's additional results. I'm changing my score based on that.
ZyAwBqJ9aP	ICLR_2025	1. The paper does not introduce a novel method, it simply applies a typical graph neural network (GAT) and protein language model (ESM) to solve a binary classification problem. If the paper is to be improved, it will need to introduce a novel approach that brings with it significantly improved performance on this task. 2. The task is presented as a new task, however much work has been done predicting the substrates of CYPs (a leaderboard can be found on the Therapeutics Data Commons) including novel approaches to solve this problem. This paper takes a subset of the data used on the ESP model (Kroll et al) and applies a very similar architecture to a subset of the data. 3. The authors compare the performance of their model to the performance of other models and do not achieve the best results across all of the Isoforms. They claim this is because the DeepP450 model is trained on a smaller dataset which might have an impact on generalizability but this claim is not substantiated with evidence. 4. Almost two pages of the paper are dedicated to explaining how Graph Attention Networks work (which were not developed by the authors). Do you think that modeling the mechanics of P450 reactions could be a better approach?	3. The authors compare the performance of their model to the performance of other models and do not achieve the best results across all of the Isoforms. They claim this is because the DeepP450 model is trained on a smaller dataset which might have an impact on generalizability but this claim is not substantiated with evidence.
NIPS_2020_1671	NIPS_2020	I have some major concerns with the evaluation part of the paper. 1. Paper compared their method with influence functions and representer selection. A simple baseline could be a loss based selection method. Simply select training points based on loss change. A recent paper [DataLens IJCNN 20] shows that a simple loss based selection outperforms both influence functions and representer selection on mislabelled data identification when the mislabeled data is small. As the fraction of mislabelled data increases, influence function works better than loss based method. 2. Paper doesn't show the performance of TrackIn with varying amounts of mislabelled data. As pointed above, I expect TrackIn to perform poorly when we increase the mislabelled data. 3. Checkpoint ensembling is a widely used technique in machine translation [MT ensemble WMT16], semi-supervised learning [Temporal Ensembling ICLR17], knowledge distillation [KD Distillation NAACL 19] so it's not surprising that it helps in TrackIn. One can argue that influence functions can also benefit from the checkpoint ensembling. The authors should explain that. Also, the paper should cite prior work related to checkpoint ensembling as a motivation for picking multiple checkpoints.	1. Paper compared their method with influence functions and representer selection. A simple baseline could be a loss based selection method. Simply select training points based on loss change. A recent paper [DataLens IJCNN 20] shows that a simple loss based selection outperforms both influence functions and representer selection on mislabelled data identification when the mislabeled data is small. As the fraction of mislabelled data increases, influence function works better than loss based method.
4wAKqlfV5t	EMNLP_2023	1. The authors say that ''Although previous methods have proposed multimodal representations and achieved promising results, most of them focus on forming positive and negative pairs, neglecting the variation in sentiment scores within the same class'', do you mean previous MSA research mainly use contrastive learning for representation learning? 2. I can't find equation 3.2 because of the format mismatch. Do you mean equation (4)? 3. The benefit of the loss function defined in equation (11) should be detailed. Why it can model the volume of the difference of a pair of samples? 4. The modality-losing problem has been solved by several previous works, but the authors don't discuss them. For example, Tag-assisted Multimodal Sentiment Analysis under Uncertain Missing Modalities and Missing Modality Imagination Network for Emotion Recognition with Uncertain Missing Modalities. 5. More recent baselines should be compared, such as Counterfactual Reasoning for Out-of-distribution Multimodal Sentiment Analysis. 6. The performance improvement in the experimental section is not significant.	4. The modality-losing problem has been solved by several previous works, but the authors don't discuss them. For example, Tag-assisted Multimodal Sentiment Analysis under Uncertain Missing Modalities and Missing Modality Imagination Network for Emotion Recognition with Uncertain Missing Modalities.
ACL_2017_395_review	ACL_2017	My main concern with the paper is the magnification of its central claims, beyond their actual worth. 1) The authors use the term "deep" in their title and then several times in the paper. But they use a skip-gram architecture (which is not deep). This is misrepresentation. 2) Also reinforcement learning is one of the central claims of this paper. However, to the best of my understanding, the motivation and implementation lacks clarity. Section 3.2 tries to cast the task as a reinforcement learning problem but goes on to say that there are 2 major drawbacks, due to which a Q-learning algorithm is used. This algorithm does not relate to the originally claimed policy. Furthermore, it remains unclear how novel their modular approach is. Their work seems to be very similar to EM learning approaches, where an optimal sense is selected in the E step and an objective is optimized in the M step to yield better sense representations. The authors do not properly distinguish their approach, nor motivative why RL should be preferred over EM in the first place. 3) The authors make use of the term pure-sense representations multiple times, and claim this as a central contribution of their paper. I am not sure what this means, or why it is beneficial. 4) They claim linear-time sense selection in their model. Again, it is not clear to me how this is the case. A highlighting of this fact in the relevant part of the paper would be helpful. 5) Finally, the authors claim state-of-the-art results. However, this is only on a single MaxSimC metric. Other work has achieved overall better results using the AvgSimC metric. So, while state-of-the-art isn't everything about a paper, the claim that this paper achieves it - in the abstract and intro - is at least a little misleading.	4) They claim linear-time sense selection in their model. Again, it is not clear to me how this is the case. A highlighting of this fact in the relevant part of the paper would be helpful.
ICLR_2022_1895	ICLR_2022	1.It is obvious that this paper applies CVAE to the OOD data detection. The question is why to select CVAE as the efficient model to generate the OOD data. What is the motivation? 2.This paper claims that we can already produce comparable results to existing SOTA contrastive learning models but much more efficient. Why? The detailed explanation is necessary. 3.The contribution is mainly the metrics.	1.It is obvious that this paper applies CVAE to the OOD data detection. The question is why to select CVAE as the efficient model to generate the OOD data. What is the motivation?
NIPS_2019_573	NIPS_2019	of the paper: - no theoretical guarantees for convergence/pruning - though experiments on the small networks (LeNet300 and LeNet5) are very promising: similar to DNS [16] on LeNet300, significantly better than DNS [16] on LeNet5, the ultimate goal of pruning is to reduce the compute needed for large networks. - on the large models authors only compare GSM to L-OBS. No motivation given for the choice of the competing algorithm. Based on the smaller experiments it should be DNS [16], the closest competitor, rather than L-OBS, showed quite poor performance compared to others. - Authors state that GSM can be used for automated pruning sensitivity estimation. 1) While graphs (Fig 2) show that GSM indeed correlates with layer sensitivity, it was not shown how to actually predict sensitivity, i.e. no algorithm that inputs model, runs GSM, processes GSM result and output sensitivity for each layer. 2) Authors don't explain the detail on how the ground truth of sensitivity is achieved, lines 238-239 just say "we first estimate a layer's sensitivity by pruning ...", but no details on how actual pruning was done. comments: 1) Table 1, Table 2, Table 3 - "origin/remain params\|compression ratio\| non-zero ratio" --- all these columns duplicate the information, only one of the is enough. 2) Figure 1 - plot 3, 4 - two lines are indistinguishable (not even sure if there are two, just a guess), would be better to plot relative error of approximation, rather than actual values; why plot 3, 4 are only for one value of beta while plot 1 and 2 are for three values? 3) All figures - unreadable in black and white 4) Pruning majorly works with large networks, which are usually trained in distributed settings, authors do not mention anything about potential necessity to find global top Q values of the metric over the average of gradients. This will potentially break big portion of acceleration techniques, such as quantization and sparsification.	- Authors state that GSM can be used for automated pruning sensitivity estimation.
NkmJotfL42	ICLR_2024	I find the formal results stated in Sections 5 and 6 to be extremely difficult to follow. While the informal statements in Section 2 are understandably vague, Sections 5 and 6 failed to clarify my confusions from Section 2. I think this is due to two issues: 1. Section 4 did a poor job at explaining the formal notation. I feel that Definitions 1 and 2 are not particularly well-motivated (more on this later). And shoving everything else into the Appendix does not help either. 2. Certain points in the intro were not explained properly in later sections. a) The term "vacuous" was never formally defined, b) the connection between tightness of generalization bound (eq 1) and the notion of estimability is also not discussed in depth (why are they equivalent? I know the argument is not hard, but this is provides important contexts for the main theorems). 3. The theorem statements are pretty mouthful themselves and except for Theorem 2, the authors did not offer discussions that helps with parsing the theorem statements. Next, some major gaps I found in the paper: 4. The definition of over-parameterizaton (Definition 2) in this paper is not standard and my impression is that they are phrased to make the proofs simpler. While Definition 2 does intuitively fit the idea of over-parameterization, I feel strongly that the author should add: a) detailed discussions on why this definition is consistent with the standard setting in the literature, b) examples. 5. Similar to the previous point, in Theorem 3, the condition on TV distance is unmotivated and seems to only exist to make the problem easier. 6. In the proof of Theorem 1, the authors did not show the existence of Bayes-like Random ERM. A few minor comments regarding the proofs (I did not check Appendix G or H): Page 18: the theorem statement is about Theorem 2, not 1. Page 19: the final sentence should start with "the second equality holds" Page 20: the result "Theorem 1 in Angel & Spinka (2021)" is just a standard fact on the existence of coupling, so it is better to just say that directly. Page 20: the big equation block looks atrocious, please left align the lines and use indentation to make the + sign on the second line more visible. Page 20: Please define what event $B$ is before that big equation block. Also in the definition of $B$, the final inequality $L_{D_{I_1}}(A(S_1)) \ge \alpha$ is missing its RHS.	2. Certain points in the intro were not explained properly in later sections. a) The term "vacuous" was never formally defined, b) the connection between tightness of generalization bound (eq
NIPS_2021_776	NIPS_2021	weakness: 1 The theoretical parts (Section 3.2 and 3.3) are a bit hard to follow. 1-1. too many symbols are used. It would be more clear if the table list of all symbols is provided. 1-2. I am not sure if the following assumption is really valid: 1-2-1. lines 143-144: "Besides, under mild assumptions, if (E) ! 0 144 then % ! 0". 1-2-2. lines 148--149: "Though it is hard to conduct task-agnostic analysis on (t) term, we believe that perfect alignment is still an adequate criteria in minimizing (t) term". For the assumption made at lines 148--149, I'd like to know (t) is really reduced by using a perfect alignment encoder (PA-SF) in the robotic arm control benchmarks. also, lines 145--146 (and lines 631--636 in appendix): "Theoretically speaking, when # is a perfect alignment encoder, an goal-conditioned RL policy trained 146 over the encoded space {z(s)}s2S will minimize (E) to 0." I think this statement is not valid if multiple optimal policies \pi_G exist. (say that there are two optimal policies \pi_{G,1} and \pi_{G,2} and that \pi is converged to \pi_{G,1}, \epsilon^{e_i}(\pi \|\| \pi_{G, 2}) is still can be greater than zero.) Minor comments: Typos?: line 145: an goal-conditioned -> a goal-conditioned line 149: criteria -> criterion line 330: is formally formulated -> is formulated line 369: mdps -> {MDP}s Reference style is not consistent (e.g., the abbreviated style of conference name is used in some references, and not in the others). ---Edit after reading the author response and the other reviews:------ I would like to improve my score, as the author has basically adequately addressed my concerns: WR->WA I still have some concerns about the assumptions made in the theoretical analysis. For example, I saw the author's response regarding the term (t), but was not sufficiently convinced. I also think that the presentation of the theoretical analysis section needs to be improved (e.g., as suggested by Reviewer S2fL). Yes. # Discussion about societal impact is not contained in the paper, but I think it is not really necessary for this research.	1 The theoretical parts (Section 3.2 and 3.3) are a bit hard to follow. 1-1. too many symbols are used. It would be more clear if the table list of all symbols is provided. 1-2. I am not sure if the following assumption is really valid: 1-2-1. lines 143-144: "Besides, under mild assumptions, if (E) !
ICLR_2023_1584	ICLR_2023	Weakness: 1. The proposed method relies on a pretrained object detection network that contains the sufficient semantic information for the in-distribution data. When the semantic of in-distribution data like medical images is not covered by the object detection network (pre-trained on natural image), the built semantic graph can be incomplete or even erroneous. If we use additional annotations to train a sufficiently strong object detection network, the effort will be extremely expensive comparing the existing methods. 2. The paper is not polished and not ready to publish, with missing details in related work / experiment / writing. See more in "Clarity, Quality, Novelty And Reproducibility".	1. The proposed method relies on a pretrained object detection network that contains the sufficient semantic information for the in-distribution data. When the semantic of in-distribution data like medical images is not covered by the object detection network (pre-trained on natural image), the built semantic graph can be incomplete or even erroneous. If we use additional annotations to train a sufficiently strong object detection network, the effort will be extremely expensive comparing the existing methods.
hn0B3jTlwE	EMNLP_2023	- The paper lacks some results on the performance of the models on other tasks. While perplexity does not increase significantly, it would be interesting to know the impact of the method on other probing tasks. - A limitation of the method is its dependency on human-annotated datastores. The paper lacks some results on the quality of the datastores data on the final performance of the model.	- A limitation of the method is its dependency on human-annotated datastores. The paper lacks some results on the quality of the datastores data on the final performance of the model.
ICLR_2022_1648	ICLR_2022	Weakness/concerns: 1. The theoretical analysis is limited to regular graph for influence score and non-attribute graph for expressiveness of link representation. Could them be generalized to more applicable graphs? 2. The authors concatenate node representation as link representation. In this way, the expressiveness of link representation is highly related to the expressiveness of node representation. Therefore, it seems powerful GNNs for node representation or node classification can be directly used for link representation or link prediction. But it seems that P-GNN conflicts with this claim as the performance of P-GNN is really bad though the authors mention some concerns about it in supplements. 3. As the experimental results show, virtual nodes do not always benefit the link prediction, such as on Cora and Pubmed. Although the authors give some analysis, readers may still be confused about in what situations virtual nodes are recommended and vice versa. I would appreciate if the authors can give a table to further explain on it, especially clarify ambiguous expression in the article. For example, what “cora/pubmed have no cluster structure” means, when both Cora and Pubmed have clearly defined classes and previous works have shown that their data points have underlying clusters. 4. The proposed method seems to rely heavily on sophisticated hyperparameter searching. (as shown in Appendix D)	4. The proposed method seems to rely heavily on sophisticated hyperparameter searching. (as shown in Appendix D)
ICLR_2023_1400	ICLR_2023	- While the paper shows improvements on CIFAR derivatives, it lacks analysis or results on other datasets (e.g., ImageNet derivatives). Verifying the effectiveness of the framework on ImageNet-1k or even ImageNet-100 is important. These results ideally can be presented in the main paper. - The authors should add some details on how to solve the optimization in the main paper. It's an important piece of information currently lacking in the paper. - Some baselines such as [1] are not considered and should be added. I feel that influence function can be replaced by other influence estimation methods such as datamodels[2] or tracin[3]. It will be beneficial to understand if the updated framework results in better pruning than the baselines. I am assuming it would result in better pruning results, however it would be beneficial to understand which influence based methods are particularly suitable for pruning. [1]. https://arxiv.org/pdf/2107.07075 [2]. https://arxiv.org/abs/2202.00622 [3]. https://arxiv.org/abs/2002.08484	- Some baselines such as [1] are not considered and should be added. I feel that influence function can be replaced by other influence estimation methods such as datamodels[2] or tracin[3]. It will be beneficial to understand if the updated framework results in better pruning than the baselines. I am assuming it would result in better pruning results, however it would be beneficial to understand which influence based methods are particularly suitable for pruning. [1]. https://arxiv.org/pdf/2107.07075 [2]. https://arxiv.org/abs/2202.00622 [3]. https://arxiv.org/abs/2002.08484
NIPS_2016_482	NIPS_2016	of the method (see above) would clearly help in making the case for its impact. Clarity: The paper is very clearly written and easy to follow. It would be interesting to see a version of Fig. 1 including error bars estimated from the method - it seems that currently only the estimated means are ever used. More emphasis could be put on explaining the big picture of when the method is actually useful too. Other comments/questions: 1. In Eq. (1), why does y depend on theta but not f? 2. It would be nice to know the source of the variance seen in Fig. 1.	2. It would be nice to know the source of the variance seen in Fig.
ICLR_2022_3056	ICLR_2022	. It is claimed that the generated OOD samples cover larger diversity ranges. If the generated OOD samples all belong to the same classes as those in the training set, how to quantify such diversity? . The experiments show that with more synthesized data the OOD generalization improves. Then the question is what if we use only synthesized data for training classifiers? What's the performance? Is there any tradeoff between using real data and synthesized data? . The authors emphasized the superior performance of the proposed algorithm. However, with more synthetic data the training time would be significantly increased. It's better to add discussion on both the advantage and limitations of the proposed algorithm for a fair comparison with benchmarks. . In Algorithm 1 line 3: how to pick \theta_i to update \theta_j? To achieve a better performance, do you have to carefully pick a network for fine-tuning? . How to set the interpolation coefficients (Equ. 1) in your experiments? How do these parameters affect the performance of OOD generalization? . In Table 1, what's the difference between in distribution check mark and cross mark? What's the OOD data here? Are the results averaged over different OOD datasets or for some particular OOD dataset? . Results on colored fashion MNIST: the described numbers are incorrect. They are from Table 1, not Table 2.	. The authors emphasized the superior performance of the proposed algorithm. However, with more synthetic data the training time would be significantly increased. It's better to add discussion on both the advantage and limitations of the proposed algorithm for a fair comparison with benchmarks.
HZtBP6DZah	ICLR_2024	One major weakness is in clarity and presentation. This is a complicated model with many components, and I found it difficult to follow. Specific suggestions: - The authors may consider rewriting or reorganizing the last two paragraphs in Introduction. - Please provide a table of notations and variables in Appendix. - (1) does not make sense mathematically. Is Group_j a set or a vector or a scalar? For others, please see the list of questions below.	- (1) does not make sense mathematically. Is Group_j a set or a vector or a scalar? For others, please see the list of questions below.
l8zRnvD95l	ICLR_2025	1.The dataset was compiled from multiple sources with various modalities, which may introduce inconsistency or OOD samples when doing model training. Careful data analysis can be helpful 2. The experiment shows the proposed EcoPerceiver outperformed the current SOTA approach for most IGBP types especially WET, WAT, and ENF. However, the paper did not include an ablation study to showcase why the proposed model achieved this performance. 3.There is only one baseline compared and there is no model with single modalities.	3.There is only one baseline compared and there is no model with single modalities.
en3NwykrHW	ICLR_2025	1. There is no experiment. 2. In the upper bound of Theorem 7, the last three terms dominate. In contrast, the abstract and the introduction claim that the upper bound is determined by the first term. The authors should clarify under what conditions, if any, the first term dominates. If the first term is not asymptotically dominant, the authors should explain why they only focus on the first term in the abstract and the introduction. 3. The introduction claims that the developed algorithm for RL with trajectory feedback achieves the same asymptotically optimal regret bound as the standard RL. The authors should explain why trajectory feedback does not lead to a worse regret bound and what properties of their algorithm allow them to overcome the information disadvantage of only receiving trajectory feedback. 4. Section 3 should clarify that how the expected trajectory reward is a linear function of the state-action visitation frequencies. 5. Some mathematical derivations are not intuitive. The authors can add explanations about what the mathematical properties mean and how they are derived. Here are some examples. 5.1 The second key observation on page 5. 5.2 Inequality (3). 5.3 The equation in (8). 6. There are a few typos: The third term of the upper bound in Theorem 7. P1 in Line 4 of Algorithm 2. D2 in Line 6 of Algorithm 2.	3. The introduction claims that the developed algorithm for RL with trajectory feedback achieves the same asymptotically optimal regret bound as the standard RL. The authors should explain why trajectory feedback does not lead to a worse regret bound and what properties of their algorithm allow them to overcome the information disadvantage of only receiving trajectory feedback.